Method and node for managing requests for hardware acceleration by means of an accelerator device
Through the hardware acceleration management node dynamically selecting the hardware acceleration device, the problem of inflexible access of the acceleration device in the prior art is solved, and transparent access to multiple acceleration devices and efficient resource utilization is achieved.
Patent Information
- Application Number
- CN201880093252.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2018-05-08
- Publication Date
- 2024-11-08
- Estimated Expiration
- 2038-05-08
AI Technical Summary
In the prior art, the use of a remote acceleration device requires pre-configuration of a computing unit to be statically associated with a specific acceleration device, lacks flexibility and cannot dynamically and transparently access multiple hardware acceleration devices.
Through the hardware acceleration management node, hardware acceleration requests from the application are received and mediated, and the appropriate hardware acceleration device is dynamically selected for processing, achieving transparent access to multiple hardware acceleration devices.
Flexible access to multiple hardware acceleration devices is achieved, resource utilization efficiency is improved, and hardware acceleration devices are allowed to be shared among multiple applications, making acceleration services completely transparent to the computing unit.
Smart Images

Figure CN112041817B_ABST
Abstract
Description
Technical Field
[0001] Embodiments herein relate to computer systems, such as networks of computer resources, cloud systems, etc. In particular, methods and hardware acceleration management nodes for managing requests for hardware acceleration by means of at least one accelerator device managed by a hardware acceleration management node are disclosed. Corresponding computer programs and computer program carriers are also disclosed. Background Art
[0002] Hardware acceleration has been known for decades, for example with the advent of graphics processing units (GPUs). GPUs are useful in accelerating specialized tasks such as image processing of still pictures and / or video.
[0003] Additionally, more general-purpose GPUs have been introduced that help central processing units (CPUs) more efficiently handle heavy computing workloads, including video processing and other processing related to, for example, big data analytics.
[0004] As the next step in the development of hardware acceleration, field programmable gate arrays (FPGAs) have been introduced. FPGAs have attracted a lot of attention, for example, because they are an efficient way to accelerate computing tasks. FPGAs are configurable with the help of software, such as to adapt the FPGA's hardware to different computing tasks.
[0005] In recent years, the term "hardware accelerator" or "hardware acceleration device" has been used to refer to any hardware acceleration providing entity, such as a GPU, FPGA, etc. Similar to the initial introduction of GPUs, hardware acceleration devices can be used as co-processors, as full replacements for CPU processors, or as instruction set architecture (ISA) extensions in one or more CPUs.
[0006] In the past few years, several reconfigurable architectures have been proposed for the acceleration of cloud computing applications in data centers using FPGAs, and the interest is ramping up.
[0007] Recently, there has been a lot of work in attempting to enable remote use of acceleration devices, sometimes referred to as Acceleration as a Service (AaaS). Some available solutions for enabling flexible and reconfigurable use of acceleration devices are Disaggregated FPGAs, as of April 2018 (see for example http: / / ieeexplore.ieee.org / document / 7830659 / , https: / / sec.ch9.ms / sessions / c1f9c808- 82bc-480a-a930-b340097f6cc1 / build / 2017 / B8063.pptx and http: / / ieeexplore.ieee.org / stamp / stamp.jsp?tp=&arnumber=7783710 and Remote Compute Unified Device Architecture (rCUDA) http: / / www.rcuda.net / pub / SC12_SLC.pdf, where CUDA is an application programming interface (API) that applications can call to take advantage of acceleration.
[0008] These solutions using remote acceleration devices enable a conventional computing unit to remotely access one or more acceleration devices.
[0009] In a further example, US8250578 discloses how to pipeline a hardware accelerator of a computing system. A hardware address is associated with at least one processing unit (PU) or at least one logical partition (LPAR) of the computing system. A work request for the associated hardware accelerator address is received. Then, the work request for the hardware accelerator using the associated hardware accelerator address is queued. Summary of the invention
[0010] In view of the above, it is noted that existing solutions using one or more remote acceleration devices enable a computing device to remotely access the acceleration device. A disadvantage of these existing solutions is that in order to use the acceleration device(s), the computing unit is required to be pre-configured, i.e., have the correct driver configuration that statically associates the computing unit with one or more specific remote acceleration devices, pre-allocation of these devices, addressing, etc.
