Flexible switch usable with CXL memory devices
The switch device addresses memory and communication challenges in AI/ML systems by converting packets between devices with different protocols, improving efficiency and performance through memory mapping and management.
Patent Information
- Application Number
- PCT/US2026/027465
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2025-05-08
- Filing Date
- 2026-05-08
- Publication Date
- 2026-11-12
AI Technical Summary
Artificial intelligence and machine learning systems face memory shortages and communication challenges due to processors, accelerators, and memory devices not sharing a common protocol, leading to inefficiencies in data transfer.
A switch device with multiple ports configured to communicate using different protocols and a switch fabric that converts packets between these protocols, enabling communication between devices with disparate protocols, and performs functions like memory mapping, interleaving, and collision management.
Facilitates efficient communication and memory sharing across devices with different protocols, enhancing performance and parallelism in AI/ML systems.
Smart Images

Figure US2026027465_12112026_PF_FP_ABST
Abstract
Description
PATENT APPLICATION Attorney Docket No.: MP12955WO Customer No. 184889 FLEXIBLE SWITCH USABLE WITH CXL MEMORY DEVICESCross References to Related Applications
[0001] This application claims the benefit of U.S. Provisional Patent App. No.63 / 802,356, entitled “FLEXIBLE SWITCH USABLE WITH ACCELERATORS AND CXL MEMORY DEVICES,” filed on May 8, 2025, the disclosure of which is expressly incorporated herein by reference in its entirety for all purposes.Field of Technology
[0002] The present disclosure relates generally to computing systems, and more particularly to path selection in a computing system.Background
[0003] The approaches described in this section are approaches that could be pursued, but not necessarily approaches that have been previously conceived or pursued.Therefore, unless otherwise indicated, it should not be assumed that any of the approaches described in this section qualify as prior art merely by virtue of their inclusion in this section.
[0004] Artificial intelligence and machine learning (AI / ML) systems utilize accelerators (e.g., Graphics Processing Units (GPUs), Neural Processing Units (NPUs), Intelligence Processing Units (IPUs), other processors, etc.) to perform complex tasks. These tasks often necessitate significant amounts of memory to complete efficiently, making memory shortage in AI / ML systems a common problem. Memory may be supplemented by discrete memory devices, offering systems designers some flexibility when building AI / ML systems. However, processors, accelerators, and memory devices may not share a common protocol. This may cause problems when transferring data between devices with differing protocols. Consequently, improved techniques for communication are desired.PATENT APPLICATION Attorney Docket No.: MP12955WO Customer No. 184889 Summary
[0005] In an embodiment, a switch device comprises: a plurality of ports configured to communicatively couple to a plurality of network devices, first ports among the plurality of ports configured to communicate at least according to a first protocol, and second ports among the plurality of ports configured to communicate at least according to a second protocol that is different than the first protocol; and a switch fabric configured to communicatively couple the plurality of ports. Each of at least some of the first ports is configured to send first packets formatted according to the first protocol to second ports via the switch fabric. Each of at least some of the second ports includes conversion circuitry configured to: convert each first packet received via the switch fabric to one or more other packets having a second format that corresponds to the second protocol, and send the one or more other packets to a network device communicatively coupled to the second port according to the second protocol.
[0006] In another embodiment, a method for interconnecting a plurality of network devices includes: receiving, at a plurality of first ports of a switch device, first memory access requests from at least first network devices amongst the plurality of network devices according to a first protocol; sending, by the plurality of first ports, first packets corresponding to the first memory access requests to a plurality of second ports via a switch fabric of the switch device; converting, at the plurality of second ports, the first packets to second memory access requests having a second format different than the first format, the second format corresponding to a second protocol different than the first protocol; and sending, by the plurality of second ports, the second memory access requests to at least first second devices amongst the plurality of network devices according to the second protocol.PATENT APPLICATION Attorney Docket No.: MP12955WO Customer No. 184889 Brief Description of the Drawings
[0007] Fig. 1A is a simplified diagram of an example switch device that communicatively couples computing devices to memory devices 104, according to an embodiment.
[0008] Fig. IB is a simplified diagram of an example computing system that utilizes multiple switches that have the same or similar structure of the switch device of Fig. 1A, according to an embodiment.
[0009] Fig. 2 is a simplified block diagram illustrating an example poll circuit of the switch device of Fig. 1A, according to an embodiment.
[0010] Fig. 3 is a simplified block diagram illustrating another example port circuit of the switch device of Fig. 1A, according to an embodiment.
[0011] Fig. 4 is a simplified block diagram of an example mapping circuit of the port circuit of Fig. 2, according to an embodiment.
[0012] Fig. 5 is a simplified block diagram of an example interrupt management circuit of the port circuit of Fig. 2, according to an embodiment.
[0013] Fig. 6 is a simplified block diagram of an example translation circuit of the port circuit of Fig. 3, according to an embodiment.
[0014] Fig. 7 is a flow diagram of an example method performed by the translation circuit of Fig. 6, according to an embodiment.
[0015] Fig. 8 is a flow diagram of another example method performed by the translation circuit of Fig. 6, according to an embodiment.
[0016] Fig. 9 is a simplified block diagram of an example collision management circuit of the port circuit of Fig. 3, according to an embodiment.
[0017] Fig. 10 is a flow diagram of an example method performed by the collision management circuit of Fig. 9, according to an embodiment.
[0018] Figs. 11A-B are simplified diagrams of an illustrative example in which the switch device of Fig. 1A is configured to provide direct memory access (DMA) memoryPATENT APPLICATION Attorney Docket No.: MP12955WO Customer No. 184889 operations between a host computer and a computing device, according to an embodiment.
[0019] Figs. 12A-B are simplified diagrams of another illustrative example in which the switch device of Fig. 1A is configured to provide DMA memory operations between a host computer and a computing device, according to another embodiment.
[0020] Fig. 13 is a simplified diagram of an illustrative example in which the switch device of Fig. 1A is configured to facilitate multiple computing devices sharing each of one or more memory devices, according to an embodiment.
[0021] Fig. 14 is a simplified diagram of an illustrative example in which the switch device of Fig. 1A is configured to facilitate memory expansion and / or pooling for multiple computing devices, according to an embodiment.
[0022] Fig. 15 is a simplified diagram of the switch device of Fig. 1A configured to provide compute express link (CXL) memory pooling and / or sharing, according to another embodiment.
[0023] Fig. 15 is a simplified diagram of the switch device of Fig. 1A configured to provide CXL memory pooling and / or sharing, according to another embodiment.Detailed Description
[0024] In some embodiments described herein, a switch includes a first of ports and a switch fabric. At least some ports of the plurality of ports are configured to be coupled with a plurality of computing devices (e.g., artificial intelligence / machine learning (Al / ML) accelerators, graphics processing units (GPUs), neural processing units (NPUs), intelligence processing units (IPUs), other suitable computer processors, etc.) using a first communication protocol. At least some ports of the plurality of ports are configured to be coupled with a plurality of memory devices using a second communication protocol that is different than the first communication protocol. The switch fabric communicatively couples the plurality of ports. The switch is configured to communicate with the plurality of computing devices using the first communication protocol and to communicate withPATENT APPLICATION Attorney Docket No.: MP12955WO Customer No. 184889 the plurality of memory devices using the second communication protocol. For example, the switch one or more of: i) allows computing device(s) to communicate with a host computer via the memory device(s); ii) allows computing devices to share memory device(s); iii) provides expanded memory capacity for the computing device(s), etc., according to various embodiments. The switch is configured to perform one or more of: i) perform memory mapping; ii) interleave data from the computing devices(s) to at least two memoiy devices of the plurality of memoiy devices; iii) account for differences between the first communication protocol and the second communication protocol (e.g., translate tags appended to packets exchanged between the plurality of computing devices and the plurality of memory devices); iv) perform collision management, etc., according to various embodiments. In other embodiments, the switch additionally or alternatively performs other suitable operations.
[0025] Fig. 1A is a simplified diagram of an example switch device (sometimes referred to herein as a “switch”) 100 that communicatively couples computing devices (“accelerators”) 102 to memory devices 104, according to an embodiment. Additionally, the switch 100 communicatively couples host computers (sometimes referred to herein as “hosts”) 106 to the accelerators 102 and / or the memory devices 104. In some embodiments, at least some computing devices 102 include memory that can be accessed by other devices via the switch 100, and / or at least some memory devices 104 include compute processors that can process data and / or access memory of other devices via the switch 100. Thus, the computing devices 102 and the memoiy devices 104 are sometimes referred to as “network devices.”
[0026] The switch 100 includes upstream port circuits (sometimes referred to as “upstream ports” or “USPs”) 110 (only two of which are shown), downstream accelerator ports (sometimes referred to as “downstream ports” or “DSPs”) 112, (only one of which is labeled), downstream memory device ports (sometimes referred to as “downstream ports” or “DSPs”) 114 (only one of which is labeled), an embedded processor (sometimesPATENT APPLICATION Attorney Docket No.: MP12955WO Customer No. 184889 referred to herein as an “embedded central processing unit” or “embedded CPU”) 120, and a switch fabric 130.
[0027] The hosts 106 are coupled to the upstream ports 110. Similarly, the accelerators 102 are coupled to the accelerator ports 112. The memory devices 104 are coupled to the memory device ports 114.
[0028] In an embodiment, the accelerators 102 are peripheral component interconnect express (PCIe) devices (i.e., each includes a respective communication interface that is configured to communicate according to the PCIe protocol) and the accelerator ports 112 each include a respective communication interface that is configured to communicate according to the PCIe protocol.
[0029] In an embodiment, the memory devices 104 are compute express link (CXL) memory devices (i.e., each includes a respective communication interface that is configured to communicate according to the CXL protocol), and memory device ports 114 each include a respective communication interface that is configured to communicate according to the CXL protocol.
[0030] In an embodiment, the hosts 106 are CXL devices (i.e., each includes a respective communication interface that is configured to communicate according to the CXL protocol), and upstream ports 110 each include a respective communication interface that is configured to communicate according to the CXL protocol.
[0031] Though described in the context of PCIe and CXL, other suitable communication protocols and / or suitable communication interfaces are used in other embodiments. Additionally or alternatively, the switch 100 includes other ports and / or connects to other suitable devices, in some embodiments. For example, in some embodiments, another switch (not shown in Fig. 1A) is coupled with a port of the switch 100 (e.g., an upstream port analogous to ports 110, coupled in an analogous manner as hosts 106).
[0032] The switch 100 facilitates communication between hosts 106, accelerators 102 and memory devices 104. The switch 100 is configured to communicate with thePATENT APPLICATION Attorney Docket No.: MP12955WO Customer No. 184889 accelerators 102 using a first protocol (e.g., PCIe in the above example) and to communicate with the memory devices 104 using a second protocol (e.g., CXL). In an embodiment, the switch 100 is configured to communicate with the hosts 106 using the second protocol (e.g., CXL). The switch 100 facilitates the accelerators 102 communicating with the hosts 106 via the memory devices 104, facilitates the accelerators 102 sharing the memory devices 104, and / or provides expanded memory capacity for the accelerators 102, according to various embodiments.
[0033] In various embodiments, the switch 100 performs one or more of: i) memory mapping; ii) interleaving data from an accelerator 102 to at least two memory devices 104; iii) accounting for differences between the first protocol and the second protocol (e.g., translating tags); iv) performing collision management, etc. The switch additionally or alternatively performs other suitable operations, in some embodiments.
