Regulation, sub-node grouping and balancing in rack-based architectures
By using orchestration layers and BMC in the rack for dynamic resource regulation and subdomain division, the impact of thermal and power events on SLA in the rack-based architecture is solved, and efficient operation and performance improvement of the processor core is achieved.
Patent Information
- Application Number
- CN201810216974.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Priority Date
- 2017-03-29
- Filing Date
- 2018-03-16
- Publication Date
- 2024-11-01
- Estimated Expiration
- 2038-03-16
AI Technical Summary
In a rack-mounted architecture, when power budgets or thermal events are used, the system linearly adjusts the computing nodes, affecting the implementation of the Service Level Agreement (SLA), and conventional operating systems cannot sense voltage changes in the processor core, resulting in core overheating and performance bottlenecks.
By introducing orchestration layers and BMCs into the rack, monitoring temperature and electrical characteristics with sensors, dynamically adjusting resource allocation of computing nodes, including processor frequency and voltage, to achieve thermal/power area regulation, and dividing the computing nodes into multiple subs domain to achieve processing balance.
Effectively reduces the performance impact of SLA, avoids core overheating, improves system flexibility and efficiency, allowing each processing core to operate at different voltages and frequencies, thereby improving overall performance.
Smart Images

Figure CN108694088B_ABST
Abstract
Description
Technical Field
[0001] The present disclosure relates generally to computer rack systems (also referred to as blade or SLED) including multiple nodes, and more particularly to thermal / power regulation, sub-node grouping, and processing balancing in such rack-based systems. Background Art
[0002] Disaggregated computing is an emerging field based on resource pooling. One disaggregated computing solution is known as rack-scale architecture (RSA).
[0003] In today's systems, when a power budget or thermal event occurs in a rack, the system linearly adjusts rack components (e.g., compute nodes). For most systems, this linear adjustment may affect the achievement of service level agreements (SLAs). For example, if a critical component (e.g., a storage node) is adjusted, the adjustment will affect the performance of all nodes in the node.
[0004] Furthermore, in conventional rack-type architectures, computing nodes are constructed at a bare metal level. Therefore, a rack owner provides at least one of the computing nodes and pooled system components, such as storage devices or network bandwidth, to a constructed system user. In addition, as technology advances, the number of processing cores in a processor in a system continues to increase. Therefore, some of the constructed system users may not need all of the cores in the processor.
[0005] Additionally, due to inherent manufacturing variations, individual cores may operate at different voltages. Conventional operating system (OS) schedulers are not aware of these variations in individual cores.
[0006] Therefore, conventionally, the system constrains all cores in a socket or core within a die to operate at the lowest common core voltage and frequency of all cores available.Thus, the OS scheduler places the workload evenly across the die space.
[0007] As a result, adjacent cores may overheat. This overheating can lead to core temperatures that cause performance bottlenecks. BRIEF DESCRIPTION OF THE DRAWINGS
[0008] Figure 1 An implementation of a rack according to one implementation of the present disclosure is shown;
[0009] Figure 2 An exemplary algorithm executed by an orchestration layer or BMC for regulation based on thermal / power zones according to one implementation of the present disclosure is shown;
[0010] Figure 3 An example processing core according to one implementation of the present disclosure is shown;
[0011] Figure 4 An example of a conventional computing node is shown;
[0012] Figure 5 An example of a computing node according to an implementation of the present disclosure is shown;
[0013] Figure 6 Another example of a computing node according to one implementation of the present disclosure is shown;
[0014] Figure 7 An algorithm for determining a configuration of a computing node according to one implementation of the present disclosure is shown;
[0015] Figure 8 An algorithm for selecting a subdomain to meet a job SLA according to one implementation of the present disclosure is shown;
[0016] Fig. 9 An example showing the frequency of compute nodes within a drawer of a rack according to one implementation of the present disclosure; and
[0017] Fig.10 An algorithm for assigning nodes based on a VF graph according to one implementation of the present disclosure is shown. DETAILED DESCRIPTION
[0018] Figure 1 An implementation of a rack 100 is shown according to one implementation of the present disclosure.
[0019] In many implementations, rack 100 operates in a software defined infrastructure (SDI). In SDI, the applications being executed and their service levels define the system requirements. SDI enables data centers to achieve greater flexibility and efficiency by enabling dynamic "optimization" of application resource allocation, enabling provisioning of services in minutes, and significantly reducing costs.
[0020] The rack 100 interfaces with an orchestration layer. The orchestration layer is implemented in software running on top of a POD manager in the disclosed rack-based design context. The POD manager manages a POD, which is a grouping of one or more racks that are typically managed by the POD manager.
[0021] The orchestration software provides, manages, and assigns resources based on data provided to the orchestration software by the service assurance layer. More specifically, the orchestration software is responsible for providing the following resources, such as computing resources, network resources, storage resources, and database resources, as well as composing and starting applications or workloads and monitoring hardware and software. Although the orchestration layer need not be included in the rack 100, the orchestration layer is indeed included in at least one implementation. The orchestration layer includes hardware logic or is executed by hardware logic. Hardware logic is an example of an orchestration unit.
[0022] Intelligent monitoring of infrastructure capacity and application resources helps orchestration software make decisions about workload placement based on actual, current data rather than estimates based on historical data or static models of average consumption demand.
[0023] The rack includes a plurality of drawers 110. Each drawer 110 includes a node slot 120, a sensor, and a node 130. Each of the nodes in the rack 100 is implemented at least in part in hardware. In this example, the node 130 is a computing node. However, the node may be a storage node, a field programmable gate array (FPGA), etc.
[0024] Node slot 120 accepts a compute node 130 for insertion. Figure 1 Two drawers 110 are shown, which include a total of one empty node slot 120 and three node slots occupied by computing nodes 130. Of course, this illustration is only for exemplary purposes and does not limit the implementation of the present disclosure in any way. For example, all of the node slots in the node slots can be filled by computing nodes 130. In addition, each drawer 110 can have fewer or more slots.
[0025] The node slot 120 includes a structure for mounting the computing node 130 within the drawer 110. The node slot 120 further includes wires for supplying power and for transmitting signals with the computing node 130.
[0026] Node slots 120 include sensors 140 that indicate when and whether a compute node 130 has been inserted into a corresponding node slot 120. Sensors 140 may send signals to the orchestration layer indicating the insertion of a compute node 130.
[0027] The sensors include sensors 150 and 160 . Figure 2 Sensors 150 and 160 are shown mounted on node 130. Additionally or alternatively, sensors 150 and 160 may also be mounted on node slot 120. Sensor 150 measures the temperature near compute node 130. Sensor 150 sends its measurements to controller 170. Sensor 150 is an example of a temperature sensing unit.
[0028] The sensor 160 measures electrical characteristics within the drawer 110. These electrical characteristics can be voltage, resistance, current, or a combination thereof (e.g., power). The sensor 160 can be located in multiple locations. As a result, the sensor 160 can measure, for example, the voltage difference between any two nodes in the same drawer or in some other location or the resistance across any two nodes. Similarly, the sensor 160 can determine the resistance across, for example, any wire within the drawer 110 or within the entire rack 100 or the current through any wire.
[0029] Controller 170 receives the transmitted measurements from sensors 150 and 160. Controller 170 controls various aspects of compute node 130 based on the measurements sensed by sensors 150 and 160. For example, controller 170 may implement at least a portion of the algorithm described below. Controller 170 also performs processing of jobs assigned to compute nodes 130 in the SLA. Controller 170 may transmit data (e.g., signals from sensors 150 and 160) to the orchestration layer of the rack.
[0030] Controller 170 may be a baseband management controller (BMC) or part of an orchestration layer. Controller 170 includes a cache memory. Controller 170 is an example of a processing unit.
[0031] In one implementation, computing node 130 includes non-volatile memory or a solid-state drive implemented as additional memory 180. Computing node 130 may also include networking resources. Memory 180 is an example of a storage unit.
[0032] Thermal / Power Zone Based Regulation
[0033] A service provider provides, for example, a cloud service by contracting with a user to provide a computer rack service at a defined service level. The resulting contract is called an SLA. The terms of these SLAs can be as complex as human creativity allows. However, in some instances, the orchestration layer or BMC of the rack can enforce the terms of these SLAs as a policy.