[0011] Therefore, one goal may be to enable access to remote acceleration devices in a more flexible manner.
[0012] According to an aspect, the object is achieved by a method performed by a hardware acceleration management node for managing requests for hardware acceleration "HA" by means of at least one hardware acceleration device among a plurality of hardware acceleration devices managed by the hardware acceleration management node.
[0013] The hardware acceleration management node receives a request for hardware acceleration of processing source data from an HA interfacing node. The source data is managed by an application that calls the HA interfacing node to initiate hardware acceleration of processing source data. The hardware acceleration management node sends an indication of a source memory location for storing source data to the HA interfacing node.
[0014] Furthermore, the hardware acceleration management node selects at least one hardware acceleration device from the plurality of hardware acceleration devices. The hardware acceleration management node receives a chunk of code to be accelerated from the HA interface connection node.
[0015] The hardware acceleration management node sends an indication of a source memory location and a set of acceleration instructions related to the code block to the one hardware acceleration device. The hardware acceleration management node receives an indication of a result memory location indicating result data generated by processing the source data from the hardware acceleration device.
[0016] Then, the hardware acceleration management node sends an indication of completed hardware acceleration to the HA interface connection node.
[0017] According to another aspect, the object is achieved by a hardware acceleration management node configured to manage requests for hardware acceleration "HA" by means of at least one hardware acceleration device among a plurality of hardware acceleration devices managed by the hardware acceleration management node.
[0018] The hardware acceleration management node is configured to receive a request for hardware acceleration of processing source data from the HA interface connection node. The source data is managed by an application that calls the HA interface connection node to initiate hardware acceleration of processing source data.
[0019] In addition, the hardware acceleration management node is configured to send an indication of a source memory location for storing source data to the HA interface connection node. The hardware acceleration management node is configured to select at least one hardware acceleration device from the plurality of hardware acceleration devices. The hardware acceleration management node is configured to receive a code block to be accelerated from the HA interface connection node. The hardware acceleration management node is configured to send an indication of a source memory location and a set of acceleration instructions related to the code block to the one hardware acceleration device.
[0020] Furthermore, the hardware acceleration management node is configured to receive from the hardware acceleration device an indication of a result memory location indicating result data generated by processing the source data. The hardware acceleration management node is configured to send an indication of completed hardware acceleration to the HA interface connection node.
[0021] According to a further aspect, the object is achieved by a computer program and a computer program carrier corresponding to the above aspect.
[0022] The hardware acceleration management node may be said to operate as a proxy function. This may mean that the hardware acceleration management node mediates requests from the HA interface connection node to the at least one selected hardware acceleration device, whereby the hardware acceleration management node enables the HA interface connection node to dynamically and transparently access the selected (one or more) hardware acceleration devices. Thus, the HA interface connection node requests acceleration without knowing which (or which) hardware acceleration devices will perform the acceleration.
[0023] This is in contrast to existing solutions which statically link the HA interfacing nodes to appropriate hardware acceleration devices, whereby an entity requesting acceleration directly requests acceleration to be performed by the statically linked appropriate hardware acceleration device, for example by means of an address to the appropriate hardware acceleration device.
[0024] Therefore, the embodiments herein enable the HA interface connection node to flexibly access the hardware acceleration device via the hardware acceleration management node.
[0025] Thus, an advantage is that multiple hardware acceleration devices can be shared among multiple HA interfaced nodes (e.g., applications, servers, operating systems, etc.) This advantage is achieved while being completely transparent to the applications / computing units as to which specific hardware acceleration device is used to provide acceleration. BRIEF DESCRIPTION OF THE DRAWINGS
[0026]
[0011] Various aspects of the embodiments disclosed herein, including particular features and advantages thereof, will be readily understood from the following detailed description and the accompanying drawings briefly described below.
[0027] Figure 1 is a schematic overview of an exemplary system in which embodiments herein may be implemented.