[0034] In an embodiment, each of at least one accelerator 102 and / or each of at least one host 106 utilizes multiple memory devices 104 via the switch 100. For example, the accelerator 102 A stores data to and / or reads data from the memory devices 104C and 104D. In another embodiment, multiple accelerators 102 and / or hosts 106 share a single memory device 104. For example, the accelerators 102 A and 102C both store data to and / or read data from the memory device 104G. In some such embodiments, each of at least accelerator 102 and / or each of at least one host 106 stores data to and reads data from distinct portions of the single memory device 104. For instance, the single memory device 104 implements multiple logical devices, in an embodiment.
[0035] The switch 100 includes a port routing and address mapping (“routing / mapping”) circuit 134 that is configured to map an initial address range corresponding to an initial memory request received from an accelerator 102 to one or more new address ranges within an address space(s) of one or more devices (e.g., one or more hosts 106 and / or memory devices 104), according to an embodiment. The one or more new address ranges correspond to one or more new memoiy access requests to be subsequently routed to the one or more ports, in an embodiment. For example, inPATENT APPLICATION Attorney Docket No.: MP12955WO Customer No. 184889 connection with a memory access request received from an accelerator 102, the memory request corresponding to an initial address range within an address space of the accelerator 102, the routing / mapping circuit 134 maps the initial address range to one or more respective new address ranges within an address space of one or more of the memory devices 104, in an embodiment.
[0036] The switch 100 also includes an interleaving circuit 138 that is configured to interleave a memory access request from an accelerator 102 to multiple memory access requests across multiple memory devices 104. For example, in connection with an initial write request from an accelerator 102, the interleaving circuit 138 generates multiple new write requests that that correspond to writing respective subsets of data to respective memory devices 104 such that data corresponding to the initial write request is striped across multiple memory devices 104, in an embodiment. As another example, in connection with an initial read request from an accelerator 102, the interleaving circuit 138 generates multiple new read requests that that correspond to reading respective subsets of data from respective memory devices 104 such that data corresponding to the initial read request that is striped across multiple memory devices 104 is respectively read from the multiple memory devices 104, in an embodiment.
[0037] In another embodiment, the interleaving circuit 138 is additionally or alternatively configured to interleave multiple sequential memory access requests from an accelerator 102 to multiple memory devices 104. Interleaving multiple sequential memory access requests from an accelerator 102 to multiple memory devices 104 increases parallelism and / or performance, in an embodiment.
[0038] The switch 100 also includes an interrupt management circuit 140 that is configured to manage message signaled interrupts (MSIs) sent by the switch 100 in connection with completion of memory access requests. For example, when an accelerator 102 issues multiple new memory requests as part of a direct memory access (DMA) operation, the interrupt management circuit 140 controls the switch 100 to hold an MSI sent by the accelerator 102 in connection with the DMA operation until after allPATENT APPLICATION Attorney Docket No.: MP12955WO Customer No. 184889 of the multiple new memory requests are completed by one or more memory devices 104, in an embodiment.
[0039] The switch 100 also includes a translation circuit 144 that is configured to i) translate a first memory access request in a first format (e.g., a format that corresponds to a PCIe memory access request) to one or more second memory access requests in a second format (e.g., a format that corresponds to a CXL memory access request), in an embodiment. For example, in connection with a first memory access request having a first tag that corresponds to a PCIe memory access request, the translation circuit 144 i) generates, based on the first tag, one or more second tags that corresponds to one or more CXL memory access requests, and ii) generates one or more second memory access requests that correspond to the first memoiy access request with the first tag replaced by the one or more second tags, in an embodiment.
[0040] The translation circuit 144 is also configured to i) translate one or more first memory access responses in the second format (e.g., a format that corresponds to a CXL memory access response) to a second memory access response in the first format (e.g., a format that corresponds to a PCIe memoiy access response), in an embodiment. For example, in connection with one or more memory access responses having a second tag that corresponds to a CXL memory access request, the translation circuit 144 i) generates, based on the second tag, a first tag that corresponds to a PCIe memory access response, and ii) generates a second memory access response that corresponds to the first memory access request(s) with the second tag replaced by the first tag, in an embodiment.
[0041] The switch 100 further includes a collision management circuit 148 that is configured to prevent a new memory access request corresponding to a particular address range in a corresponding memory device 104 from being sent to the memory device 104 prior to the memory device 104 completing processing of a current memory access request corresponding to the particular address range, in an embodiment. For example, the collision management circuit 148 includes, for each of a least some of the memory devices 104, respective buffers corresponding to respective address ranges in thePATENT APPLICATION Attorney Docket No.: MP12955WO Customer No. 184889 corresponding memory device 104; in response to receiving a new memory access request corresponding to the particular address range in the corresponding memory device 104, the collision management circuit 148 stores the new memory access request in the buffer at least when the memory device 104 has not yet completed processing of a current memory access request corresponding to the particular address range, in an embodiment. When the memory device 104 completes processing of the current memory access request corresponding to the particular address range, the collision management circuit 148 retrieves a new memory access request from the buffer and sends the new memory access request to the memory device 104, in an embodiment.
[0042] Fig. IB is a simplified diagram of an example computing system 150 that utilizes multiple switches that have the same or similar structure of the switch 100 of Fig.1A, according to an embodiment. The computing system 150 is one of multiple computing systems in a data center, for example, each of the multiple computing systems having a same or similar structure, according to an embodiment. In some embodiments, the multiple computing systems are communicatively coupled and exchange data.
[0043] The computing system 150 corresponds to and / or is a component of an AI / ML system, in an embodiment. For example, the computing system 150 implements, or is one of a plurality of computing systems that implement, a neural network associated with an AI / ML system, in an embodiment.
[0044] The computing system 150 includes a set of y computing devices 102 (“accelerators”), where y is a suitable positive integer greater than two. Each accelerator 102 in the set is configured to perform computational operations and exchange data (e.g., associated with an AI / ML system and / or another suitable application) with other computing devices (including other accelerators 102) via a plurality of network switches 158, in an embodiment. The accelerators 102 include suitable processors configured to perform the computational operations and to exchange data, e.g., associated with an AI / ML system and / or another suitable application. In at least some embodiments in which the computing system 150 is associated with an AI / ML system, the acceleratorsPATENT APPLICATION Attorney Docket No.: MP12955WO Customer No. 184889 102 include Al accelerator processors such as GPUs, NPUs, Al accelerator processors implemented using field programmable gate arrays (FPGAs), Al accelerator processors implemented using application-specific integrated circuits (ASICs), etc.
[0045] The computing system 150 also includes N network switches 1 8, where N is a suitable positive integer greater than one. The network switches 158 are communicatively connected to the accelerators 102 via a plurality of network links 162. In an embodiment, each accelerator 102 is communicatively connected to each of the network switches 158 via a respective network link 162, i.e., each accelerator 102 is communicatively connected to the N network switches 158 via N network links 162. In an embodiment, each network switch 158 is communicatively connected to each of the accelerators 102 via a respective network link 162, i.e., each network switch 108 is communicatively connected to the y accelerators 102 via y network links 162.
[0046] In an embodiment, each of the network links 162 illustrated in Fig. IB corresponds to a single physical network link (e.g., corresponds to a single electrical or optical cable). In another embodiment, each of at least some of the network links 162 illustrated in Fig. IB corresponds to a set of multiple physical network links (e.g., corresponds to a set of multiple electrical or optical cables).
[0047] Each accelerator 102 is communicatively coupled to each other accelerator 102 via N alternative paths, in an embodiment. For example, the accelerator 102-1 is communicatively coupled to the accelerator 102-7 via N alternative paths, i.e., a first path via the network switch 158-1, a second path via the network switch 158-2, a third path via the network switch 158-3, and so on. Thus, when a first accelerator 102 is to transmit data to a second accelerator 102, the data can be transmitted via N alternative paths, each alternative path corresponding to a respective network switch 158, in an embodiment.
[0048] Although Fig. IB illustrates four accelerators 102 connected to each switch 100, a suitable quantity of accelerators 102 different than four is connected to each switch 100, in some embodiments. In some embodiments, two or more switches 100 are connected to different quantities of accelerators 102. Although Fig. IB illustrates fourPATENT APPLICATION Attorney Docket No.: MP12955WO Customer No. 184889 memory devices 104 connected to each switch 100, a suitable quantity of memory devices 104 different than four is connected to each switch 100, in some embodiments. In some embodiments, two or more switches 100 are connected to different quantities of memory devices 104. Although Fig. IB illustrates two hosts 106 connected to each switch 100, a suitable quantity of hosts 106 different than two is connected to each switch 100, in some embodiments. In some embodiments, two or more switches 100 are connected to different quantities of hosts 106.
[0049] In some embodiments, at least some of the ports 112 are configured to communicate according to both the first protocol and the second protocol. For example, when a port 112 is connected to a device that operates according to the first protocol, the port 112 communicates with the device according to the first protocol; whereas when the port 112 is connected to a device that operates according to the second protocol, the port 112 communicates with the device according to the second protocol, in an embodiment. Similarly, in some embodiments, at least some of the ports 114 are configured to communicate according to both the first protocol and the second protocol. For example, when a port 114 is connected to a device that operates according to the first protocol, the port 114 communicates with the device according to the first protocol; whereas when the port 114 is connected to a device that operates according to the second protocol, the port 112 communicates with the device according to the second protocol, in an embodiment.
[0050] Fig. 2 is a simplified block diagram illustrating an example port circuit (sometimes referred to as a “port”) 200 of a switch that is configured to communicatively couple at least two of i) computing devices (e.g., accelerators 102), memory devices 104 (e.g., memoiy devices 104), and iii) host computers (e.g., hosts 106), according to an embodiment. In an embodiment, the port 200 is included in each of at least some of the ports 112 of Fig. 1A, and Fig. 2 is described with reference to Fig. 1A for ease of explanation. In other embodiments, the port 200 is included in another suitable switch different than the switch 100 of Fig. 1A. In other embodiments, each of at least some of the ports 112 includes another suitable port circuit different than the port 200.PATENT APPLICATION Attorney Docket No.: MP12955WO Customer No. 184889
[0051] The port 200 includes PCIe interface circuitry (“PCIe interface”) 204 that is configured to communicate according to the PCIe protocol. The port 200 also includes fabric interface circuitry (“fabric interface”) 208 that is configured to interface with a switch fabric (e.g., the switch fabric 130) of the switch that includes the port 200.
[0052] The port 200 includes a routing / mapping circuit 212. In an embodiment, the routing / mapping circuits 212 for multiple ports 200 collectively correspond to the routing / mapping circuit 134 of Fig. 1A. In another embodiment, the routing / mapping circuit 134 includes one or more suitable routing / mapping circuits different than the routing / mapping circuit 212.
[0053] The routing / mapping circuit 212 is configured to map an initial address range corresponding to an initial memory request received via the PCIe interface 204 to one or more new address ranges within an address space(s) of one or more devices (e.g., one or more hosts 106 and / or memory devices 104), in an embodiment. The one or more new address ranges correspond to one or more new memory access requests to be subsequently routed to the one or more ports, in an embodiment. For example, in connection with a memory access request received from an accelerator 102, the memory request corresponding to an initial address range within an address space of the accelerator 102, the routing / mapping circuit 212 maps the initial address range to one or more respective new address ranges within an address space of one or more of the memory devices 104, in an embodiment.