[0034] For example, a rack owner may define an SLA on a compute node basis. In such a case, the rack owner assigns at least a portion of one or more compute nodes of the rack to a user to perform a job. Optionally or alternatively, the rack owner may define an SLA based on a job duration, wherein the rack owner agrees to allow the user to perform the job within a defined duration (e.g., 30 minutes). The job may be for itself to use at least one compute node for the duration. The job may also be for completing a specific task regardless of the number of nodes within the duration. Optionally or alternatively, the SLA may require an amount of bandwidth, or a number of frames processed per second, or a number of instructions executed per second.
[0035] Optionally or alternatively, the rack owner may define an SLA based on the treatment of errors. For example, an error may be defined as an interrupted upload, and the corresponding treatment may be to re-execute the upload. In another case, an error may be defined as a program not running continuously during a job; the corresponding treatment for this error may be to re-execute the job. In another case, this process may be designated as a critical task. In this case, if any specified error occurs, the rack owner may be liable to the user to compensate for the financial loss.
[0036] During the implementation of these SLAs, errors (especially power-based errors) or thermal events may occur, for which the orchestration layer, BMC, basic input / output system (BIOS), microcode, or OS adjusts resources. These resources may be, for example, memory bandwidth, processor interconnect bandwidth, voltage applied to a processing core, frequency applied to a processing core, or the number of processor operations or memory accesses.
[0037] The reduction in bandwidth usually relates to the bandwidth provided between the computing node and an external network (such as the Internet). However, the reduction in bandwidth can also be related to the internal network bandwidth or the bus bandwidth within the node or rack.
[0038] The BMC, BIOS, orchestration layer, or OS scheduler can perform adjustments in a variety of ways. As an example, the BMC, BIOS orchestration layer, or OS scheduler can reduce or slow down the memory access bandwidth. That is, the BMC, BIOS orchestration layer, or OS scheduler can reduce the processor core clock. As another example, the BMC, BIOS, orchestration layer, or OS scheduler can reduce the voltage applied to the processor core. This reduction in system voltage or frequency can cause a reduction in the number of accesses or operations performed over a period of time, and thus reduce the bandwidth or power required by the system.
[0039] Typically processor, memory, or network resources are regulated. In the case where the BMC or orchestration layer regulates processors, accesses to the processor are related to the rate at which instructions are executed by the processor within a predetermined time. In the case where the BMC or orchestration layer regulates the number of memory accesses, these accesses are related to the number of reads and writes performed by the memory within a predetermined time period.
[0040] In one embodiment, the POD manager or orchestration layer first creates a map of nodes (e.g., node 130) in a rack (e.g., rack 100) based on signals received from sensors (e.g., sensor 140) in drawers (e.g., drawer 110). For example, the POD manager or orchestration layer determines the number of drawers in the rack and the location of each occupied node in each drawer. Next, the POD manager or orchestration layer associates each occupied node with the terms of the SLA corresponding to the node. In many cases, the POD manager or orchestration layer itself assigns the SLA to the node. The POD manager or orchestration layer then determines the errors that can be tolerated by each node without violating the terms of the SLA for the node.
[0041] When the system is operational, the POD manager or orchestration layer generates a thermal and / or power map of the rack based on signals received from sensors (e.g., sensors 150 and / or 160). The POD manager or orchestration layer may generate the map periodically. For example, the BMC or orchestration layer may generate (or update) the map every five seconds. Of course, the POD manager or orchestration layer may generate (or update) the map more frequently or less frequently, at random intervals, or at dynamically adjusted intervals.
[0042] Figure 2 An exemplary algorithm executed by the orchestration layer or BMC for regulation based on thermal / power zones is shown. Figure 2 The subject matter can be, but does not necessarily have to be, a combination of Figure 1 The theme is achieved.
[0043] The algorithm starts at S200 and proceeds to S210. In S210, the POD manager, the orchestration layer, or the sensor determines whether a regulation event occurs in a region of the rack. Additionally, the POD manager or the orchestration layer determines a region in which the regulation event occurs. The region may be a drawer, a node, a nearest portion (e.g., half, one third, one quarter) closest to a node, or some other defined region.
[0044] The regulation event may be a thermal event or a power event. A thermal event is an event in which the orchestration layer, POD manager, or sensor determines that the sensor 150 has sensed that the temperature in the area of the drawer has exceeded a predetermined threshold. A power event is an event in which the orchestration layer, POD manager, or sensor determines that the sensor 160 has sensed that the voltage difference or resistance has become zero or the current or power has exceeded a predetermined threshold in the area of the rack. In any case, the regulation event is not necessarily a thermal event or a power event, but may be a different event.
[0045] If the POD manager, orchestration layer or sensor determines that no adjustment event has occurred in the area, the algorithm returns to S210. On the contrary, if the POD manager, orchestration layer or sensor determines that a adjustment event has occurred, the algorithm proceeds to S230.
[0046] At S220, the orchestration layer or POD manager or BMC determines what resources are to be adjusted. As described above, these resources may be bandwidth, power, or the number of accesses to memory or processors. The algorithm then proceeds to S230.
[0047] In S230, the system determines the terms of each SLA in the region. The terms of these SLAs can usually be simplified to bandwidth, power consumption, or the number of accesses to memory or processors. Therefore, the system determines the specific resources suitable for adjustment. The algorithm then proceeds to S240.
[0048] In S240, the system determines the resources of the computing nodes. In one implementation, the system determines the resources of each computing node that performs non-mission-critical jobs. More specifically, the system determines the bandwidth, power consumption, or the number of accesses to memory or processors performed by the computing node. The algorithm then proceeds to S250.
[0049] In S250, the system determines the head room of resources for each computing node in the region. This determination is made with respect to the terms determined in S230 and the resources determined in S240. More specifically, the system determines whether the resources determined in S240 exceed the terms of the SLA determined in S230. If so, these excess resources can be adjusted by the system. The algorithm then proceeds to S260.
[0050] In S260, the system adjusts the appropriate resources of the computing nodes in the region based on the headroom determined in S250. For example, the system adjusts bandwidth, power consumption, or the number of accesses to memory or processors performed by the computing nodes. In one implementation, this adjustment is performed by reducing the system clock or system voltage. The algorithm then proceeds to S270.
[0051] In S270, the POD manager, BMC, or orchestration layer determines whether sufficient regulation has occurred. In one implementation, the POD manager, BMC, or orchestration layer determines whether the thermal event or power event has ended. If the thermal event or power event has not ended, the POD manager, BMC, or orchestration layer may determine that sufficient regulation has not occurred. If the system determines that sufficient regulation has not occurred, the algorithm returns to S220. On the other hand, if the POD manager, BMC, orchestration layer, or sensor determines that sufficient regulation has occurred, the algorithm ends at S280.
[0052] exist Figure 2 In some implementations of the algorithm, the BMC is replaced by a management controller. Examples of management controllers include a management engine (ME) and an innovation engine (IE) in a PCH.
[0053] Therefore, in the case of a regulation event, sufficient performance of certain nodes can be maintained. As a result, the impact on the SLA from the regulation event can be minimized, or even avoided.
[0054] Subdomain combination
[0055] Implementations according to the present disclosure partition a compute node into subdomains that allow multiple OSes to be booted at the bare metal level. Partitioning a compute node into subdomains can be based on capacity, such as the amount of memory, processor speed, and cache size.
[0056] The Platform Controller Hub (PCH) is an evolution of the previous Northbridge / Southbridge design. The PCH is typically directed to the Southbridge functionality. However, the PCH can also contain some Northbridge functionality, such as clocking. The PCH can also control the data path and support functions, such as the Direct Media Interface (DMI). In some implementations, the PCH can control the Flexible Display Interface (FDI). One implementation of the PCH is PCH.
[0057] At least a portion of the PCH functionality may be emulated by microcode within one or more processor cores and / or system management interrupts (SMIs) and some field programmable gate array (FPGA) components to embed the hypervisor into the BIOS itself.
[0058] Before turning to additional activities associated with the present disclosure, some basic information is provided to aid understanding. The number of cores in a system is not necessarily equal to the number of slots. A slot corresponds to a physical portion of a physical board (e.g., a printed circuit board [PCB]). Therefore, a SLED can host a varying number of slots. A slot includes multiple processing cores, an input / output controller, a memory controller, and a link controller.