[0028] Figure 2 is a combined signaling and flow chart illustrating the methods herein.
[0029] Figure 3 is a block diagram illustrating an embodiment of a hardware accelerated management node. DETAILED DESCRIPTION
[0030] Throughout the following description, where applicable, like reference numerals have been used to denote like features, such as nodes, acts, modules, circuits, parts, items, elements, units, etc. In the figures, features that appear in some embodiments are indicated by dashed lines.
[0031] Figure 1 An exemplary hardware system 100 is depicted in which embodiments herein may be implemented. Hardware system 100 may be a so-called disaggregated hardware system, a cloud system, a data center, a network of computers, or the like.
[0032] The hardware system 100 may include a hardware acceleration (HA) management node 110. The HA management node 110 may be configured to statically or dynamically manage one or more hardware acceleration devices (HA devices) 120-122, such as GPUs, FPGAs, accelerator devices, etc. As previously described, the accelerator device may be exemplified by one or more ISA extensions executed on one or more CPUs. Further implementations of the accelerator device are known in the art.
[0033] The HA management node 110 may act as an interface for the HA interfacing node 130 towards the computing unit 140. The HA interfacing node 130 may handle the communication between the application 150 executed on the computing unit 140 and the HA management node 110. The HA interfacing node 130 may be a so-called simulator simulating a physical hardware accelerator device, or the HA interfacing node 130 may include a so-called accelerator library including features of embodiments herein.
[0034] In general, the hardware system 100 may rely on the principle of hardware resource decomposition. The principle of hardware decomposition regards central processing units (CPUs), memory, and network resources as independent and modular components. Resources are usually organized in pools such as blades, sleds, etc. That is, there are pools of CPU units and pools of memory units. Resources may include many other different types of pools, such as storage pools, pools of network interfaces, etc., but throughout this disclosure, due to relevance, the pools of CPUs and pools of memory units are discussed. The terms "CPU pool", "memory pool", etc. will be used herein to refer to the pools of CPUs and the pools of memory units. When utilizing the hardware system, the host may be composed of a subset of units / resources in one or more pools. The host may be assembled by selecting a subset of resources in one or more pools. Each host typically includes at least one computing unit 140. In this way, multiple resources are allocated to the host. In the relevant literature, the host may be referred to by the terms "logical hardware machine", "virtual hardware machine", etc.
[0035] Furthermore, the system 100 includes a memory 160, which is accessible by the HA devices 120-122 as well as the application 150 and / or the computing unit 140. The memory 160 may be any kind of data storage medium, such as a hard disk drive (HDD), a solid state drive, a volatile or non-volatile memory unit, a random access memory (RAM), etc. The memory 160 may be used to store information to be processed by one or more HA devices 120-122 and / or to store information resulting from processing by one or more HA devices 120-122.
[0036] In addition, the system 100 may include a binary repository 170, such as a database, etc. The binary repository 170 may include one or more different representations of the process desired to be accelerated. In this way, the process to be accelerated may be executed on multiple different types of accelerator devices, which may require different formats for execution. For example, code in the Open Computing Language (OpenCL) may be compiled to adapt to different types of accelerator devices. The application 150 may provide different representations to the binary repository 170.
[0037] Figure 1The entity can be at least Figure 1 , for example, via an application programming interface (API) 180 .
[0038] Now that it has been described Figure 1 The entity shown in the figure, then the following provides Figure 2 An exemplary overview of the process. The HA management node 110 may be responsible for assigning a memory area to an application based on the application's request using, for example, cudaMalloc(). In addition, the HA management node 110 may receive a trigger from an application 150 to run an acceleration task. In the trigger, the application 150 may also tell the HA management node 110 who will consume the results of the work, for example, it may be the application 150 itself that triggers the start of the work, or it may be another application (not shown). In addition, if necessary, the HA management node may select the corresponding code binary from the binary repository and load it to the device, or forward instructions from the computing unit 140 and trigger the start of the work. The HA management node 110 may also notify the application 150 (or the other application) that will use the results of the accelerated work of the end of the work, and refer to the memory range where the result data is stored.