[0054] The port 200 also includes an interleaving circuit 216. In an embodiment, interleaving circuits 216 for multiple ports 200 collectively correspond to the interleaving circuit 138 of Fig. 1A. In another embodiment, the interleaving circuit 138 includes one or more suitable interleaving circuits different than the interleaving circuit 216.
[0055] The interleaving circuit 216 is configured to convert a memory access request from an accelerator 102 to multiple memory access requests across multiple memory devices 104. For example, in connection with an initial write request from an accelerator 102, the interleaving circuit 216 generates multiple new write requests that thatPATENT APPLICATION Attorney Docket No.: MP12955WO Customer No. 184889 correspond to writing respective subsets of data to respective memory devices 104 such that data corresponding to the initial write request is striped across multiple memory devices 104, in an embodiment. As another example, in connection with an initial read request from an accelerator 102, the interleaving circuit 216 generates multiple new read requests that that correspond to reading respective subsets of data from respective memory devices 104 such that data corresponding to the initial read request that is striped across multiple memoiy devices 104 is respectively read from the multiple memory devices 104, in an embodiment.
[0056] In another embodiment, the interleaving circuit 216 is additionally or alternatively configured to interleave multiple sequential memory access requests from an accelerator 102 to multiple memoiy devices 104.
[0057] In an embodiment, the interleaving circuit 216 generates the multiple memory access requests using indications of addresses and, optionally, indications of memory devices received from the routing / mapping circuit 212. For example, in connection with an initial memory request received via the PCIe interface 204, the routing / mapping circuit 212 determines a plurality of new address ranges that correspond to an initial address range corresponding to the initial memory request, where the new address ranges correspond to multiple memory devices 104. In an embodiment, the routing / mapping circuit 212 also determines the multiple memory devices 104. The interleaving circuit 216 receives indications of the multiple new address ranges and, optionally, indications of the multiple memory devices, and the interleaving circuit 216 generates the multiple memory access requests using the indications of the multiple address ranges and, optionally, the indications of the multiple memory devices received from the routing / mapping circuit 212, in an embodiment.
[0058] The port 200 further includes an interrupt management circuit 220. In an embodiment, interrupt management circuits 220 for multiple ports 200 collectively correspond to the interrupt management circuit 140 of Fig. 1 A. In another embodiment,PATENT APPLICATION Attorney Docket No.: MP12955WO Customer No. 184889 the interrupt management circuit 140 includes one or more suitable interrupt management circuits different than the interrupt management circuit 220.
[0059] The interrupt management circuit 220 is configured to manage MSIs sent by the switch 100 in connection with completion of memory access requests. For example, when an accelerator 102 issues multiple new memory requests as part of a direct DMA operation, the interrupt management circuit 220 prompts the switch 100 to hold an MSI sent by the accelerator 102 in connection with the DMA operation until after all of the multiple new memory requests are completed by one or more memory devices 104, in an embodiment.
[0060] Fig. 3 is a simplified block diagram illustrating an example port circuit (sometimes referred to as a “port”) 300 of a switch that is configured to communicatively couple at least two of i) computing devices (e.g., accelerators 102), memory devices 104 (e.g., memory devices 104), and iii) host computers (e.g., hosts 106), according to an embodiment. In an embodiment, the port 300 is included in each of at least some of the ports 114 of Fig. 1A, and Fig. 3 is described with reference to Fig. 1A for ease of explanation. In other embodiments, the port 300 is included in another suitable switch different than the switch 100 of Fig. 1A. In other embodiments, each of at least some of the ports 114 includes another suitable port circuit different than the port 300.
[0061] The port 300 includes fabric interface circuitry (“fabric interface”) 304 that is configured to interface with a switch fabric (e.g., the switch fabric 130) of the switch that includes the port 300. The port 300 includes CXL interface circuitry (“CXL interface”) 308 that is configured to communicate according to the CXL protocol.
[0062] The port 300 includes a translation circuit 312. In an embodiment, translation circuits 312 for multiple ports 300 collectively correspond to the translation circuit 144 of Fig. 1 A. In another embodiment, the translation circuit 144 includes one or more suitable translation circuits different than the translation circuit 312.
[0063] The translation circuit 312 is configured to i) translate a first memory access request in a first format (e.g., a format that corresponds to a PCIe memory request) to onePATENT APPLICATION Attorney Docket No.: MP12955WO Customer No. 184889 or more second memory access requests in a second format (e.g., a format that corresponds to a CXL memory request), in an embodiment. For example, in connection with a first memory access request having a first tag that corresponds to a PCIe memory request, the translation circuit 312 i) generates, based on the first tag, one or more second tags that correspond to one or more CXL memory requests, and ii) generates one or more second memory access requests that correspond to the first memory access request with the first tag replaced by the one or more second tags, in an embodiment.
[0064] The translation circuit 312 is also configured to i) translate one or more first memory access responses in the second format (e.g., a format that corresponds to a CXL memory access response) to a second memory access response in the first format (e.g., a format that corresponds to a PCIe memory access response), in an embodiment. For example, in connection with one or more memory access responses having a second tag that corresponds to a CXL memory access request, the translation circuit 144 i) generates, based on the second tag, a first tag that corresponds to a PCIe memory access response, and ii) generates a second memory access response that corresponds to the first memory access request(s) with the second tag replaced by the first tag, in an embodiment.
[0065] The port 300 also includes a collision management circuit 316. In an embodiment, collision management circuits 316 for multiple ports 300 collectively correspond to the collision management circuit 148 of Fig. 1A. In another embodiment, the collision management circuit 148 includes one or more suitable collision management circuits different than the collision management circuit 316.
[0066] The collision management circuit 316 is configured to prevent a new memory access request corresponding to a particular address range in a corresponding memory device 104 from being sent to the memory device 104 prior to the memory device 104 completing processing of a current memory access request corresponding to the particular address range, in an embodiment. For example, the collision management circuit 316 includes a buffer corresponding to the particular address range in the corresponding memory device 104; in response to receiving a new memory access requestPATENT APPLICATION Attorney Docket No.: MP12955WO Customer No. 184889 corresponding to the particular address range, the collision management circuit 316 stores the new memory access request in the buffer at least when the memory device 104 has not yet completed processing of a current memory access request corresponding to the particular address range, in an embodiment. When the memory device 104 completes processing of the current memory access request corresponding to the particular address range, the collision management circuit 316 retrieves a new memory access request from the buffer and sends the new memoiy access request to the memoiy device 104, in an embodiment.
[0067] Referring again to Figs. 1A and 2, the routing / mapping circuit 134, 212 translates between memory addresses of first devices (e.g., the accelerators 104) that operate according to a first protocol (e.g., the PCle protocol) and corresponding memory addresses of one or more second devices (e.g., the memory devices 104 and / or the hosts 106) that operate according to a second protocol (e.g., the CXL protocol).
[0068] Referring to Figs. 2 and 3, in an embodiment in which the port 200 is configured to communicate according to the first protocol and the second protocol, the port 200 includes the CXL interface 308, the translation circuit 312, and the collision management circuit 316. In some embodiments in which the port 200 is configured to communicate according to the first protocol and the second protocol and includes the CXL interface 308, the translation circuit 312, and the collision management circuit 316, the port 200 is included in each of at least some of the ports 114 of Fig. 1A.
[0069] Fig. 4 is a simplified block diagram of an example mapping circuit 400 included in a routing / mapping circuit, according to an embodiment. The mapping circuit 400 is included in the routing / mapping circuit 134 and / or the routing / mapping circuit 212, in some embodiments, and Fig. 4 is described with reference to Figs. 1A and 2 for ease of explanation. In other embodiments, the routing / mapping circuit 134 and / or the routing / mapping circuit 212 includes a suitable mapping circuit different than the mapping circuit 400. In other embodiments, the mapping circuit 400 is included in another suitable routing / mapping circuit different than the routing / mapping circuit 134PATENT APPLICATION Attorney Docket No.: MP12955WO Customer No. 184889 and / or the routing / mapping circuit 212, and / or is included in another suitable switch different than the switch 100.
[0070] The mapping circuit 400 includes a memory device 404 that stores mapping information for mapping memory addresses of a first device (e.g., an accelerator 102) that operates according to a first protocol (e.g., the PCIe protocol) to corresponding memory addresses of one or more second devices (e.g., the memory devices 104 and / or the hosts 106) that operate according to a second protocol (e.g., the CXL protocol). In an embodiment, the memory 404 comprises a content addressable memory (CAM), such as a ternary CAM (TCAM). In other embodiments, the memory 404 comprises one or more other suitable memory devices different than a CAM.
[0071] The mapping circuit 400 also includes mapping logic 408 that is configured to use the mapping information stored in the memory device 404 to map memory addresses of the first device (e.g., the accelerator 102) that operates according to the first protocol (e.g., the PCIe protocol) to corresponding memory addresses of one or more second devices (e.g., the memory devices 104 and / or the hosts 106) that operate according to a second protocol (e.g., the CXL protocol).
[0072] In some embodiments in which the memory 404 comprises a CAM (e.g., a TCAM), the CAM stores M entries that correspond to M memory ranges of the memory addresses of the first device, where M is a suitable positive integer. As merely an illustrative example, M is 32. In other embodiments, M is a suitable positive integer different than 32. The mapping logic 408 provides an address corresponding to a memory access request from the first device, and applies the address as an input key to the TCAM 404. The TCAM 404 compares the input key to entries in TCAM 404 in parallel and, when a match is detected, the TCAM 404 outputs an index that indicates the entry to which the input key matched. The mapping logic 408 uses the index output by the TCAM 404 to determine i) an address range corresponding to one or more second devices, and ii) optionally, an indicator(s) of the one or more second devices.PATENT APPLICATION Attorney Docket No.: MP12955WO Customer No. 184889
[0073] In some embodiments, the mapping logic 408 is implemented using hardware circuitry. For example, the mapping logic 408 includes a hardware state machine configured to perform actions such as described above. In other embodiments, the mapping logic 408 is implemented additionally or alternatively using a processor that executes machine readable instructions, where the machine readable instructions, when executed by the processor, cause the processor to perform actions such as described above. In an embodiment, an embedded processor of the switch (e.g., the CPU 120 (Fig.1A)) at least partially implements the mapping logic 408.
[0074] In an embodiment, an embedded processor of the switch (e.g., the CPU 120 (Fig. 1 A)) configures the mapping circuit 400 at least by storing mapping information in the memory 408.
[0075] In some embodiments, at least some ordering rules of the first protocol (e.g., PCIe) are different than (e.g., not compatible with) at least some ordering rules of the second protocol (e.g., CXL). Thus, the switch 100 (Figs. 1A-B) is configured to provide ordering logic that enables satisfying ordering requirements of at least some memory access operations of the first protocol (e.g., PCIe) while sustaining memory load-store performance levels expected of the second format (e.g., CXL). For example, the switch 100 is configured to satisfy ordering requirements of one or more of write-after- write (WAW) operations, read-after-write (RAW) operations, and write-after-read (WAR) operations of PCIe while sustaining memory load-store performance levels expected of CXL, in some embodiments.