[0059] Figure 3 An exemplary processor 300 is shown according to one implementation of the present disclosure. Processor 300 includes an integrated memory controller (IMC) 310, a platform controller hub (PCH) 320, a BIOS 330, a hypervisor 340, and a processing core 350. Many implementations include more than one processing core 350.
[0060] The IMC 310 controls a memory for the processing core 350. If the processing core 350 is split into a plurality of sub-domains, the IMC 310 controls a memory for each of the sub-domains.
[0061] The platform controller hub (PCH) 320 supports subdomains. Typically, there is one PCH per processing core 350. In some implementations of the present disclosure, the PCH may be emulated. Thus, virtual subdomains may be created without the need for additional PCHs.
[0062] BIOS 330 is a set of instructions stored in memory that are executed by a processor. More specifically, BIOS 330 is a type of firmware that performs hardware initialization during a boot process (e.g., power-on startup). BIOS 330 also provides runtime services for operating systems and programs.
[0063] BIOS 330 can be used for embedded hypervisor 340. When hypervisor 340 is embedded in BIOS, BIOS acts as a trusted virtual machine monitor (VMM) in which multiple OS can be loaded. BIOS 330 is considered a trusted component. Once a new BIOS image is signed, hypervisor 340 is also signed to ensure that the integrity of embedded hypervisor 340 is not compromised. The internal channel established by hypervisor 340 using BIOS 330 is trusted. Therefore, this embedding of hypervisor 340 allows the establishment of trusted communication to BMC.
[0064] Figure 4 An example of a conventional compute SLED 400 is shown. Compute SLED 400 includes processing sockets 405, 415, 425, and 435. Each of processing sockets 405, 415, 425, and 435 is implemented in hardware and may have multiple cores.
[0065] Processing socket 405 communicates with processing socket 415 via crosslink 410, with processing socket 435 via crosslink 430, and with processing socket 425 via crosslink 420. Processing socket 415 communicates with processing socket 435 via crosslink 445 and with processing socket 425 via crosslink 440. Processing socket 435 communicates with processing socket 425 via crosslink 450. Crosslinks 410, 420, 430, 440, 445, and 450 are intra-processor crosslinks.
[0066] Although computing SLED 400 is shown with four processing slots, computing SLED 400 may include additional or fewer processing slots. In this case, each individual processing slot may communicate with each other processing slot via a crosslink. However, a particular processing slot does not need to communicate with each other processing slot via a crosslink.
[0067] Figure 5An example of a compute SLED 500 according to an implementation of the present disclosure is shown. Compute SLED 500 includes processing slots 505, 515, 525, and 535. Each of processing slots 505, 515, 525, and 535 is implemented in hardware.
[0068] exist Figure 5 , crosslinks 510, 520, 530, 540, and 545 have been disabled by hypervisor 340. Thus, processing socket 505 does not communicate with processing socket 515 via crosslink 510, does not communicate with processing socket 535 via crosslink 530, and does not communicate with processing socket 525 via crosslink 520. Processing socket 515 does not communicate with processing socket 535 via crosslink 545, and does not communicate with processing socket 525 via crosslink 540. However, as Figure 5 As shown, processing socket 535 communicates with processing socket 525 via crosslink 550. Crosslinks 510, 520, 530, 540, 545, and 550 are intra-processor crosslinks.
[0069] In this manner, processing slots 505 and 515 operate independently of each other and independently of processing slots 525 and 535. Processing slots 525 and 535 operate in coordination. Figure 5 Three domains are shown, the first being processing slot 505, the second being processing slot 515, and the third being processing slots 525 and 535. Because these three domains do not communicate with each other, they are considered isolated.
[0070] Figure 5 One implementation of crosslinks disabled by hypervisor 340 is shown. In another implementation, crosslinks 510, 520, 530, 540, 545, and 550 are all disabled by hypervisor 340. Thus, each of processing slots 505, 515, 525, and 535 operates independently of one another. In yet another implementation, crosslinks 520, 530, 540, and 545 are disabled by hypervisor 340, but crosslinks 510 and 550 are not disabled by hypervisor 340. Thus, compute SLED 500 may include two domains, a first being processing slots 550 and 515, and a second being processing slots 525 and 535.
[0071] Figure 6 An example of a compute SLED 600 according to one implementation of the present disclosure is shown. Compute SLED 600 includes processing slots 605, 615, 625, and 635. Each of processing slots 605, 615, 625, and 635 is implemented in hardware.
[0072] Additionally, each of processing sockets 605, 615, 625, and 635 includes a hypervisor. Thus, compute SLED 600 effectively has five domains. Crosslinks 610, 620, 630, 640, 645, and 650 are intra-processor crosslinks.
[0073] Figure 6 The hypervisor implementation is shown as being implemented by the hypervisor 340 Figure 5 However, such illustration is only for explanation purposes. Figure 6 The hypervisor implementation can be found in Figure 5 In addition, including the cross-linking disabled by the hypervisor 340 Figure 5 Implementations that disable crosslinking do not need to include Figure 6 Similarly, including Figure 6 The hypervisor-implemented implementation need not include the hypervisor 340 implementing Figure 5 Cross-disabled.
[0074] Each of the cores of processor sockets 605, 615, 625, and 635 can be further subdivided when the system is partitioned. For example, assume that each of processor sockets 605, 615, 625, and 635 contains 10 cores. In this case, computing SLED 600 can be cooperatively operated as a single entity including 4 sockets and 10 cores per socket. Computing SLED 600 can also be operated as two entities, each of which includes 2 sockets and 10 cores per socket. In addition, computing SLED 600 can be operated as four entities, each of which includes 1 socket and 10 cores per socket. Such an implementation is an example of socket-level partitioning. In addition, the socket can be subdivided. Therefore, in one example, computing SLED 600 can be operated as eight entities, each of which includes one socket and five cores. Of course, the socket can be subdivided into any number of cores. An operating system can be run on each subdivided socket.
[0075] Figure 7 An algorithm for determining a configuration of a computing node according to one implementation of the present disclosure is shown.
[0076] The algorithm starts at S700 and proceeds to S710, where the BMC or orchestration layer receives a target number of domains. In some implementations, the rack owner specifies the target number of domains. In various implementations, the application or job itself can provide the target number of domains.
[0077] In other implementations, the BMC or orchestration layer receives the SLA. The BMC or orchestration layer can determine a set of resources from the SLA, such as the number of cores, the amount of memory, or the amount of network bandwidth. In at least one implementation, the BMC or orchestration layer receives a set of resources, such as the number of cores, the amount of memory, or the amount of network bandwidth. The compute node then determines a minimum set of resources (e.g., processor cores) to provide these resources.
[0078] Furthermore, in many embodiments, the target number of domains is a power of 2. However, the target number of domains is not limited thereto. The algorithm then proceeds to S720.
[0079] In S720, hypervisor 340 determines the crosslinks to be disabled (eg, crosslinks 510, 520, 530, 540, 545, and 550). The algorithm then proceeds to S730.
[0080] In S730, the hypervisor 340 disables the plurality of crosslinks. The algorithm then proceeds to S740.
[0081] In S740, the BMC or orchestration layer determines a processing core to implement the hypervisor. The use of the hypervisor slightly degrades overall performance. Therefore, the use of the hypervisor has a lower priority than other operations for changing the number of subdomains. The identified processing core implements the hypervisor. The algorithm then proceeds to S750.
[0082] In S750, the BMC or orchestration layer determines the processing cores for splitting its IMCs and splits multiple IMCs. These processing cores are the cores that are split using the hypervisor. The algorithm then proceeds to S760.
[0083] Thus, by subdividing the processor core at the socket level in S730 and implementing a hypervisor at S740 , the processor core is divided into the target number of sub-domains received at S710 .
[0084] In S760, the processing core uses the processor, IMC and hypervisor to process data (eg, execute the job defined in the SLA). The algorithm then ends in S770.
[0085] Figure 8 An algorithm for selecting a subdomain for a job to meet an SLA according to one implementation of the present disclosure is shown.
[0086] The algorithm starts at S800 and proceeds to S810, where the server receives a job request associated with an SLA. The algorithm then proceeds to S820.
[0087] In S820, a rack-based design (RSD) server receives a resource request. For example, a job may use predefined processing or memory resources (e.g., 8 cores or 2TB of memory). The RSD server receives an indication of these resources from the application. Thus, depending on the job, the system knows the number of processors to be provided. The algorithm then proceeds to S830.