[0039] Figure 2 Shows that when Figure 1 An exemplary method according to embodiments herein when implemented in the system 100 .
[0040] The HA management node 110 performs a method for managing a request for hardware acceleration by at least one HA device 120, 121, 122 among a plurality of hardware acceleration devices 120-122 managed by the HA management node 110. The disaggregated hardware system 100 may include a plurality of HA devices 120-122.
[0041] One or more of the following actions may be performed in any suitable order.
[0042] Action A010
[0043] Initially, the application 150 may call the HA interface connection node 130 in order to obtain acceleration of a code block for processing source data. The call may be exemplified by the application 150 sending one or more instructions, etc. to the HA interface connection node 130. However, this is performed according to a known manner while, for example, utilizing commands known from the CUDA API (like cudaMalloc, etc.) mentioned in the background section.
[0044] In some examples, application 150 may further specify a result memory location where the processing results of the source data are to be stored. This may mean that the instruction may further specify a result memory location. The result memory location may point to one or more locations within memory 160.
[0045] In response to the call from the application 150 , the HA interface connection node 130 sends a request for hardware acceleration of processing source data to the hardware acceleration management node 110 .
[0046] As described above, the result storage location may be specified by the application 150. Thus, the request sent to the HA management node 110 may include an indication of the result storage location. In this manner, the request may provide information about the consumer of the result, such as based on the identity of the consumer, the result storage location from which the consumer may retrieve the result, and the like.
[0047] Action A020
[0048] Following action A010, the HA management node 110 receives a request for hardware acceleration of processing source data from the HA interface connection node 130. The source data is managed by an application that calls the HA interface connection node 130 to initiate hardware acceleration of processing source data.
[0049] Action A030
[0050] In order to let HA interface connection node 130 know where the source data to be processed by the HA device to be selected in the following action A040 is stored, HA management node 110 sends an indication of the source memory location for storing the source data to HA interface connection node 130. The source memory location may point to one or more locations within memory 160.
[0051] Sometimes, such as when a user has not yet been specified, a result memory location can be derived from a source memory location, such as by the HA management node 110. As an example, the result memory location can be the same as the source memory location, directly after the source memory location, or after a similar location is found to be suitable for any particular use case. In a further example, the HA management node 110 determines the result memory location, such as without considering the source memory location. Typically, the HA management node 110 finds an available memory area and determines, for example, that the starting point of the found memory area is the result memory location.
[0052] When a user has been specified, it is possible that the application 150 and / or the HA interfacing node 130 has derived the result memory location from the source memory location in a manner similar to that mentioned directly above.
[0053] Action A040
[0054] Since the HA management node 110 has received the request in action A020, the HA management node 110 selects at least one HA device 120, 121, 122 from the plurality of HA devices 120-122. In this way, the HA management node 110 can select at least one available HA device from the plurality of HA devices 120-122. In this context, available may mean that the currently selected HA device 120, 121, 122 is not assigned to perform the accelerated task.
[0055] Action A050
[0056] Following action A030, the HA interface connection node 130 may receive an indication of a source memory location. Upon receiving the indication of the source memory location, the HA interface connection node 130 may inform the application 150 of the source memory location. In response, the application 150 may store the source data at the source memory location indicated by the indication of the source memory location. As an example, the application 150 may use cudaMemcpy() known from the CUDA API.
[0057] The indication of the source memory location may be a physical or virtual address.
[0058] Action A060
[0059] In action A050, the source memory location is informed to the application 150. In order to execute the code block to be accelerated, the application 150 may provide the code block to the HA interfacing node 130.
[0060] In response thereto, the HA interface connection node 130 may transmit the code block to the HA management node 110 .
[0061] Additionally, it is possible that a code block is accompanied by some metadata, such as:
[0062] binary_type = OpenCL, open accelerator (openACC), CUDA,
[0063] Invocation = synchronous or asynchronous
[0064] Priority = high, medium, ...
[0065] version = <version of function (func)>,
[0066] wait.
[0067] binary_type specifies the binary format, which can be OpenCL, openACC, CUDA, etc.