[0076] Referring again to Figs. 1 A-B, with PCIe, a sender of a memory access request can issue an MSI to signal that processing of the memory access request has commenced. With a system such as illustrated in Figs. 1A-B, however, this may cause problems with maintaining ordering of memory access requests, for example when an initial memoiy request from an accelerator 102 is converted to multiple memory access requests sent to multiple memory devices 104. Thus, when a port issues multiple new memory requests to multiple memory devices 104 in connection with one initial memory request from anPATENT APPLICATION Attorney Docket No.: MP12955WO Customer No. 184889 accelerator 102, the interrupt management circuit 140, 220 (Figs. 1A, 2) determines when the multiple memory devices 104 have completed processing of the multiple memory access requests and prompts the switch 100 to hold an MSI until after all of the multiple new memory requests are completed by one or more memory devices 104, in an embodiment.
[0077] Fig. 5 is a simplified block diagram of an example interrupt management circuit 500, according to an embodiment. The interrupt management circuit 500 is included in the interrupt management circuit 140 and / or the interrupt management circuit 220, in some embodiments, and Fig. 5 is described with reference to Figs. 1 A and 2 for ease of explanation. In other embodiments, the interrupt management circuit 140 and / or the interrupt management circuit 220 includes a suitable interrupt management circuit different than the interrupt management circuit 500. In other embodiments, the interrupt management circuit 500 is included in another suitable interrupt management circuit different than the interrupt management circuit 140 and / or the interrupt management circuit 220, and / or is included in another suitable switch different than the switch 100.
[0078] In some embodiments, operation of the interrupt management circuit 500 facilitates satisfying ordering requirements of at least some memory access operations of the first protocol (e.g., PCIe) while sustaining memory load-store performance levels expected of the second format (e.g., CXL). For example, operation of the interrupt management circuit 500 facilitates satisfying ordering requirements of one or more of WAW operations, RAW operations, and WAR operations of PCIe while sustaining memory load-store performance levels expected of CXL, in some embodiments.
[0079] The interrupt management circuit 500 includes a buffer 504 and management logic 508. In connection with a port 112 receiving an initial write request from an accelerator 102, the port 112 may also receive an MSI from the accelerator in connection with the initial write request. The port sends multiple write requests (which correspond to the initial write request) to multiple ports 114 connected to multiple memory devices 104. Additionally, the interrupt management circuit 500 receives the MSI, and thePATENT APPLICATION Attorney Docket No.: MP12955WO Customer No. 184889 management logic 508 stores the MSI in the buffer 504. Subsequently, the port 112 receives indications from the multiple ports 114 that indicate the multiple write requests have been processed by the multiple memory devices 104. In response to the management logic 508 determining that indications from all of the multiple ports 114 (that received the multiple write requests that corresponded to the initial write request) have been received, the indications indicating that the multiple write requests have been processed by the multiple memory devices 104, the management logic 508 retrieves the MSI from the buffer 504 and prompts the switch 100 to transmit the MSI. For example, the management logic 508 sends the MSI to the switch fabric 130, in an embodiment.
[0080] In some embodiments and / or scenarios, ports 114 do not send any indications that indicate write requests have been processed by the memory devices 104. In some such embodiments / scenarios, the port 112 forwards an MSI from the accelerator 102 to the host in response to receiving the MSI from the accelerator 102.
[0081] Referring again to Figs. 1A-B and 3, a PCIe memory access request can request access to a larger memory space as compared to a CXL memory request. Thus, the switch 100 (e.g., the hanslation circuit 144 and / or the hanslation circuit 312) is configured to convert one PCIe memory request that corresponds to requesting access to a first memory size into multiple CXL memory requests that respectively correspond to requesting access to second memory sizes that are each a fraction of the first memory size, at least in some circumstances, according to an embodiment. Similarly, the switch 100 is configured to convert multiple CXL memory access responses that respectively correspond to the second memory sizes into a single PCIe memory access response that corresponds to the first memory size, at least in some circumstances, according to an embodiment.
[0082] Fig. 6 is a simplified block diagram of an example hanslation circuit 600, according to an embodiment. The hanslation circuit 600 is included in the hanslation circuit 144 and / or the hanslation circuit 312, in some embodiments, and Fig. 6 is described with reference to Figs. 1A and 3 for ease of explanation. In otherPATENT APPLICATION Attorney Docket No.: MP12955WO Customer No. 184889 embodiments, the translation circuit 140 and / or the translation circuit 312 includes a suitable translation circuit different than the translation circuit 600. In other embodiments, the translation circuit 600 is included in another suitable translation circuit different than the translation circuit 144 and / or the translation circuit 312, and / or is included in another suitable switch different than the switch 100.
[0083] The translation circuit 600 is configured to i) translate a first memory access request in a first format (e.g., a format that corresponds to a PCIe memory request) to one or more second memory access requests in a second format (e.g., a format that corresponds to a CXL memory request), in an embodiment. For example, in connection with a first memory access request having a first tag that corresponds to a PCIe memory request, the translation circuit 600 i) generates, based on the first tag, one or more second tags that correspond to one or more CXL memory requests, and ii) generates one or more second memory access requests that correspond to the first memory access request with the first tag replaced by the one or more second tags, in an embodiment.
[0084] The translation circuit 600 includes translation logic 604 that is configured to receive memory access requests in the first format (e.g., PCIe) and, in response, generate memory access requests in the second format (e.g., CXL). In some embodiments, the translation logic 604 is also configured to receive memory access responses in the second format (e.g., CXL) and, in response, generate memory access responses in the first format (e.g., PCIe).
[0085] The translation circuit 600 includes a first-in-first-out (FIFO) buffer 608 that is configured to store at least portions of available tags in the second format (e.g., CXL), which are used by the translation logic 604 to generate memory access requests in the second format (e.g., CXL). An original tag table 612 is configured to store tags of memory access requests in the first format (e.g., PCIe) that are received by the translation circuit 600.
[0086] In some circumstances, the translation circuit 600 generates multiple memory access requests in the second format (e.g., CXL) for a corresponding one memory accessPATENT APPLICATION Attorney Docket No.: MP12955WO Customer No. 184889 request in the first format (e.g., PCIe), in an embodiment. For example, a PCIe memory access request can request access to a larger memory space as compared to a maximum memory space that can be accessed by a CXL memory request. In a merely illustrative embodiment, CXL memory access requests can access a maximum memoiy space size of 64 bytes (aligned to 64-byte memory blocks), whereas a PCIe memory access request can access a memory space size of 128 bytes or more. Thus, the translation circuit 600 is configured to convert one memory access request in the first format (e.g., PCIe) that corresponds to requesting access to a first memory size into multiple memory access requests in the second format (e.g., CXL) that respectively correspond to requesting access to second memory sizes that are each a fraction of the first memory size, at least in some circumstances, according to an embodiment. Similarly, the translation circuit 600 is configured to convert multiple memory access responses in the second format (e.g., CXL) that respectively correspond to the second memory sizes into a single memory access response in the first format (e.g., PCIe) that corresponds to the first memoiy size, at least in some circumstances, according to an embodiment.
[0087] When the translation circuit 600 converts one memory access request in the first format (e.g., PCIe) into multiple memory access requests in the second format (e.g., CXL), the translation circuit 600 assigns the multiple memory access requests in the second format with respective sequence numbers.
[0088] The translation circuit 600 includes a memory (e.g., a sequence number bit vector table) 616 that is configured to store, for each memoiy access request in the first format (e.g., PCIe) received by the translation circuit 600, information that indicates a quantity of memory access requests in the second format (e.g., CXL) generated by the translation circuit. In an embodiment, the information that indicates the quantity of memory access requests in the second format (e.g., CXL) comprises a bit vector with bits corresponding to respective sequence numbers; the translation logic 604 sets one or more bits corresponding to one or more sequence numbers of memory access requests in the second format that are generated to a first logic value (e.g., logic one), and the translationPATENT APPLICATION Attorney Docket No.: MP12955WO Customer No. 184889 logic 604 sets one or more bits corresponding to one or more sequence numbers of memory access requests in the second format that were not generated to a second logic value (e.g., logic zero).
[0089] Operation of the translation circuit 600 is discussed with reference to Figs. 7 and 8. Fig. 7 illustrates an example method 700 performed by the translation circuit 600 in response to receiving a memory access request in the first format (e.g., PCIe), according to an embodiment.
[0090] At block 704, the translation logic 604 retrieves (e.g., pops) an available tag in the second format (e.g., CXL) from the available new tag FIFO 608. The available tag retrieved from the FIFO 608 at block 704 includes an index value that the translation logic 604 subsequently uses to access the original tag table 612 and the sequence number bit vector table 616. At block 708, the translation logic 604 removes an original tag in the first format (e.g., PCIe) from the received memory access request in the first format, and stores the original tag in the original tag table 612. For example, the hanslation logic 604 stores the original tag at an entry in the original tag table 612 that corresponds to the index value of the available tag retrieved from the FIFO 608 at block 704.
[0091] At block 712, the translation logic 604 determines a quantity of memory access requests in the second format (e.g., CXL) that are to be generated in connection with the memory access request in the first format. As merely an illustrative example, when i) the maximum size of a memory space that can be accessed by a memory access request in the second format is X (e.g., 64 bytes or another suitable size), and ii) the memory space that is accessed by the memory access request in the first format is Y, the hanslation logic 604 determines the quantity of memoiy access requests in the second format (e.g., CXL) that are to be generated according to:k = Y / X Equation 1PATENT APPLICATION Attorney Docket No.: MP12955WO Customer No. 184889 at least when Y is an integer multiple of X. In another embodiment in which Y may be a value that is not an integer multiple of X, the translation logic 604 determines the quantity of memory access requests in the second format (e.g., CXL) that are to be generated according to:k = ceil(Y / X) Equation 2
[0092] At block 716, the translation logic 604 generates k (which is determined at block 712) memory access requests in the second format (e.g., CXL) using the available tag in the second format retrieved (e.g., popped) from the available new tag FIFO 608 at block 704. In an embodiment, generating each memoiy access requests in the second format comprises i) modifying the available tag in the second format retrieved (e.g., popped) from the available new tag FIFO 608 by adding a corresponding sequency number to generate a modified tag in the second format, and ii) adding the modified tag in the second format to the received memory access request in the first format from which the original tag in the first format was removed at block 708. In this way, the translation logic 604 generates a memory access request in the second format from the original memory access request in the first format at least by replacing the original tag in the first format with the modified tag in the second format.
[0093] At block 720, the translation logic 604 sets one or more bits in an entry of the sequence number bit vector table 616, the entry corresponding to the index value of the available tag retrieved from the FIFO 608 at block 704, and the one or more bits that are set indicating the quantity of the memory access requests in the second format that are generated at block 716. In an embodiment, bits of the entry correspond to respective sequence numbers, and the one or more bits that are set also indicate the sequence number(s) of the memory access requests in the second format that are generated at block 716. In an embodiment, setting the one or more bits at block 720 comprises setting the one or more bits to the first logic value (e.g., logic one).PATENT APPLICATION Attorney Docket No.: MP12955WO Customer No. 184889
[0094] Fig. 8 is a flow diagram an example method 800 performed by the translation circuit 600 in response to receiving a memoiy access response in the second format (e.g., CXL), according to an embodiment. In an embodiment, the received memory access response in the second format includes the same tag as was added to a corresponding memory access request in the second format, such as a memory access request generated using the method 700 (Fig. 7) or another suitable method.