[0088] In S830, the RSD server notifies the BMC of the job and the resources to be provided. The algorithm then proceeds to S840.
[0089] In S840, the BMC determines available resources, as described later in Fig. 9 Then, the algorithm proceeds to S840.
[0090] In S850, resources are selected. In one implementation, the BMC selects the minimum resources. More specifically, the BMC selects the subdomain with the minimum resources that meet the SLA. In another implementation, the BMC provides a list of available resources to the POD manager. The POD manager then selects the resources. Therefore, the algorithm can use fewer resources for each SLA. The algorithm then proceeds to S860, where the algorithm ends.
[0091] Thus, some implementations of the present disclosure allow a compute node to expose the system as a configurable system that implements multiple subdomains. A user can request a certain configuration, such as the number of cores and the amount of memory. The system can be reconfigured, for example by using a hypervisor, to match the request and expose the system as bare metal components to the system user as a rack-based component.
[0092] The present disclosure also enables rack owners to use their racks efficiently by not over-committing resources to any one user.
[0093] Reveal Core VF diagram
[0094] Manufacturing yield defects prevent some processing cores within a socket of a computing node from operating at the same maximum voltage as other processing cores. Typically, all cores in a socket are capped at the same maximum voltage and frequency to avoid damaging some of the processing cores.
[0095] Typically, the OS scheduler or hypervisor scheduler randomly selects a core to run a job. In addition, the POD manager (PODM) assumes that the maximum voltage and frequency for all cores in the core are the same. Therefore, as long as the PODM knows the maximum voltage and frequency, the PODM can calculate the overall performance of the socket. For example, if the core performance cp = V x F and there are 10 cores, the overall performance is 10x cp.
[0096] The POD Manager includes firmware and software application interfaces (APIs) that implement management resources and policies and expose standard interfaces to the hardware below the POD Manager and the orchestration layer above it. The POD Manager API allows for the use of Rack-Scale Architecture (RSA) system resources in a flexible manner and allows for integration with the ecosystem in which RSA is used. The POD Manager implements health monitoring and troubleshooting (e.g., fault location and isolation) as well as physical location features.
[0097] The implementation of the present disclosure collects the possible voltage and frequency of the operating state of each processor core in the BIOS, and provides this data to the PODM to effectively schedule the workload to obtain better performance. Therefore, in one implementation of the present disclosure, the P-state (or performance state) of the processing core is evaluated. By exposing the voltage and frequency of each processor core to the OS or hypervisor and the PODM, each core can perform processing operations at different voltages and / or frequencies to improve performance, rather than fixing the same voltage / frequency for operation performance to all cores.
[0098] The system is aware of the geometry of the processing cores and the location of the processing cores within the compute node. Therefore, some embodiments avoid overheating by splitting jobs among higher voltage processing cores. For example, instead of the OS scheduler selecting high voltage cores that are adjacent to each other, the OS scheduler can select cores that are located farther from each other to reduce thermal effects and thus produce thermal balance at the socket level.
[0099] Fig. 9 An example of the frequency of cores within a compute node for a rack according to one implementation of the present disclosure is shown. Fig. 9 As shown, the cores in the upper row can handle maximum frequencies ranging from 3.8 to 4.0 GHz. Conversely, the cores in the lower row can handle maximum frequencies ranging from 3.7 to 4.0 GHz. Therefore, in a conventional node, the cores operate at a maximum of 3.7 GHz. The reduced maximum limit prevents most of the cores in the core from operating at their full capacity.
[0100] Fig.10 An algorithm for assigning processing jobs to processing cores based on a VF graph is shown according to one implementation of the present disclosure.
[0101] At boot time, the processor executes the BIOS. The BIOS executed by the processor determines the position of each SLED in the drawer. The BIOS executed by the processor can determine these positions based on, for example, signals received from sensors that monitor each position of the SLED.
[0102] For each SLED, in S1110, the BIOS executed by the processor collects the voltage or frequency of the operating state allowed by each of the multiple processor cores of the node, the maximum transactions per second (TPCC) allowed by the processor core, and the location of the processing core. The BIOS sends this information to the orchestration layer.
[0103] The maximum frequency of each core is determined at manufacturing time, and the maximum setting is fused during manufacturing. A PCU (Power Control Unit) or similar power controller can read the fused setting and control power to each core. Typically, the performance state is exposed to the OS, and the OS can set the performance state of a single core by setting model specific registers (MSRs) in each core.
[0104] At S1120, the POD manager orchestration selects a computing node and pushes the workload to the computing node based on the first SLA or TPCC capability. However, the POD manager orchestration does not necessarily control which core in the socket will execute the workload.
[0105] At S1130, the OS or hypervisor within the selected computing node receives the first processing job and the first SLA corresponding to the first job. Therefore, the OS scheduler or hypervisor scheduler assigns the first processing job to the first core in the socket. Specifically, the OS scheduler or hypervisor scheduler determines to assign the first processing job to the first processing core based on the first SLA and the voltage or frequency allowed by each core. More specifically, the OS scheduler or hypervisor scheduler determines to assign the first processing job to the first core based on the duration (e.g., number of minutes) or TPCC in the SLA.
[0106] In addition, the OS or hypervisor attempts to minimize the duration and / or TPCC above that agreed to in the first SLA. For example, a processing job may require 75 million instructions per second. The first processing core is capable of processing 100 million transactions per second. The second processing core is capable of processing 80 million transactions per second. In this case, the OS or hypervisor selects the second processing core because its 5 million transactions per second is less than the first processing core's additional 20 million transactions per second.
[0107] Similarly, the first processing core can complete the job in 10 minutes. In contrast, the second processing core can complete the job in 12 minutes. In this case, the OS or hypervisor selects the first processing core because the first processing core can complete the job in a shorter time.
[0108] Of course, the OS or hypervisor can weigh TPCC against completion time. This trade-off can occur in many ways and is outside the scope of this disclosure.
[0109] When assigning the first processing job to the first core, the OS or hypervisor sets the performance state of the first core by setting the MSR in the corresponding core.
[0110] At S1140, as described above, the POD manager orchestration selects a computing node and pushes the workload based on the second SLA and TPCC capabilities. Likewise, the POD manager orchestration does not necessarily control which core in the socket executes the workload.
[0111] In S1150, the OS or hypervisor in the selected computing node receives the second processing job and the second SLA corresponding to the second processing job, and the computing node may be the same or different from the node selected for the first processing job. Of course, the second SLA may be the same SLA as the first SLA. Therefore, the OS scheduler or hypervisor scheduler assigns the second processing job to the second core based on the second SLA, the frequency of the second core, and the positions of the first core and the second core.
[0112] In particular, the OS scheduler or hypervisor scheduler can assign the second processing job to the second core based on similar SLA and voltage and frequency criteria as the OS or hypervisor assigned the first processing job to the first core. In addition, the OS scheduler or hypervisor scheduler can assign the second processing job to the second core based on the location of the first core and the second core. In particular, the OS scheduler schedules non-adjacent cores on the same die to run the workload, thereby avoiding overheating caused by running workloads on cores that are adjacent to each other on the same die.
[0113] When assigning the second processing job to the second core, the OS or hypervisor sets the performance state of the second core by setting the MSR in the corresponding core.
[0114] The algorithm then ends at S1160.
[0115] Some implementations of the present disclosure enable each core to individually advertise its maximum voltage and frequency of operation and the location of the core within the die. Therefore, some implementations of the present disclosure can give better performance than traditional sockets that are capped at the lowest frequency across all cores within the die.
[0116] The OS scheduler itself can sense this data and can run the cores at their respective maximum voltage or frequency, or make individual cores automatically run at their respective maximum voltage or frequency to obtain better performance. For example, the POD manager can calculate the overall performance as, for example, cpl+cp2+...+cp10 (=V1xF1+V2xF2+...+V10xF10). This approach can achieve additional performance and the POD manager can accurately determine the performance, rather than the traditional approach of capping voltage and frequency for all processors.
[0117] Additionally, the OS may be aware of the location of each core and may schedule jobs so that the OS evenly distributes the jobs to keep the temperature evenly distributed.