[0068] The call may specify whether to call the acceleration synchronously or asynchronously, where synchronous and asynchronous have their conventional meanings in the field of computer programming, i.e., synchronous means that the processing of the computing unit may wait for the acceleration of the code to complete before continuing to execute other code, and asynchronous execution allows for processing other code while the acceleration is executing.
[0069] The priority levels may be used by the HA management node 110 to prioritize between different code blocks.
[0070] The version may be a number indicating the version of the code block in order to enable the HA management node 110 to find a suitable binary from the binary repository 170 , for example.
[0071] Action A070
[0072] Following action A060 , the HA management node 110 receives the code block to be accelerated from the HA interface connection node 130 .
[0073] The code block may be an intermediate representation or a final representation of the code. As an example, when the code block is a final representation, the HA devices 120, 121, 122 may use the code block directly. As an alternative example, the HA devices 120, 121, 122 may be required to transform (such as translate, compile, recompile, etc.) the code block so that it can be converted into a format executable by the HA devices 120, 121, 122.
[0074] In some examples, the code block may be a request for one of the functions stored in binary repository 170. Figure 1 With respect to the description of binary repository 170, it may be noted that a code block may represent a process that is desired to be accelerated.
[0075] Action A080
[0076] The HA management node 110 sends an indication of the source memory location(s) and a set of acceleration instructions related to the code block to the one HA device 120 , 121 , 122 .
[0077] In this manner, the HA management node 110 may be able to enable processing of source data stored at a source memory location by the one HA device 120 , 121 , 122 according to a set of acceleration instructions.
[0078] In some cases, the set of acceleration instructions may include a code block (i.e., no translation or compilation is required) so that execution of the code block by the one HA device 120, 121, 122 may be enabled. Thus, in some examples, the set of acceleration instructions may be a code block. In such examples, the terms "set of acceleration instructions" and "code block" may be interchangeable.
[0079] Action A090
[0080] Following action A080, the HA device 120 may receive a set of acceleration instructions or simply referred to as "acc.instr." as shown in the figure.
[0081] The set of acceleration instructions may be accompanied by an indication of the source memory location(s). This may mean that the HA management node 110 may send an indication of the source memory location to the selected HA device 120, 121, 122. Thus, the HA device 120 is made aware of both the code block to be executed and the source data to be processed when executing the code block.
[0082] At this stage, the HA device 120 may process the source data by executing a set of acceleration instructions. When completed, the following action A100 is performed.
[0083] Action A100
[0084] Thus, when HA device 120 has completed processing the source data by executing the set of acceleration instructions, HA device 120 may send an indication of the result memory location(s) indicating the result data. Thus, the result data is generated by the processing of the source data.
[0085] Action A110
[0086] Following action A100, the HA management node 110 receives an indication of a result memory location from said HA device 120, 121, 122. The indication of the result memory location may be a physical or virtual memory address.
[0087] Action A120
[0088] In order to let the HA interfacing node 130 and possibly also the application 150 know that the processing of the source data has been completed, the HA management node 110 sends an indication of the completed hardware acceleration to the HA interfacing node 130 .
[0089] In some examples, the indication of completed hardware acceleration may include a message.
[0090] When the result memory location has been determined or derived by the HA management node 110, the message may further include an indication of the result memory location. The indication of the result memory location may also be sent separately from the indication of the completed hardware acceleration. In this way, the HA interface connection node 130 may know where to find the results of the processing. Finally, the result memory location may also be notified to the application 150 by the HA interface connection node 130.
[0091] In this manner, HA management node 110 enables remote hardware acceleration as a service (ie, AaaS) in a manner that is truly transparent to application 150 and / or computing unit 140. In other words, there is no "hard wiring" between application 150 requesting acceleration and HA devices 120-122.
[0092] From the perspective of the application 150, the application 150 may interact with the HA interface connection node 130 without being aware that the HA interface connection node implements acceleration by using the HA management node 110 providing AaaS. This means that the computing unit 140 and / or the application 150 are unaware of which specific HA device 120, 121, 122 provides acceleration.
[0093] Furthermore, the HA management node 110 may be able to achieve more efficient resource utilization, for example, because the HA devices 120 - 122 may be shared among many applications / servers / operating systems.