[0095] At block 804, the hanslation logic 604 looks up an entry in the sequence number bit vector table 616 using information in a tag of the memory access response in the second format (e.g., CXL). At block 808, the translation logic 604 clears a bit (e.g., sets the bit to the second logic value (e.g., logic zero)) of the entry in the sequence number bit vector table 616 that was looked up at block 804.
[0096] At block 812, the hanslation logic 604 determines whether all memory access responses in the second format (e.g., CXL) corresponding to all memory access requests in the second format that were previously sent to the memory device 104 have been received. For example, the hanslation logic 604 determines whether all bits in the entry in the sequence number bit vector table 616 that was looked up at block 804 are cleared (e.g., set to the second logic value (e.g., logic zero)), in an embodiment.
[0097] In response to determining at block 812 that all memory access responses in the second format (e.g., CXL) corresponding to all memory access requests in the second format that were previously sent to the memory device 104 have not been received (e.g., determining that all bits in the entry in the sequence number bit vector table 616 that was looked up at block 804 are not cleared (e.g., set to the second logic value (e.g., logic zero)), the flow ends.
[0098] On the other hand, in response to determining at block 812 that all memory access responses in the second format (e.g., CXL) corresponding to all memoiy access requests in the second format that were previously sent to the memory device 104 have been received (e.g., determining that all bits in the entry in the sequence number bitPATENT APPLICATION Attorney Docket No.: MP12955WO Customer No. 184889 vector table 616 that was looked up at block 804 are cleared (e.g., set to the second logic value (e.g., logic zero)), the flow proceeds to block 816.
[0099] At block 816, the translation logic 604 retrieves, from the original tag table 612, the original tag in the first format corresponding to the received memory access response. For example, the translation logic 604 uses information in the tag of the received memory access response to look up an entry in the original tag table 612 that stores the original tag in the first format.
[0100] At block 820, the translation logic 604 removes the tag in the second format of the received memory access response; and at block 824, the translation logic 604 stores (e.g., pushes) at least a portion of the removed tag to the available new tag FIFO 608. In an embodiment, the sequence number in the removed tag is removed prior to storing the removed tag to the available new tag FIFO 608.
[0101] At block 828, the translation logic 604 generates a memory access response in the first format using the original tag retrieved at block 816.
[0102] Referring again to Figs. 1A and 2, the collision management circuit 148, 316 prevents a new memoiy access request corresponding to a particular address range in a corresponding memory device 104 from being sent to the memory device 104 prior to the memory device 104 completing processing of a current memory access request corresponding to the particular address range, in an embodiment.
[0103] Fig. 9 is a simplified block diagram of an example collision management circuit 900, according to an embodiment. The collision management circuit 900 is included in the collision management circuit 148 and / or the collision management circuit 316, in some embodiments, and Fig. 9 is described with reference to Figs. 1 A and 2 for ease of explanation. In other embodiments, the collision management circuit 148 and / or the collision management circuit 316 includes a suitable collision management circuit different than the collision management circuit 900. In other embodiments, the collision management circuit 900 is included in another suitable collision management circuitPATENT APPLICATION Attorney Docket No.: MP12955WO Customer No. 184889 different than the collision management circuit 148 and / or the collision management circuit 316, and / or is included in another suitable switch different than the switch 100.
[0104] In some embodiments, operation of the collision management circuit 900 facilitates satisfying ordering requirements of at least some memory access operations of the first protocol (e.g., PCIe) while sustaining memory load-store performance levels expected of the second format (e.g., CXL). For example, operation of the collision management circuit 900 facilitates satisfying ordering requirements of one or more of WAW operations, RAW operations, and WAR operations of PCIe while sustaining memory load-store performance levels expected of CXL, in some embodiments.
[0105] The collision management circuit 900 includes one or more memory devices 904 (sometimes referred to herein as “the collision management memory 904” for brevity). The collision management memory 904 is configured to store collision management information related to memory access requests that are to be sent to a memory device associated with the collision management circuit 900 (e.g., a memory device 104 connected to a port 114, 300 that includes the collision management circuit 900) at least for purposes of handling situations in which multiple memory access requests corresponding to a same memory segment of the memory device are received during a time period that is less than a time period required by the memory device to process one of the multiple memory access requests, or less than another suitable time period. A scenario in which multiple memory access requests corresponding to a same memory segment of the memory device being received during a time period that is less than a time period required by the memory device to process one of the multiple memory access requests (or another suitable time period) is sometimes referred to herein as a “collision.”
[0106] The collision management circuit 900 also includes collision management logic 908 that is configured to maintain the collision management information stored in the collision management memory 904 and to use the collision management informationPATENT APPLICATION Attorney Docket No.: MP12955WO Customer No. 184889 stored in the collision management memory 904 to manage collisions of memory access requests.
[0107] Table 1 illustrates an example of collision management information stored in the collision management memory 904.Table 1Region # Base Limit Read / Write Pending DMA Pending (I / O) Count Count Requests 0 0 63 0 1 641 64 63 0 2 1282 128 63 N / A 0 0
[0108] Region number (“Region #”) is an indicator of a particular region in the memory 104. Base is an indicator of a starting address of the region, and Limit is an indicator of an ending address of the region. Read / Write is an indicator of whether a current memory access request accessing the region is a write operation or a read operation. For example, a logic one indicates the region is being accessed by a current read operation, and a logic zero indicates the region is being accessed by a current write operation. In another embodiment, a logic one indicates the region is being accessed by a current write operation, and a logic zero indicates the region is being accessed by a current read operation. Pending Count indicates a number of remaining memory access operations that are left to process the current memory access request. DMA count indicates an amount of data remaining to be transferred for the current memory access request. Pending Requests are memory access requests for the region that have been received and that have not yet been processed.
[0109] In an embodiment, the collision management logic 908 is configured to maintain collision management information such as in Table 1 and to use such collisionPATENT APPLICATION Attorney Docket No.: MP12955WO Customer No. 184889 management information to ensure that memory access requests are processed according to an order.
[0110] Table 2 illustrates another example of collision management information stored in the collision management memory 904. The collision management information of Table 2 is similar to the collision management information of Table 1, but DMA Count is omitted.Table 2Region # Base Limit Read / Write Pending Pending(I / O) Count Requests0 0 63 0 11 64 63 0 12 128 63 N / A 0
[0111] In an embodiment, the collision management logic 908 is configured to maintain collision management information such as in Table 2 and to use such collision management information to ensure that memory access requests are processed according to an order.
[0112] The collision management memory 904 includes FIFOs 912 for storing collision management information for respective regions in the memory 104. Each FIFO 912 stores pending memory access requests (if any) in the second format (e.g., CXL) for the respective region in the memory 104. In an embodiment, each FIFO entry corresponding to a memory access request in the second format stores at least a portion of the memory access request.
[0113] The collision management information in the collision management memory 904 also includes an outstanding request vector (ORV) 916 that stores indications of memory access requests in the second format currently being processed by the memory 104. Elements of the ORV 916 correspond to respective regions in the memoiy 104, inPATENT APPLICATION Attorney Docket No.: MP12955WO Customer No. 184889 an embodiment. In an embodiment, each element of the ORV 916 stores at least a portion of an outstanding memory access request in the second format that is currently being processed by the memory 104 (if any) for the respective region of the memory 104.
[0114] In an embodiment, the collision management logic 908 is configured to maintain collision management information such as in the FIFOs 912 and / or the ORV 916 and to use such collision management information to ensure that memory access requests are processed according to an order.
[0115] In operation, when the collision management circuit 900 receives a new memory access request in the second format (e.g., CXL), the collision management logic 908 stores the new memory access request to the memory 904, in an embodiment. For example, the collision management logic 908 stores the new memory access request to a FIFO 912 that corresponds to a region of the memory 104 to be accessed when the new memory access request is executed by the memory 104.
[0116] In connection with the collision management circuit 900 outputting a memoiy access request in the second format (e.g., CXL) for sending to the memory 104, the collision management logic 908 updates the ORV 916 to indicate that the memory access request is currently being processed by the memory 104, in an embodiment. For example, the collision management logic 908 stores at least a portion of the memory access request in an entry of the ORV 916 that corresponds to a region of the memory 104 to be accessed when the new memory access request is executed by the memory 104.
[0117] When the collision management circuit 900 receives memory access response sent by the memory 104, the collision management logic 908 uses a tag of the memory access response to identify an entry of the ORV 916 that corresponds to the memoiy access request that prompted the memory access response, and the collision management logic 908 clears the entry of the ORV 916, in an embodiment. Clearing the entry comprises setting a value of a valid bit in the entry to a logic value that indicates the entry does not include valid data (e.g., indicates the entry is not occupied).PATENT APPLICATION Attorney Docket No.: MP12955WO Customer No. 184889
[0118] Fig. 10 is a flow diagram of an example method 1000 performed by the collision management circuit 900, according to an embodiment. In an embodiment, the method 1000 is performed by the collision management circuit 900 repeatedly. In an embodiment, the method 1000 is performed by the collision management circuit 900 periodically.
[0119] At block 1004, the collision management logic 908 determines memory ranges (if any) for which i) a cunent memory access request is not being processed, and ii) there are one or more pending memory access requests waiting to be executed. In an embodiment, block 1004 comprises determining elements of the ORV 916 that are not occupied, which indicates a memory access request for the corresponding memory region is not currently being processed. In another embodiment, block 1004 additionally or alternatively comprises determining memory segments for which both the pending count and DMA count (Table 1) are zero, which indicates a memory access request for the corresponding memory region is not currently being processed. In another embodiment, block 1004 additionally or alternatively comprises determining memory segments for which the pending count (Table 2) is zero, which indicates a memory access request for the corresponding memory region is not currently being processed.
[0120] In another embodiment, block 1004 additionally comprises, at least for each memory region for which a current memory access request is not being processed, determining whether there are any pending memory access requests for the memory region.
[0121] At block 1008, the collision management logic 908 selects a next pending memory access request to be processed from amongst pending memoiy access requests corresponding to memory regions for which a current memory access request is not being processed. In an embodiment, block 1008 comprises the collision management logic 908 selecting a particular memory region for which a current memory access request is not being processed according to an ordering of memory regions, and the collision management logic 908 selecting a next pending memory access for the selected memoryPATENT APPLICATION Attorney Docket No.: MP12955WO Customer No. 184889 region. In an embodiment, selecting the particular memory region for which a current memory access request is not being processed according to the ordering of memory regions comprises selecting the particular memory region for which a current memory access request is not being processed according to a round robin selection technique.
[0122] In an embodiment, selecting a next pending memory access request for the selected memory region comprises the collision management logic 908 retrieving (popping) a pending memoiy access request from a FIFO 912 corresponding to the selected memory region.
[0123] At block 1012, the collision management logic 908 prompts the port that includes the collision management logic 908 to issue the pending memory access request selected at block 1008.