[0118] Therefore, benchmark data (e.g., TPC-C, TPC [Transaction Processing Performance Council] Benchmark E, TPCx-V, SPECVirt, SPECjExt) that describes the core VF information can be exposed to the OS or hypervisor. Thus, the OS or hypervisor can select the system to run the workload that can maximize overall performance and reduce TCO.
[0119] Revise
[0120] Many modifications of the teachings of this disclosure are possible. For example, Figure 2 S230 is described as occurring before S240. In some implementations of the present disclosure, S240 is performed before S230.
[0121] In an example embodiment, the circuit of the accompanying drawings can be implemented on a board of an electronic device. The board can be a general circuit board that holds various components of the internal electronic system of the electronic device, and further provides connectors for other peripheral devices. More specifically, the board can provide an electrical connection, and other components of the system can communicate electrically through the electrical connection. Processors (including digital signal processors, microprocessors, and supporting chipsets) and computer-readable non-transient memory elements can be coupled to the board based on configuration requirements, processing requirements, and computer design. Other components (such as external storage devices, additional sensors, controllers for audio / video display, and peripheral devices) can be attached to the board via a cable as a plug-in card or integrated into the board itself. In various embodiments, the functions described herein can be implemented in a simulated form as software or firmware running in one or more configurable (e.g., programmable) elements arranged in a structure that supports these simulated functions. The software or firmware providing simulation can be provided on a non-transient computer-readable storage medium including instructions for allowing the processor to perform those functions.
[0122] In another example embodiment, the circuit of the accompanying drawings can be implemented as a stand-alone module (e.g., a device having components and circuits for performing a specific application or function), or as an insert module of application-specific hardware to an electronic device. Specific embodiments of the present disclosure may be partially or completely included in a system-on-chip (SOC) package. SOC represents an IC that integrates the components of a computer or other electronic system into a single chip. The SOC may contain digital, analog, mixed signal, and generally radio frequency functions, all of which may be provided on a single chip substrate. Other embodiments may include a multi-chip module (MCM) having multiple independent ICs located within a single electronic package and closely interacting with each other through the electronic package. In various other embodiments, the digital filter may be implemented in one or more silicon cores in an application-specific integrated circuit (ASIC), a field programmable gate array (FPGA), and other semiconductor chips.
[0123] The provisions, dimensions and relationships (e.g., number of processors, logical operations) outlined herein are provided for purposes of example and instruction only. Without departing from the spirit of the present disclosure or the scope of the appended claims, this information may vary considerably. These provisions apply only to a non-limiting example, and therefore should be understood as such. In the foregoing description, example embodiments have been described with reference to specific processors and / or component arrangements. Various modifications and changes may be made to such embodiments without departing from the scope of the appended claims. Therefore, the description and drawings are considered to be illustrative rather than restrictive.
[0124] Utilizing many examples provided herein, interaction can be described in terms of two, three, four or more electronic components. However, this description is made only for the purpose of clarity and example. The system can be combined in any way. Connecting similar design alternatives, any of the components, modules and elements shown in the accompanying drawings can be combined in various possible configurations, all of which are clearly within the scope of the present disclosure. In some cases, it may be easier to describe one or more of the functions of a given set of processes by referring only to a limited number of electrical components. The circuits of the accompanying drawings and their teachings are easy to scale and can accommodate a large number of components and more difficult / complex arrangements and configurations. Therefore, the examples provided should not limit the scope of the circuit or limit the teachings of the circuit to potentially being applied to a large number of other architectures.
[0125] In the present disclosure, references to various features (e.g., elements, structures, modules, components, steps, operations, characteristics, etc.) included in “one implementation,” “an example implementation,” “implementation,” “another implementation,” “some implementations,” “various implementations,” “other implementations,” etc.) are intended to mean that any such features are included in one or more implementations of the present disclosure, but may or may not be combined in the same implementation.
[0126] Without departing from the scope of the present disclosure, some of the operations may be deleted or removed, or these operations may be modified or changed appropriately, where appropriate. In addition, the timing of these operations may be changed appropriately. For purposes of example and discussion, the aforementioned operation flow is provided. A large amount of flexibility is provided by the implementation described herein, because any suitable arrangement, time sequence, configuration, and timing mechanism may be provided without departing from the teachings of the present disclosure.
[0127] Other changes, substitutions, changes, substitutions and modifications may be ascertained by those skilled in the art, and the present disclosure includes all such changes, substitutions, changes, substitutions and modifications that fall within the scope of the claims. Optional features of the apparatus and methods described above may also be implemented with respect to the methods or processes described herein, and the details in the examples may be used anywhere in one or more implementations.
[0128] Example 1 is a device for performing event regulation, the device comprising: a plurality of resources; one or more sensors; and a controller for regulating at least one of the resources if the controller determines that a thermal event or a power event has occurred, wherein the regulation is based at least in part on one or more service level agreements (SLAs) associated with the resources.
[0129] In Example 2, the apparatus of Example 1 may optionally include the following feature: the regulated at least one resource is associated with a lower level of service, the lower level of service being lower than the level of service associated with the unregulated resource.
[0130] In Example 3, the apparatus described in any one of Examples 1-2 may optionally include the following feature: the controller is used to determine the capability of a computing node among the plurality of computing nodes that is assigned to the one or more SLAs.
[0131] In Example 4, the apparatus described in any one of Examples 1-3 may optionally include the following feature: the controller is used to determine the at least one resource of a computing node among multiple computing nodes in the area.
[0132] In Example 5, the apparatus described in any one of Examples 1-4 may optionally include the following feature: the controller is used to determine the margin space of the at least one resource that can be used for adjustment.
[0133] In Example 6, the apparatus described in any one of Examples 1-5 may optionally include the following feature: the controller is a BMC, a management controller, or part of an orchestration layer.
[0134] In Example 7, the apparatus of any one of Examples 1-6 may optionally include the following feature: the at least one resource is network bandwidth, a number of accesses to a memory, or a number of operations performed by a processor.
[0135] In Example 8, the apparatus of any one of Examples 1-7 may optionally include an orchestrator for tracking a plurality of resources working on tasks assigned to the one or more SLAs and grouping the plurality of resources into regions.
[0136] In Example 9, the apparatus described in any one of Examples 1-8 may optionally include the following feature: the adjusted resources are all in an area with a first SLA, the area with the first SLA is lower than resources in other areas associated with a second SLA, and the second SLA is higher than the first SLA.
[0137] In Example 10, the apparatus described in Example 9 may optionally include the following feature: before resources are adjusted in a region with a higher SLA, all resources in the region with the first SLA are adjusted.
[0138] In Example 11, the apparatus of any one of Examples 1-10 may optionally include the following feature: the apparatus is a computing system.
[0139] Example 12 is a device for performing event regulation, the device comprising: a plurality of resources; one or more sensors; and a unit for regulating at least one of the resources if the unit for regulating determines that a thermal event or a power event has occurred, wherein the regulation is based at least in part on one or more service level agreements (SLAs) associated with the resources.
[0140] In Example 13, the apparatus of Example 12 may optionally include the following feature: the regulated at least one resource is associated with a lower level of service, the lower level of service being lower than the level of service associated with the unregulated resource.
[0141] In Example 14, the apparatus of any one of Examples 12-13 may optionally include the following feature: the unit for adjusting determines a capability of a computing node of the plurality of computing nodes that is assigned to the one or more SLAs.
[0142] In Example 15, the apparatus of any one of Examples 12-14 may optionally include the following feature: the means for adjusting determines the at least one resource of a computing node among a plurality of computing nodes in the area.
[0143] In Example 16, the apparatus described in any one of Examples 12-15 may optionally include the following feature: the unit for adjusting determines a margin space of the at least one resource that can be used for adjustment.
[0144] In Example 17, the apparatus of any one of Examples 12-16 may optionally include the following feature: the unit for adjusting is a BMC, a management controller, or a part of an orchestration layer.
[0145] In Example 18, the apparatus of any one of Examples 12-17 may optionally include the following feature: the at least one resource is network bandwidth, a number of accesses to a memory, or a number of operations performed by a processor.
[0146] In Example 19, the apparatus of any one of Examples 12-18 may optionally include an orchestration unit for tracking the plurality of resources working on tasks assigned to the one or more SLAs and for grouping the plurality of resources into regions.
[0147] In Example 20, the apparatus described in any one of Examples 12-17 may optionally include the following feature: the adjusted resources are all in a region with a first SLA, the region with the first SLA is lower than resources in other regions associated with a second SLA, and the second SLA is higher than the first SLA.
[0148] In Example 21, the apparatus of any one of Examples 20 may optionally include the following feature: all resources in the region having the first SLA are adjusted before resources in the region having a higher SLA are adjusted.
[0149] In Example 22, the apparatus of any one of Examples 12-21 may optionally include the following feature: the apparatus is a computing system.
[0150] Example 23 is a method for performing event regulation, the method comprising: if a controller determines that a thermal event or a power event has occurred, regulating at least one of a plurality of resources, wherein the regulation is based at least in part on one or more service level agreements (SLAs) associated with the resource.
[0151] In Example 24, the method of Example 23 may optionally include the following feature: the regulated at least one resource is associated with a lower level of service, the lower level of service being lower than the level of service associated with the unregulated resource.
[0152] In Example 25, the method of any one of Examples 23-24 may optionally include determining a capability of a computing node of the plurality of computing nodes that is assigned to the one or more SLAs.
[0153] In Example 26, the method of any one of Examples 23-25 may optionally include determining at least one resource of a computing node among a plurality of computing nodes in a region.
[0154] In Example 27, the method of any one of Examples 23-26 may optionally include determining a headroom available for adjustment of the at least one resource.
[0155] In Example 28, the method of any one of Examples 23-27 may optionally include the following feature: the controller is a BMC, a management controller, or part of an orchestration layer.
[0156] In Example 29, the method of any one of Examples 23-28 may optionally include the following feature: the at least one resource is network bandwidth, a number of accesses to a memory, or a number of operations performed by a processor.
[0157] In Example 30, the method of any one of Examples 23-29 may optionally include tracking, with the orchestrator, the plurality of resources working on tasks assigned to the one or more SLAs; and grouping, with the orchestrator, the plurality of resources into regions.
[0158] In Example 31, the method described in any one of Examples 23-30 may optionally include the following feature: the adjusted resources are all in a region with a first SLA, the region with the first SLA is lower than resources in other regions associated with a second SLA, and the second SLA is higher than the first SLA.
[0159] In Example 32, the method of Example 31 may optionally include the following feature: before resources in a region with a higher SLA are adjusted, all resources in the region with the first SLA are adjusted.
[0160] Example 33 is a machine-readable medium comprising code, which, when executed, causes the machine to perform the method according to any one of Examples 23-32.
[0161] Example 34 is a device comprising a unit for performing the method according to any of 23-32.
[0162] In Example 35, the apparatus described in Example 34 may optionally include the following features: the unit for executing the method includes a processor and a memory.
[0163] In Example 36, the apparatus described in Example 35 may optionally include the following features: the memory includes machine-readable instructions, which, when executed, cause the apparatus to perform the method.
[0164] In Example 37, the apparatus of any one of Examples 34-36 may optionally include the following feature: the apparatus is a computing system.
[0165] Example 38 is at least one computer-readable medium comprising instructions that, when executed, implement the method according to any one of Examples 23-32, or implement the apparatus according to any one of Examples 34-37.
[0166] Example 39 is a non-transitory, tangible computer-readable storage medium encoded with instructions that, when executed, cause a processing unit to perform a method comprising: if the processing unit determines that a thermal event or a power event has occurred, adjusting at least one of a plurality of resources, wherein the adjustment is based at least in part on one or more service level agreements (SLAs) associated with the resources.
[0167] In Example 40, the medium of Example 39 may optionally include the following feature: the regulated at least one resource is associated with a lower level of service, the lower level of service being lower than the level of service associated with the unregulated resource.
[0168] In Example 41, the medium of any one of Examples 39-40 may optionally include the following feature: the method further includes: determining a capability of a computing node of the plurality of computing nodes that is assigned to the one or more SLAs.
[0169] In Example 42, the medium of any one of Examples 39-41 may optionally include the following feature: the method further includes: determining at least one resource of one of the plurality of computing nodes in the region.
[0170] In Example 43, the medium of any one of Examples 39-42 may optionally include the following feature: the method further includes: determining a headroom available for adjustment of the at least one resource.
[0171] In Example 44, the medium of any one of Examples 39-43 may optionally include the following feature: the processing unit is a BMC, a management controller, or part of an orchestration layer.
[0172] In Example 45, the medium of any one of Examples 39-44 may optionally include the following feature: the at least one resource is network bandwidth, a number of accesses to a memory, or a number of operations performed by a processor.
[0173] In Example 46, the medium of any one of Examples 39-45 may optionally include the following features: the method further includes: using an orchestrator to track the multiple resources working on tasks assigned to the one or more SLAs; and using the orchestrator to group the multiple resources into regions.
[0174] In Example 47, the medium of any one of Examples 39-46 may optionally include the following feature: the adjusted resources are all in a region with a first SLA, the region with the first SLA is lower than resources in other regions associated with a second SLA, and the second SLA is higher than the first SLA.
[0175] In Example 48, the medium of Example 47 may optionally include the following feature: before resources in a region with a higher SLA are adjusted, all resources in the region with the first SLA are adjusted.
[0176] Example 49 is an apparatus for subnode combination, comprising: a memory element operable to store electronic code; and a plurality of processing cores operable to execute instructions associated with the electronic code, wherein an orchestration layer or BMC receives a target number of domains for the plurality of cores and divides the plurality of computing cores into the target number of domains.
[0177] In Example 50, the apparatus of Example 49 may optionally include the following feature: the plurality of processing cores are used to communicate via a plurality of connections, and at least one of the plurality of connections is disabled to create one of the target number of domains.
[0178] In Example 51, the apparatus of any one of Examples 49-50 may optionally include the following feature: one of the plurality of processing units executes a hypervisor to create one of the target number of domains.
[0179] In Example 52, the apparatus of any one of Examples 49-51 may optionally include the following feature: each of the target number of domains simultaneously executes jobs from different service level agreements.
[0180] In Example 53, the apparatus of any one of Examples 49-52 may optionally include the following feature: one of the multiple processing cores includes an integrated memory controller, and the integrated memory controller supports two domains of the target number of domains executed by one of the multiple processing cores.
[0181] In Example 54, the apparatus of any one of Examples 49-53 may optionally include the following feature: the job is assigned to one of the target number of domains based at least in part on a service level agreement.
[0182] In Example 55, the apparatus of any one of Examples 49-54 may optionally include the following feature: a platform controller hub is simulated to support at least one of the target number of domains.
[0183] In Example 56, the apparatus of any one of Examples 49-55 may optionally include the following feature: the apparatus is a computing system.
[0184] Example 57 is a device for sub-node combination, comprising: an execution unit for executing instructions associated with an electronic code; and a processing unit for receiving a target number of domains for the execution unit and for dividing the execution unit into the target number of domains.
[0185] In Example 58, the apparatus of Example 57 may optionally include the following feature: the execution unit communicates via a plurality of connections, and at least one of the plurality of connections is disabled to create one of the target number of domains.
[0186] In Example 59, the apparatus of any one of Examples 57-58 may optionally include the following features: an execution unit executes a hypervisor to create one of the target number of domains.
[0187] In Example 60, the apparatus of any one of Examples 57-59 may optionally include the following feature: each of the target number of domains simultaneously executes jobs from different service level agreements.
[0188] In Example 61, the apparatus of any one of Examples 57-60 may optionally include the following feature: one of the execution units includes a unit for supporting two of the target number of domains executed by one of the execution units.
[0189] In Example 62, the apparatus of any one of Examples 57-61 may optionally include the following feature: the job is assigned to one of the target number of domains based at least in part on a service level agreement.
[0190] In Example 63, the apparatus of any one of Examples 57-62 may optionally include the following feature: a platform control hub is simulated to support at least one of the target number of domains.
[0191] In Example 64, the apparatus of any one of Examples 57-63 may optionally include the following feature: the apparatus is a computing system.
[0192] Example 65 is a method for subnode combination, comprising: receiving, by a BMC or an orchestration layer, a target number of domains for the plurality of processing cores; and dividing, by the BMC or the orchestration layer, the plurality of computing cores into the target number of domains.
[0193] In Example 66, the method of Example 65 may optionally include communicating with the plurality of processing cores via a plurality of connections; and disabling at least one of the plurality of connections to create one of the target number of domains.