[0124] At block 1016, the collision management logic 908 updates collision management information regarding the memory region that corresponds to the memory access request issued at block 1012. In an embodiment, updating the collision management information at block 1016 comprises updating the collision management information regarding the memory region to indicate that the memory 104 is currently processing a memory access request corresponding to the memory region. In an embodiment, updating the collision management information at block 1016 comprises updating one or more of i) the Read / Write bit corresponding to the memory region, ii) the Pending Count, and iii) the DMA count (Table 1). In another embodiment, updating the collision management information at block 1016 comprises updating one or more of i) the Read / Write bit corresponding to the memory region, and ii) the Pending Count (Table 2).
[0125] In another embodiment in which an accelerator 102 sends an MSI to the host when the accelerator 102 has transmitted all TLPs corresponding to a write request, a collision management circuit of the switch 100 is configured to monitor read requests to an address range for which a previous TLP of a write request was sent, and to block thePATENT APPLICATION Attorney Docket No.: MP12955WO Customer No. 184889 read requests at least until the collision management circuit is informed that processing of the write request is complete.
[0126] Referring again to Fig. 1A, the switch 100 is configurable to provide direct memory access (DMA) memory operations between one or more hosts 106 and one or more accelerators 102, according to an embodiment. In an illustrative embodiment, the switch 100 is configurable to facilitate: i) each of multiple memories 104 operating as a CXL memory expander for a host 106; ii) each of multiple accelerators 102 being allocated a respective portion of each of the CXL memory expanders as a PCIe memorymapped input / output (MMIO) resource; iii) interleaving across the multiple CXL memory expanders to improve bandwidth. In such an embodiment, the transfer of large amounts of data to the CXL memory expanders reduces computing load of the host 106 and / or improves speed. Additionally or alternatively, such an embodiment facilitates software coherency between the host 106 and the multiple accelerators 102.
[0127] Figs. 11A-B are simplified diagrams of an illustrative example in which the switch 100 is configured to provide direct DMA memory operations between a host 106 and an accelerator 102, according to an embodiment. More specifically, Figs. 11A-B illustrate a scenario in which the accelerator 102 writes data to multiple memory devices 104 (operating as CXL memory expanders), and the host 106 then reading the data from the multiple CXL memory expanders 104.
[0128] The host 106 first sets up the DMA operation at least by setting up an exclusive region in the memory devices 104A and 104D (operating as CXL memoiy expanders). In an embodiment, the host 106 sets up the exclusive region in the memory devices 104A and 104D by configuring the ports 114A and 114D. For example, the host 106 communicates with the ports 114A and 114D to set up the exclusive region in the memory devices 104 A and 104D by configuring collision management information in the ports 114A and 114D, according to an embodiment. In another embodiment, the host 106 communicates with CPU 120 regarding setting up the exclusive region in the memory devices 104A and 104D, and the CPU 120 then communicates with the portsPATENT APPLICATION Attorney Docket No.: MP12955WO Customer No. 184889 114A and 114D to configure collision management information in the ports 114A and 114D, according to an embodiment.
[0129] Next, the host 106 instructs the accelerator 102 A to start a DMA write to the exclusive region in the memory devices 104A and 104D. As part of the DMA write operation, the accelerator 102A sends PCIe write requests to the port 112A. The routing / mapping circuit 134, 212 of the port 112A maps addresses in the PCIe write requests to address ranges within address space(s) of the memory devices 104A and 104D, in an embodiment. In some scenarios, the interleaving circuit 138, 216 of the port 112A converts a PCIe write request to multiple write requests across the memoiy devices 104A and 104D.
[0130] The translation circuits 144, 312 of the ports 114A and 114D translate the write requests to CXL write requests, and the collision management circuits 148, 316 of the ports 114A and 114D prompts the ports 114A and 114D to issue the CXL write requests to the memories 104A and 104D, in an embodiment.
[0131] In connection with finishing the DMA write operation, the accelerator 102 A sends an MSI to the host 106 (via the switch 100) to signal that the DMA write operation is completed. As illustrated in Fig. 1 IB, in connection with receiving the MSI from the accelerator 102A, the host 106 issues CXL read requests to the memory devices 104A and 104D to read the data that was written by the accelerator 102A.
[0132] Figs. 12A-B are simplified diagrams of another illustrative example in which the switch 100 is configured to provide direct DMA memoiy operations between a host 106 and an accelerator 102, according to an embodiment. More specifically, Figs. 12A-B illustrate a scenario in which host 160 writes data to multiple memory devices 104 (operating as CXL memory expanders), and the accelerator 104 then reading the data from the multiple CXL memory expanders 104.
[0133] The host 106 first sets up the DMA operation at least by setting up an exclusive region in the memory devices 104A and 104D (operating as CXL memory expanders). In an embodiment, the host 106 sets up the exclusive region in the memoryPATENT APPLICATION Attorney Docket No.: MP12955WO Customer No. 184889 devices 104A and 104D by configuring the ports 114A and 114D. For example, the host 106 communicates with the ports 114A and 114D to set up the exclusive region in the memory devices 104A and 104D by configuring collision management information in the ports 114A and 114D, according to an embodiment. In another embodiment, the host 106 communicates with CPU 120 regarding setting up the exclusive region in the memory devices 104A and 104D, and the CPU 120 then communicates with the ports 114A and 114D to configure collision management information in the ports 114A and 114D, according to an embodiment.
[0134] Next, the host 106 issues CXL write requests to write data to the exclusive region in the memory devices 104A and 104D, as illustrated in Fig. 12A.
[0135] Next, the host 106 instructs the accelerator 102 A to start a DMA read from the exclusive region in the memory devices 104A and 104D. The DMA read operation is illustrated in Fig. 12B.
[0136] As part of the DMA read operation, the accelerator 102A sends PCIe read requests to the port 112A. The routing / mapping circuit 134, 212 of the port 112A maps addresses in the PCIe read requests to address ranges within address space(s) of the memory devices 104 A and 104D, in an embodiment. In some scenarios, the interleaving circuit 138, 216 of the port 112A converts a PCIe read request to multiple read requests across the memory devices 104A and 104D.
[0137] The translation circuits 144, 312 of the ports 114A and 114D translate the read requests to CXL read requests, and the collision management circuits 148, 316 of the ports 114A and 114D prompts the ports 114A and 114D to issue the CXL read requests to the memories 104A and 104D, in an embodiment.
[0138] In connection with finishing the DMA read operation, the accelerator 102 A sends an MSI to the host 106 (via the switch 100) to signal that the DMA read operation is completed. In connection with receiving the MSI, the host 106 may then start another similar exchange of data with the accelerator 102A by issuing CXL write requests to thePATENT APPLICATION Attorney Docket No.: MP12955WO Customer No. 184889 memory devices 104 A and 104D to write data to the exclusive region in the memory devices 104A and 104D.
[0139] Referring again to Fig. 1A, the switch 100 is configurable to provide DMA memory operations between multiple accelerators 102, according to an embodiment. In an illustrative embodiment, the switch 100 is configurable to facilitate: i) each of multiple memories 104 operating as a CXL memory expander for a host 106; ii) each of multiple memories 104 being shared by multiple accelerators; ii) each of multiple accelerators 102 sharing a same memory range of each of the CXL memory expanders; iii) interleaving across the multiple CXL memory expanders to improve bandwidth. In such an embodiment, software coherency between the multiple accelerators 102 and between the host 106 and the multiple accelerators 102 is facilitated.
[0140] Fig. 13 is a simplified diagram of an illustrative example in which the switch 100 is configured to facilitate multiple accelerators 102 sharing each of one or more memory devices 104, according to an embodiment. More specifically, Fig. 13 illustrates a scenario in which one accelerator 102B writes data to multiple memory devices 104 (operating as CXL memoiy expanders), and another accelerator 102 A then reading the data from the multiple CXL memory expanders 104.
[0141] The host 106 first sets up the memory sharing arrangement at least by setting up an exclusive region in the memory devices 104 A and 104B (operating as CXL memory expanders). In an embodiment, the host 106 sets up the exclusive region in the memory devices 104A and 104B by configuring the ports 114A and 114B. For example, the host 106 communicates with the ports 114A and 114B to set up the exclusive region in the memory devices 104A and 104B by configuring collision management information in the ports 114A and 114B, according to an embodiment. In another embodiment, the host 106 communicates with CPU 120 regarding setting up the exclusive region in the memory devices 104A and 104B, and the CPU 120 then communicates with the ports 114A and 114B to configure collision management information in the ports 114A and 114B, according to an embodiment.PATENT APPLICATION Attorney Docket No.: MP12955WO Customer No. 184889
[0142] Next, the host 106 instructs the accelerator 102B to start a DMA write to the exclusive region in the memory devices 104A and 104B. As part of the DMA write operation, the accelerator 102B sends PCIe write requests to the port 112B. The routing / mapping circuit 134, 212 of the port 112B maps addresses in the PCIe write requests to address ranges within address space(s) of the memory devices 104A and 104B, in an embodiment. In some scenarios, the interleaving circuit 138, 216 of the port 112B converts a PCIe write request to multiple PCIe write requests across the memory devices 104 A and 104B.
[0143] The translation circuits 144, 312 of the ports 114A and 114B translate the PCIe write requests to CXL write requests, and the collision management circuits 148, 316 of the ports 114A and 114B prompt the ports 114A and 114B to issue the CXL write requests to the memories 104A and 104B, in an embodiment.
[0144] In connection with finishing the DMA write operation, the accelerator 102B sends an MSI to the host 106 (via the switch 100) to signal that the DMA write operation is completed.
[0145] In connection with receiving the MSI, the host 106 instructs the accelerator 102 A to start a DMA read from the exclusive region in the memory devices 104A and 104B.
[0146] As part of the DMA read operation, the accelerator 102A sends PCIe read requests to the port 112A. The routing / mapping circuit 134, 212 of the port 112A maps addresses in the PCIe read requests to address ranges within address space(s) of the memory devices 104 A and 104B, in an embodiment. In some scenarios, the interleaving circuit 138, 216 of the port 112A converts a PCIe read request to multiple PCIe read requests across the memory devices 104A and 104B.
[0147] The translation circuits 144, 312 of the ports 114A and 114B translate the PCIe read requests to CXL read requests, and the collision management circuits 148, 316 of the ports 114A and 114B prompt the ports 114A and 114D to issue the CXL read requests to the memories 104A and 104B, in an embodiment.PATENT APPLICATION Attorney Docket No.: MP12955WO Customer No. 184889
[0148] In connection with finishing the DMA read operation, the accelerator 102 A sends an MSI to the host 106 (via the switch 100) to signal that the DMA read operation is completed. In connection with receiving the MSI, the host 106 may then start another similar exchange of data between the accelerators 102A and 102B by instructing the accelerator 102B to start another DMA write operation.
[0149] Referring again to Fig. 1A, the switch 100 is configurable to provide memory expansion and / or pooling for each of one or more accelerators 102, according to an embodiment. In an illustrative embodiment, the switch 100 is configurable to facilitate: i) each of multiple memories 104 operating as a CXL memory expander for a host 106; ii) each of multiple accelerators 102 being allocated a respective portion of each of the CXL memoiy expanders as a PCle MM10 resource; and iii) interleaving across the multiple CXL memory expanders to improve bandwidth.
[0150] Fig. 14 is a simplified diagram of an illustrative example in which the switch 100 is configured to facilitate memory expansion and / or pooling for multiple accelerators 102, according to an embodiment. More specifically, Fig. 14 illustrates a scenario in which each of accelerator 102A and accelerator 102B writes data to and reads data from respective exclusive regions in multiple memory devices 104 (operating as CXL memory expanders).