[0194] In Example 67, the method of any one of Examples 65-66 may optionally include executing a hypervisor using one of the plurality of processing cores to create one of the target number of domains.
[0195] In Example 68, the method of any one of Examples 65-67 may optionally include the following feature: each of the target number of domains simultaneously executes jobs from different service level agreements.
[0196] In Example 69, the method of any one of Examples 65-68 may optionally include the following feature: one of the plurality of processing cores includes an integrated memory controller for supporting two of the target number of domains executed by the one of the plurality of processing cores.
[0197] In Example 70, the method of any one of Examples 65-69 may optionally include the following feature: based at least in part on a service level agreement, the job is assigned to one of the target number of domains.
[0198] In Example 71, the method of any one of Examples 65-70 may optionally include the following feature: a platform control hub is simulated for supporting at least one of the target number of domains.
[0199] Example 72 is a machine-readable medium comprising code, which, when executed, causes the machine to perform the method of any of Examples 65-71.
[0200] Example 73 is an apparatus comprising means for performing the method of any of Examples 65-71.
[0201] In Example 74, the apparatus described in Example 73 may optionally include the following features: the unit for executing the method includes a processor and a memory.
[0202] In Example 75, the apparatus described in Example 74 may optionally include the following features: the memory includes machine-readable instructions, which, when executed, cause the apparatus to perform the method.
[0203] In Example 76, the apparatus of any one of Examples 73-75 may optionally include the following feature: the apparatus is a computing system.
[0204] Example 77 is at least one computer-readable medium comprising instructions that, when executed, implement the method of any of Examples 65-71 or implement the apparatus of any of Examples 73-76.
[0205] Example 78 is a non-transitory, tangible computer-readable storage medium encoded with instructions that, when executed, cause a processing unit to perform a method comprising: receiving a target number of domains for a plurality of processing cores; and dividing a plurality of computing cores into the target number of domains.
[0206] In Example 79, the medium described in Example 78 may optionally include the following features: the method also includes communicating with the multiple processing cores via multiple connections; and disabling at least one of the multiple connections for creating one of the target number of domains.
[0207] In Example 80, the medium of any one of Examples 78-79 may optionally include the following feature: the method further includes: executing a hypervisor using one of the plurality of processing cores to create one of the target number of domains.
[0208] In Example 81, the medium of any one of Examples 78-80 may optionally include the following feature: each of the target number of domains simultaneously executes jobs from different service level agreements.
[0209] In Example 82, the medium of any one of Examples 78-81 may optionally include the following feature: one of the plurality of processing cores includes an integrated memory controller for supporting two of the target number of domains executed by one of the plurality of processing cores.
[0210] In Example 83, the medium of any one of Examples 78-82 may optionally include the following feature: the job is assigned to one of the target number of domains based at least in part on a service level agreement.
[0211] In Example 84, the medium of any one of Examples 78-83 may optionally include the following feature: a platform control hub is simulated to support at least one of the target number of domains.
[0212] Example 85 is a device for balanced processing, the device comprising: a memory element operable to store an electronic code; a first core operable to operate at a first voltage or frequency; a second core operable to operate at a second voltage or frequency different from the first voltage or frequency; and a processor operable to execute instructions associated with the electronic code to receive an indication of the first voltage or frequency and to receive an indication of the second voltage or frequency, wherein the first core performs a first processing operation at the first voltage or frequency and the second core performs a second processing operation at the second voltage or frequency.
[0213] In Example 86, the apparatus of Example 85 may optionally include the following feature: the processor is used to assign the first processing job to the first core based at least in part on the first voltage or frequency.
[0214] In Example 87, the apparatus of any one of Examples 85-86 may optionally include the following feature: the first core and the second core are in a same socket in a computing node of the apparatus.
[0215] In Example 88, the apparatus of any one of Examples 85-87 may optionally include the following feature: the apparatus includes a plurality of computing nodes, and the first core and the second core are in different computing nodes among the plurality of computing nodes.
[0216] In Example 89, the apparatus of any one of Examples 85-88 may optionally include the following features: the processor is used to receive a first position of the first core and a second position of the second core, and is used to assign the second processing job to the second core based at least in part on the first position and the second position.
[0217] In Example 90, the apparatus of any one of Examples 85-89 may optionally include the following features: the processor is used to assign the first processing job to the first core based at least in part on a first service level agreement, and is used to assign the second processing job to the second core based at least in part on a second service level agreement, and the first service level agreement limits a number of transactions per second that is greater than the number of transactions per second limited by the second service level agreement.
[0218] In Example 91, the apparatus of any one of Examples 85-90 may optionally include the following feature: the performance of the first core and the second core is calculated by summing the product of the first voltage and the first frequency and the product of the second voltage and the second frequency.
[0219] In Example 92, the apparatus of any one of Examples 85-91 may optionally include the following feature: a set of voltage and frequency pairs storing an operating state of each memory core.
[0220] In Example 93, the apparatus of Example 92 may optionally include the following features: the set of voltage and frequency pairs is stored in the BIOS, and the set of voltage and frequency pairs is provided to a POD manager that schedules workloads based at least in part on the set of voltage and frequency pairs.
[0221] In Example 94, the apparatus of any one of Examples 85-93 may optionally include the following feature: the apparatus is a computing system.
[0222] Example 95 is a device for balanced processing, the device comprising: a first unit for operating at a first voltage or frequency; a second unit for operating at a second voltage or frequency different from the first voltage or frequency; and a processing unit for receiving an indication of the first voltage or frequency and for receiving an indication of the second voltage or frequency, wherein the first unit performs a first processing operation at the first voltage or frequency and the second unit performs a second processing operation at the second voltage or frequency.
[0223] In Example 96, the apparatus of Example 95 may optionally include the following feature: the processing unit assigns the first processing job to the first unit at least in part based on the first voltage or frequency.
[0224] In Example 97, the apparatus of any one of Examples 95-96 may optionally include the following feature: the first unit and the second unit are in the same slot in a computing node of the apparatus.
[0225] In Example 98, the apparatus of any one of Examples 95-97 may optionally include the following feature: the apparatus includes a plurality of computing nodes, and the first unit and the second unit are in different computing nodes among the plurality of computing nodes.
[0226] In Example 99, the apparatus of any one of Examples 95-98 may optionally include the following feature: the processing unit receives a first position of the first unit and a second position of the second unit, and assigns the second processing job to the second unit based at least in part on the first position and the second position.
[0227] In Example 100, the apparatus described in any one of Examples 95-99 may optionally include the following features: the processing unit assigns the first processing job to the first unit at least in part based on a first service level agreement, and assigns the second processing job to the second unit at least in part based on a second service level agreement, and the first service level agreement limits a number of transactions per second that is greater than the number of transactions per second limited by the second service level agreement.
[0228] In Example 101, the apparatus of any one of Examples 95-100 may optionally include the following feature: the performance of the first unit and the second unit is calculated by summing the product of the first voltage and the first frequency and the product of the second voltage and the second frequency.
[0229] In Example 102, the apparatus of any one of Examples 95-101 may optionally include the following feature: a set of voltage and frequency pairs storing the operating state of each memory core.
[0230] In Example 103, the apparatus of any one of Examples 95-102 may optionally include the following features: the set of voltage and frequency pairs is stored in the BIOS, and the set of voltage and frequency pairs is provided to a POD manager that schedules workloads based at least in part on the set of voltage and frequency pairs.
[0231] In Example 104, the apparatus of any one of Examples 95-103 may optionally include the following feature: the apparatus is a computing system.
[0232] Example 105 is a method for balanced processing implemented by a rack, the method comprising: operating a first core at a first voltage or frequency; operating a second core at a second voltage or frequency different from the first voltage or frequency; receiving an indication of the first voltage or frequency; and receiving an indication of the second voltage or frequency, wherein the first core performs a first processing job at the first voltage and frequency, and the second core performs a second processing job at the second voltage or frequency.
[0233] In Example 106, the method of Example 105 may optionally include assigning the first processing job to the first core based at least in part on the first voltage or frequency.
[0234] In Example 107, the method of any one of Examples 105-106 may optionally include the following feature: the first core and the second core are in a same slot in a compute node of the rack.
[0235] In Example 108, the method of any one of Examples 105-107 may optionally include the following feature: the rack includes a plurality of computing nodes, and the first core and the second core are in different computing nodes among the plurality of computing nodes.
[0236] In Example 109, the method of any one of Examples 105-108 may optionally include receiving a first position of the first core and a second position of the second core; and assigning the second processing job to the second core based at least in part on the first position and the second position.
[0237] In Example 110, the method described in any of Examples 105-109 may optionally include assigning the first processing job to the first core based at least in part on a first service level agreement; and assigning the second processing job to the second core based at least in part on a second service level agreement, wherein the first service level agreement limits a number of transactions per second that is greater than the number of transactions per second limited by the second service level agreement.
[0238] In Example 111, the method of any one of Examples 105-110 may optionally include calculating the performance of the first core and the second core by summing the product of the first voltage and the first frequency and the product of the second voltage and the second frequency.
[0239] In Example 112, the method of any one of Examples 105-111 may optionally include the following feature: storing a set of voltage and frequency pairs of the operating state of each memory core.
[0240] In Example 113, the method described in Example 112 may optionally include the following features: the set of voltage and frequency pairs is stored in the BIOS, and the set of voltage and frequency pairs is provided to a POD manager that schedules workloads based at least in part on the set of voltage and frequency pairs.
[0241] Example 114 is a machine-readable medium comprising code, which, when executed, causes the machine to perform the method of any of Examples 105-113.
[0242] Example 115 is an apparatus comprising means for performing the method of any of Examples 105-113.
[0243] In Example 116, the apparatus described in Example 115 may optionally include the following features: the unit for performing the method includes a processor and a memory.
[0244] In Example 117, the apparatus described in Example 116 may optionally include the following features: the memory includes machine-readable instructions, which, when executed, cause the apparatus to perform the method.
[0245] In Example 118, the apparatus of any one of Examples 115-117 may optionally include the following feature: the apparatus is a computing system.
[0246] Example 119 is at least one computer-readable medium comprising instructions that, when executed, implement the method of any one of Examples 105-113 or implement the apparatus of any one of Examples 115-118.
[0247] Example 120 is a non-transitory, tangible computer-readable storage medium encoded with instructions that, when executed, cause a processing unit to perform a method comprising: operating a first core of a rack at a first voltage or frequency; operating a second core at a second voltage or frequency different from the first voltage or frequency; receiving an indication of the first voltage or frequency; and receiving an indication of the second voltage or frequency, wherein the first core performs a first processing job at the first voltage or frequency and the second core performs a second processing job at the second voltage or frequency.
[0248] In Example 121, the medium described in Example 120 may optionally include the following features: the method also includes: assigning the first processing job to the first core based at least in part on the first voltage or frequency.
[0249] In Example 122, the medium of any one of Examples 120-121 may optionally include the following feature: the first core and the second core are in a same slot in a compute node of the rack.
[0250] In Example 123, the medium of any one of Examples 120-122 may optionally include the following feature: the rack includes a plurality of computing nodes, and the first core and the second core are in different computing nodes among the plurality of computing nodes.
[0251] In Example 124, the medium described in any one of Examples 120-123 may optionally include the following features: the method also includes: receiving a first position of the first core and a second position of the second core; and assigning the second processing job to the second core at least in part based on the first position and the second position.
[0252] In Example 125, the medium described in any one of Examples 120-124 may optionally include the following features: the method also includes: assigning the first processing job to the first core at least in part based on a first service level agreement; and assigning the second processing job to the second core at least in part based on a second service level agreement, wherein the first service level agreement limits a number of transactions per second that is greater than the number of transactions per second limited by the second service level agreement.
[0253] In Example 126, the medium of any one of Examples 120-125 may optionally include the following feature: the method also includes: calculating the performance of the first core and the second core by summing the product of the first voltage and the first frequency and the product of the second voltage and the second frequency.
[0254] In Example 127, the medium of any one of Examples 120-126 may optionally include the following feature: a set of voltage and frequency pairs storing an operating state of each memory core.
[0255] In Example 128, the medium described in Example 127 may optionally include the following features: the set of voltage and frequency pairs is stored in the BIOS, and the set of voltage and frequency pairs is provided to a POD manager that schedules workloads based at least in part on the set of voltage and frequency pairs.
Claims
1. A rack system, include: a plurality of resources of a plurality of computing nodes in a region of the rack system; one or more sensors; as well as A controller for determining headroom available for adjustment of at least one resource of one of a plurality of computing nodes in the region and adjusting the at least one resource if it is determined that a thermal event or a power event has occurred in the region, wherein the adjustment is based at least in part on one or more service level agreements (SLAs) associated with the resources, and wherein the determination of the headroom includes determining whether the at least one resource exceeds the SLA associated with the at least one resource, wherein the thermal event includes an event in which a temperature in the region exceeds a first predetermined threshold, and the power event includes an event in which a voltage difference or resistance in the region becomes zero or a current or power exceeds a second predetermined threshold.
2. The rack system according to claim 1, in, The throttled at least one resource is associated with a lower level of service than a level of service associated with the unthrottled resource.
3. The rack system according to any one of claims 1 to 2, in, The controller is configured to determine a capability of a computing node of the plurality of computing nodes that is assigned to the one or more SLAs.
4. The rack system according to any one of claims 1 to 2, in, The controller is a BMC, a management controller, or part of an orchestration layer.
5. The rack system according to any one of claims 1 to 2, in, The at least one resource is network bandwidth, a number of accesses to memory, or a number of operations performed by a processor.
6. The rack system according to any one of claims 1 to 2, further comprising: include: An orchestrator is configured to track the plurality of resources working on tasks assigned to the one or more SLAs and group the plurality of resources into regions.
7. The rack system according to any one of claims 1 to 2, in, The adjusted resources are all in a region with a first SLA that is lower than resources in other regions associated with a second SLA that is higher than the first SLA.
8. The rack system according to claim 7, in, All resources in a region with the first SLA are adjusted before resources in a region with a higher SLA are adjusted.
9. A rack system, include: a plurality of resources of a plurality of computing nodes in a region of the rack system; one or more sensors; as well as A unit for determining a margin available for adjustment of at least one resource of one of a plurality of computing nodes in the region and adjusting the at least one resource if it is determined that a thermal event or a power event has occurred in the region, wherein the adjustment is based at least in part on one or more service level agreements (SLAs) associated with the resources, and wherein the determination of the margin includes determining whether the at least one resource exceeds the SLA associated with the at least one resource, wherein the thermal event includes an event in which the temperature in the region exceeds a first predetermined threshold, and the power event includes an event in which a voltage difference or resistance in the region becomes zero or a current or power exceeds a second predetermined threshold.
10. A method for performing event conditioning in a rack system, the method include: If it is determined that a thermal event or a power event has occurred in a region of the rack system including a plurality of computing nodes having a plurality of resources, a headroom available for adjustment of at least one resource of one of the plurality of computing nodes in the region is determined and the at least one resource is adjusted, wherein the adjustment is based at least in part on one or more service level agreements (SLAs) associated with the resources, and wherein the determination of the headroom includes determining whether the at least one resource exceeds the SLA associated with the at least one resource, wherein the thermal event includes an event in which a temperature in the region exceeds a first predetermined threshold, and the power event includes an event in which a voltage difference or resistance in the region becomes zero or a current or power exceeds a second predetermined threshold.
11. The method according to claim 10, in, The throttled at least one resource is associated with a lower level of service than a level of service associated with the unthrottled resource.
12. The method according to any one of claims 10-11, in, The at least one resource is network bandwidth, a number of accesses to memory, or a number of operations performed by a processor.
13. The method according to any one of claims 10 to 11, further comprising: include: tracking, with an orchestrator, the plurality of resources working on tasks assigned to the one or more SLAs; as well as The plurality of resources are grouped into zones using the orchestrator.
14. The method according to any one of claims 10-11, in, The adjusted resources are all in a region with a first SLA that is lower than resources in other regions associated with a second SLA that is higher than the first SLA.
15. A non-transitory, tangible computer-readable storage medium encoded with instructions that, when executed, cause a processing unit to perform the method of any one of claims 10-14.
16. A computer program product comprising instructions which, when executed, cause a processing unit to perform the method according to any one of claims 10-14.
17. A computing device, include: A memory for storing instructions; as well as A processor coupled to the memory, wherein the instructions, when executed by the processor, cause the processor to perform the method according to any one of claims 10-14.