[0151] The host 106 first sets up the memory expansion and / or pooling arrangement at least by setting up respective exclusive regions in the memory devices 104A-D (operating as CXL memory expanders) for accelerator 102 A and accelerator 102D. In an embodiment, the host 106 sets up the exclusive regions in the memory devices 104A-D by configuring the ports 114A-D. For example, the host 106 communicates with the ports 114A-D to set up the exclusive regions in the memory devices 104A-D by configuring collision management information in the ports 114A-D, according to an embodiment. In another embodiment, the host 106 communicates with CPU 120 regarding setting up the exclusive regions in the memory devices 104A-D, and the CPUPATENT APPLICATION Attorney Docket No.: MP12955WO Customer No. 184889 120 then communicates with the ports 114A-D to configure collision management information in the ports 114A-D.
[0152] Next, the host 106 instructs the accelerator 102 A to start a DMA write to a first exclusive region in the memory devices 104A-D. As part of the DMA write operation, the accelerator 102A sends PCIe write requests to the port 112A. The routing / mapping circuit 134, 212 of the port 112A maps addresses in the PCIe write requests to address ranges within address space(s) of the memory devices 104A-D, in an embodiment. In some scenarios, the interleaving circuit 138, 216 of the port 112A converts a PCIe write request to multiple PCIe write requests across the memory devices 104A-D.
[0153] The translation circuits 144, 312 of the ports 114A-D translate the PCIe write requests to CXL write requests, and the collision management circuits 148, 316 of the ports 114A-D prompt the ports 114A-D to issue the CXL write requests to the memories 104A-D, in an embodiment.
[0154] In connection with finishing the DMA write operation, the accelerator 102 A sends an MSI to the host 106 (via the switch 100) to signal that the DMA write operation is completed.
[0155] In connection with receiving the MSI, the host 106 instructs the accelerator 102 A to start a DMA read from the first exclusive region in the memory devices 104A-D.
[0156] As part of the DMA read operation, the accelerator 102A sends PCIe read requests to the port 112A. The routing / mapping circuit 134, 212 of the port 112A maps addresses in the PCIe read requests to address ranges within address space(s) of the memory devices 104A-D, in an embodiment. In some scenarios, the interleaving circuit 138, 216 of the port 112A converts a PCIe read request to multiple PCIe read requests across the memory devices 104A-D.
[0157] The translation circuits 144, 312 of the ports 114A-D translate the PCIe read requests to CXL read requests, and the collision management circuits 148, 316 of the ports 114A-D prompt the ports 114A-D to issue the CXL read requests to the memories 104A and 104B, in an embodiment.PATENT APPLICATION Attorney Docket No.: MP12955WO Customer No. 184889
[0158] In connection with finishing the DMA read operation, the accelerator 102 A sends an MSI to the host 106 to signal that the DMA read operation is completed. In connection with receiving the MSI, the host 106 may then start another similar exchange of data between the accelerator 102A and the memories 104A-D by instructing the accelerator 102 A to start another DMA write operation.
[0159] Concurrently with the data transfers related to the accelerator 102A discussed above, the host 106 instructs the accelerator 102D to perform DMA write operations to and DMA read operations from a second exclusive region in the memory devices 104A-D in a similar manner. In other embodiments, the data transfers related to the accelerator 102 A and the data transfers related to the accelerator 102D are not performed concurrently.
[0160] Fig. 15 is a simplified diagram of the switch 100 configured to provide CXL memory pooling and / or sharing, according to another embodiment. The memory device 104A acts as a CXL memoiy resource for the host 106A, and the memory device 104B acts as a CXL memory resource for the host 106B. The memory device 104 A is shared between the accelerators 102A and 102B, and the memory device 104B is shared between the accelerators 102C and 102D. The accelerators 102A and 102B share a same memory range in the memory device 104 A, and the accelerators 102B and 102C share a same memory range in the memory device 104B. In such an embodiment, software coherency between the multiple accelerators 102 and between the host 106 and the multiple accelerators 102 is facilitated. In some embodiments, interleaving across multiple CXL memory expanders is used to improve bandwidth.
[0161] Fig. 16 is a simplified diagram of the switch 100 configured to provide CXL memory pooling and / or sharing, according to another embodiment. The memory device 104B acts as a CXL memoiy resource for the host 106 A, and the memory device 104C acts as a CXL memory resource for the host 106B. The memory device 104A acts as a shared CXL memory resource for both the host 106A and host 106B.PATENT APPLICATION Attorney Docket No.: MP12955WO Customer No. 184889
[0162] The memory device 104B is shared between the accelerators 102A and 102B, and the memory device 104C is shared between the accelerators 102C and 102D. The accelerators 102A and 102B share a same memory range in the memory device 104B, and the accelerators 102B and 102C share a same memory range in the memory device 104C.
[0163] Embodiment 1: A switch device, comprising: a plurality of ports configured to communicatively couple to a plurality of network devices, first ports among the plurality of ports configured to communicate at least according to a first protocol, and second ports among the plurality of ports configured to communicate at least according to a second protocol that is different than the first protocol; and a switch fabric configured to communicatively couple the plurality of ports; wherein each of at least some of the first ports is configured to send first packets formatted according to the first protocol to second ports via the switch fabric; and wherein each of at least some of the second ports includes conversion circuitry configured to: convert each first packet received via the switch fabric to one or more other packets having a second format that corresponds to the second protocol, and send the one or more other packets to a network device communicatively coupled to the second port according to the second protocol.
[0164] Embodiment 2: The switch device of embodiment 1, wherein the corresponding conversion circuitry of the each of the at least some second ports is configured to, for each first packet having the first format received by the second port, convert the first packet to the one or more other packets at least by: removing a first tag from the first packet to generate a modified first packet, the first tag corresponding to the first format; storing the first tag in a memory of the second port; and generating the one or more other packets based on the first packet, including generating each of the one or more other packets as a respective second tag added to the modified first packet, each respective second tag corresponding to the second format.
[0165] Embodiment 3: The switch device of either of embodiments 1 or 2, wherein: each of the at least some second ports includes a collision management circuit having aPATENT APPLICATION Attorney Docket No.: MP12955WO Customer No. 184889 plurality of first-in-first-out (FIFO) buffers corresponding to respective segments of a memory space of a memory device of the network device communicatively coupled to the second port; and the collision management circuit of the each of the at least some second ports is configured to: for each of the one or more other packets, store the other packet in a FIFO corresponding to a segment to be accessed when the memory device processes the other packet, retrieve other packets from the plurality of FIFOs, and prompt the second port to send other packets retrieved from the plurality of FIFOs to the network device communicatively coupled to the second port.
[0166] Embodiment 4: The switch device of embodiment 3, wherein the corresponding collision management circuit of the each of the at least some second ports is configured to: determine first segments of the address space for which the memory device is not currently processing memory access requests; select one FIFO from amongst non-empty FIFOs that correspond to the first segments of the address space; retrieve an other packet from the one FIFO; and prompt the second port to send the other packet retrieved from the one FIFO to the network device communicatively coupled to the second port.
[0167] Embodiment 5: The switch device of any of embodiments 1-4, wherein: each of the at least some second ports is configured to send second packets that have the first format to the at least some first ports via the switch fabric, the second packets corresponding to memory access responses corresponding to memory access requests received by the second port via the switch fabric.
[0168] Embodiment 6: The switch device of embodiment 5, wherein the corresponding conversion circuitry of the each of the at least some second ports is configured to, for each second packet to be sent by the second port via the switch fabric: receive one or more memory access responses from the memoiy device communicatively coupled to the second port according to the second protocol, the one or more memory access responses corresponding to one memory access request; and generate a secondPATENT APPLICATION Attorney Docket No.: MP12955WO Customer No. 184889 packet that corresponds to the one memory access request based on the one or more memory access responses, the second packet having the first format.
[0169] Embodiment 7: The switch device of embodiment 6, wherein the corresponding conversion circuitry of the each of the at least some second ports is configured to, for each second packet to be sent by the second port via the switch fabric, generate the second packet at least by: locating a first tag stored in a memory of the second port using one or more second tags of the one or more memory access responses, the first tag having been previously removed from the one memory access request, the first tag corresponding to the first format and the one or more second tags corresponding to the second format; retrieving the first tag from the memory of the second port; and generating the second packet based on the one or more memory access responses, the second packet including the first tag.
[0170] Embodiment 8: The switch device of embodiment 7, wherein the corresponding conversion circuitry of the each of the at least some second ports is configured to, for each second packet to be sent by the second port via the switch fabric, generate the second packet at least by: replacing the second tag from a first memory access response from among the one or more memory access responses with the first tag.
[0171] Embodiment 9: The switch device of any of embodiments 1-8, wherein each of at least some of the first ports includes interrupt management circuitry configured to: receive a message signaled interrupt (MSI) from a network device communicatively coupled to the first port, the MSI corresponding to a memoiy access request from the network device communicatively coupled to the first port; store the MSI in a memory of the first port while the memory access request is processed; and send the MSI to the switch fabric in response to determining that processing of the memory access request is completed.
[0172] Embodiment 10: The switch device of embodiment 9, wherein the corresponding interrupt management circuitry of the each of at least some of the first ports is configured to: determine that processing of the memory access request isPATENT APPLICATION Attorney Docket No.: MP12955WO Customer No. 184889 completed in response to receiving, via the switch fabric, one or more memory access responses corresponding to the memory access request.
[0173] Embodiment 11: A method for interconnecting a plurality of network devices, the method comprising: receiving, at a plurality of first ports of a switch device, first memory access requests from at least first network devices amongst the plurality of network devices according to a first protocol; sending, by the plurality of first ports, first packets corresponding to the first memory access requests to a plurality of second ports via a switch fabric of the switch device; converting, at the plurality of second ports, the first packets to second memory access requests having a second format different than the first format, the second format corresponding to a second protocol different than the first protocol; and sending, by the plurality of second ports, the second memory access requests to at least first second devices amongst the plurality of network devices according to the second protocol.
[0174] Embodiment 12: The method of embodiment 11, wherein converting the first packets to the second memory access requests comprises, for each first packet received by each second port: removing a first tag from the first packet to generate a modified first packet, the first tag corresponding to the first format; storing the first tag in a memory of the second port; and generating one or more second memory access requests based on the first packet, including generating each of the one or more second memory access requests as a respective second tag added to the modified first packet, each respective second tag corresponding to the second format.
[0175] Embodiment 13: The method of embodiment 12, further comprising, for each second port: for each of one or more second memory access requests generated based on each first packet, store the second memory access request in a first-in-first-out (FIFO) corresponding to a segment to be accessed when the memory device processes the second memory access requests; and for each second port, retrieving second memory access requests from the plurality of FIFOs; wherein sending the second memory access requestsPATENT APPLICATION Attorney Docket No.: MP12955WO Customer No. 184889 to the plurality of memory devices comprises sending second memory access requests that were retrieved from the plurality of FIFOs of each second port.
[0176] Embodiment 14: The method of embodiment 13, further comprising, for each second port: determining first segments of the address space for which the corresponding memory device is not currently processing memory access requests; selecting one FIFO from amongst non-empty FIFOs that correspond to the first segments of the address space; retrieving a second memory access request from the one FIFO; and prompting the second port to send the second memory access request retrieved from the one FIFO to the memory device.
[0177] Embodiment 15: The method of any of embodiments 11-14, further comprising: sending, by the plurality of second ports, second packets that have the first format to the plurality of first ports via the switch fabric, the second packets corresponding to memory access responses corresponding to the memory access requests.
[0178] Embodiment 16: The method of embodiment 15, further comprising, in connection with each second packet to be sent by each second port via the switch fabric: receiving one or more memoiy access responses from the corresponding memoiy device communicatively coupled to the second port according to the second protocol, the one or more memory access responses corresponding to one memory access request; and generating a second packet that corresponds to the one memory access request based on the one or more memory access responses, the second packet having the first format.
[0179] Embodiment 17: The method of embodiment 16, wherein, in connection with each second packet to be sent by each second port via the switch fabric, generating the second packet includes: locating a first tag stored in a memory of the second port using one or more second tags of the one or more memory access responses, the first tag having been previously removed from the one memoiy access request, the first tag corresponding to the first format and the one or more second tags corresponding to the second format; retrieving the first tag from the memory of the corresponding second port;PATENT APPLICATION Attorney Docket No.: MP12955WO Customer No. 184889 and generating the second packet based on the one or more memory access responses, the second packet including the first tag.
[0180] Embodiment 18: The method of embodiment 17, wherein, in connection with each second packet to be sent by each second port via the switch fabric, generating the second packet includes: replacing the second tag from a first memory access response from among the one or more memory access responses with the first tag.
[0181] Embodiment 19: The method of any of embodiments 11-18, further comprising: receiving, at each of at least some of the first ports, a message signaled interrupt (MSI) from a network device communicatively coupled to the first port, the MSI corresponding to a memory access request from the network device communicatively coupled to the first port; storing the MSI in a memory of the first port while the memory access request is processed; and sending, by the each of at least some of the first ports, the MSI to the switch fabric in response to determining that processing of the memory access request is completed.
[0182] Embodiment 20: The method of embodiment 19, further comprising: determining, by the each of at least some of the first ports, that processing of the memory access request is completed in response to receiving, via the switch fabric, one or more memory access responses corresponding to the memoiy access request.
[0183] Some of the various blocks, operations, and techniques described above may be implemented utilizing hardware, a processor executing firmware instructions, a processor executing software instructions, or any suitable combination thereof. When implemented utilizing a processor executing software or firmware instructions, the software or firmware instructions may be stored in any suitable computer readable memory. The software or firmware instructions may include machine readable instructions that, when executed by one or more processors, cause the one or more processors to perform various acts such as described above.PATENT APPLICATION Attorney Docket No.: MP12955WO Customer No. 184889
[0184] When implemented in hardware, the hardware may comprise one or more of discrete components, an integrated circuit, an application-specific integrated circuit (ASIC), a programmable logic device (PLD), etc.
[0185] While the present invention has been described with reference to specific examples, which are intended to be illustrative only and not to be limiting of the invention, changes, additions and / or deletions may be made to the disclosed embodiments without departing from the scope of the invention.
Claims
PATENT APPLICATION Attorney Docket No.: MP12955WO Customer No. 184889 What is claimed is:
1. A switch device, comprising:a plurality of ports configured to communicatively couple to a plurality of network devices, first ports among the plurality of ports configured to communicate at least according to a first protocol, and second ports among the plurality of ports configured to communicate at least according to a second protocol that is different than the first protocol; anda switch fabric configured to communicatively couple the plurality of ports; wherein each of at least some of the first ports are configured to send first packets formatted according to the first protocol to second ports via the switch fabric; and wherein each of at least some of the second ports includes conversion circuitry configured to:convert each first packet received via the switch fabric to one or more other packets having a second format that corresponds to the second protocol, andsend the one or more other packets to a network device communicatively coupled to the second port according to the second protocol.
2. The switch device of claim 1, wherein the corresponding conversion circuitry of the each of the at least some second ports is configured to, for each first packet having the first format received by the second port, convert the first packet to the one or more other packets at least by:removing a first tag from the first packet to generate a modified first packet, the first tag corresponding to the first format;storing the first tag in a memory of the second port; andPATENT APPLICATION Attorney Docket No.: MP12955WO Customer No. 184889 generating the one or more other packets based on the first packet, including generating each of the one or more other packets as a respective second tag added to the modified first packet, each respective second tag corresponding to the second format.
3. The switch device of claim 1, wherein:each of the at least some second ports includes a collision management circuit having a plurality of first-in-first-out (FIFO) buffers corresponding to respective segments of a memory space of a memory device of the network device communicatively coupled to the second port; andthe collision management circuit of the each of the at least some second ports is configured to:for each of the one or more other packets, store the other packet in a FIFO corresponding to a segment to be accessed when the memory device processes the other packet,retrieve other packets from the plurality of FIFOs, andprompt the second port to send other packets retrieved from the plurality of FIFOs to the network device communicatively coupled to the second port.
4. The switch device of claim 3, wherein the corresponding collision management circuit of the each of the at least some second ports is configured to:determine first segments of the address space for which the memory device is not currently processing memory access requests;select one FIFO from amongst non-empty FIFOs that correspond to the first segments of the address space;retrieve an other packet from the one FIFO; andPATENT APPLICATION Attorney Docket No.: MP12955WO Customer No. 184889 prompt the second port to send the other packet retrieved from the one FIFO to the network device communicatively coupled to the second port.
5. The switch device of claim 1, wherein:each of the at least some second ports is configured to send second packets that have the first format to the at least some first ports via the switch fabric, the second packets corresponding to memory access responses corresponding to memoiy access requests received by the second port via the switch fabric.
6. The switch device of claim 5, wherein the corresponding conversion circuitry of the each of the at least some second ports is configured to, for each second packet to be sent by the second port via the switch fabric:receive one or more memory access responses from the memory device communicatively coupled to the second port according to the second protocol, the one or more memory access responses corresponding to one memory access request; and generate a second packet that corresponds to the one memory access request based on the one or more memory access responses, the second packet having the first format.
7. The switch device of claim 6, wherein the corresponding conversion circuitry of the each of the at least some second ports is configured to, for each second packet to be sent by the second port via the switch fabric, generate the second packet at least by:locating a first tag stored in a memory of the second port using one or more second tags of the one or more memory access responses, the first tag having been previouslyPATENT APPLICATION Attorney Docket No.: MP12955WO Customer No. 184889 removed from the one memory access request, the first tag corresponding to the first format and the one or more second tags corresponding to the second format;retrieving the first tag from the memory of the second port; andgenerating the second packet based on the one or more memory access responses, the second packet including the first tag.
8. The switch device of claim 7, wherein the corresponding conversion circuitry of the each of the at least some second ports is configured to, for each second packet to be sent by the second port via the switch fabric, generate the second packet at least by:replacing the second tag from a first memory access response from among the one or more memory access responses with the first tag.
9. The switch device of claim 1, wherein each of at least some of the first ports includes interrupt management circuitry configured to:receive a message signaled interrupt (MSI) from a network device communicatively coupled to the first port, the MSI corresponding to a memory access request from the network device communicatively coupled to the first port;store the MSI in a memory of the first port while the memory access request is processed; andsend the MSI to the switch fabric in response to determining that processing of the memory access request is completed.
10. The switch device of claim 9, wherein the corresponding interrupt management circuitry of the each of at least some of the first ports is configured to:PATENT APPLICATION Attorney Docket No.: MP12955WO Customer No. 184889 determine that processing of the memory access request is completed in response to receiving, via the switch fabric, one or more memory access responses corresponding to the memory access request.
11. A method for interconnecting a plurality of network devices, the method comprising:receiving, at a plurality of first ports of a switch device, first memoiy access requests from at least first network devices amongst the plurality of network devices according to a first protocol;sending, by the plurality of first ports, first packets corresponding to the first memory access requests to a plurality of second ports via a switch fabric of the switch device;converting, at the plurality of second ports, the first packets to second memory access requests having a second format different than the first format, the second format corresponding to a second protocol different than the first protocol; andsending, by the plurality of second ports, the second memory access requests to at least first second devices amongst the plurality of network devices according to the second protocol.
12. The method of claim 11, wherein converting the first packets to the second memory access requests comprises, for each first packet received by each second port: removing a first tag from the first packet to generate a modified first packet, the first tag corresponding to the first format;storing the first tag in a memory of the second port; andPATENT APPLICATION Attorney Docket No.: MP12955WO Customer No. 184889 generating one or more second memory access requests based on the first packet, including generating each of the one or more second memory access requests as a respective second tag added to the modified first packet, each respective second tag corresponding to the second format.
13. The method of claim 12, further comprising, for each second port:for each of one or more second memory access requests generated based on each first packet, store the second memory access request in a first-in-first-out (FIFO) corresponding to a segment to be accessed when the memory device processes the second memory access requests; andfor each second port, retrieving second memory access requests from the plurality of FIFOs;wherein sending the second memory access requests to the plurality of memory devices comprises sending second memory access requests that were retrieved from the plurality of FIFOs of each second port.
14. The method of claim 13, further comprising, for each second port: determining first segments of the address space for which the corresponding memory device is not currently processing memory access requests;selecting one FIFO from amongst non-empty FIFOs that correspond to the first segments of the address space;retrieving a second memory access request from the one FIFO; and prompting the second port to send the second memory access request retrieved from the one FIFO to the memoiy device.PATENT APPLICATION Attorney Docket No.: MP12955WO Customer No. 184889 15. The method of claim 11, further comprising:sending, by the plurality of second ports, second packets that have the first format to the plurality of first ports via the switch fabric, the second packets corresponding to memory access responses corresponding to the memory access requests.
16. The method of claim 15, further comprising, in connection with each second packet to be sent by each second port via the switch fabric:receiving one or more memory access responses from the corresponding memory device communicatively coupled to the second port according to the second protocol, the one or more memory access responses corresponding to one memory access request; and generating a second packet that corresponds to the one memory access request based on the one or more memory access responses, the second packet having the first format.
17. The method of claim 16, wherein, in connection with each second packet to be sent by each second port via the switch fabric, generating the second packet includes:locating a first tag stored in a memory of the second port using one or more second tags of the one or more memory access responses, the first tag having been previously removed from the one memory access request, the first tag corresponding to the first format and the one or more second tags corresponding to the second format;retrieving the first tag from the memory of the corresponding second port; and generating the second packet based on the one or more memory access responses, the second packet including the first tag.PATENT APPLICATION Attorney Docket No.: MP12955WO Customer No. 184889 18. The method of claim 17, wherein, in connection with each second packet to be sent by each second port via the switch fabric, generating the second packet includes:replacing the second tag from a first memory access response from among the one or more memory access responses with the first tag.
19. The method of claim 11, further comprising:receiving, at each of at least some of the first ports, a message signaled interrupt (MSI) from a network device communicatively coupled to the first port, the MSI corresponding to a memory access request from the network device communicatively coupled to the first port;storing the MSI in a memory of the first port while the memory access request is processed; andsending, by the each of at least some of the first ports, the MSI to the switch fabric in response to determining that processing of the memoiy access request is completed.
20. The method of claim 19, further comprising:determining, by the each of at least some of the first ports, that processing of the memory access request is completed in response to receiving, via the switch fabric, one or more memory access responses corresponding to the memoiy access request.