[0094] As mentioned in the background section, US8250578 directly assigns a processor to a specific accelerator device within an accelerator field by means of the physical address of the specific accelerator device. Thus, the processor communicates directly with the assigned accelerator. When a specific accelerator is occupied and the processor attempts to access it, it will be placed in a queue.
[0095] refer to Figure 3 , showing Figure 1 A schematic block diagram of an embodiment of the HA management node 110 is shown.
[0096] The HA management node 110 may include a processing unit 301, such as a component for performing the methods described herein. The component may be embodied in the form of one or more hardware units and / or one or more software units. Therefore, according to various embodiments described below, the term "unit" may refer to a circuit, a software block, etc.
[0097] The HA management node 110 may further include a memory 302. The memory may include (such as contain or store) instructions (eg, in the form of a computer program 303), which may include computer readable code means.
[0098] According to some embodiments herein, the HA management node 110 and / or the processing unit 301 include a processing circuit 304 as an exemplary hardware unit, and the processing circuit 304 may include one or more processors. Therefore, the processing unit 301 may be embodied in the form of the processing circuit 304 or may be 'implemented' by the processing circuit 304. The instructions may be executable by the processing circuit 304, so that the HA management node 110 may be operated to perform Figure 2As another example, the instructions, when executed by the HA management node 110 and / or the processing circuit 304, may cause the HA management node 110 to perform the following steps: Figure 2 method.
[0099] In view of the above, in one example, there may be provided an HA management node 110 configured to manage requests for hardware acceleration by at least one HA device 120, 121, 122 in a plurality of hardware acceleration devices 120-122 managed by the HA management node 110. As described above, the decomposed hardware system 100 may include a plurality of HA devices 120-122. Again, the memory 302 contains instructions executable by the processing circuit 304, whereby the HA management node 110 may be operable to:
[0100] receiving a request for hardware acceleration of processing source data from the HA interface connection node 130, wherein the source data is managed by an application that calls the HA interface connection node 130 to initiate hardware acceleration of processing source data,
[0101] sending an indication of a source memory location for storing source data to the HA interface connection node 130,
[0102] Selecting one HA device 120, 121, 122 from the plurality of HA devices 120-122,
[0103] receiving a code block to be accelerated from the HA interface connection node 130,
[0104] sending an indication of the source memory location and a set of acceleration instructions related to the code block to the one HA device 120, 121, 122,
[0105] receiving from said HA device 120, 121, 122 an indication of a result memory location indicating result data generated by processing source data, and
[0106] An indication of completed hardware acceleration is sent to the HA interface connection node 130 .
[0107] Figure 3 Further shown is a carrier 305 or program carrier providing (such as comprising, mediating, supplying, etc.) a computer program 303 as described directly above. The carrier 305 may be one of an electronic signal, an optical signal, a radio signal, and a computer readable medium.
[0108] In some embodiments, the HA management node 110 and / or the processing unit 301 may include one or more of the receiving unit 310, the sending unit 320, and the selecting unit 330 as exemplary hardware units. When the term "unit" refers to a hardware unit, the term "unit" may refer to a circuit. In other examples, one or more of the exemplary hardware units described above may be implemented as one or more software units.
[0109] Furthermore, the HA management node 110 and / or the processing circuit 301 may include an input / output (I / O) circuit 306 which may be instantiated by a receiving unit and / or a transmitting unit, as applicable.
[0110] Therefore, the HA management node 110 is configured to manage requests for hardware acceleration "HA" by at least one HA device 120, 121, 122 among the plurality of hardware acceleration devices 120-122 managed by the HA management node 110. The disaggregated hardware system 100 may include a plurality of HA devices 120-122.
[0111] Therefore, according to various embodiments described above, the HA management node 110 and / or the processing circuit 304 and / or the processing unit 301 and / or the receiving unit 310 are configured to receive a request for hardware acceleration of processing source data from the HA interface connection node 130. The source data is managed by an application that calls the HA interface connection node 130 to initiate hardware acceleration of processing source data. The request may include an indication of a result memory location. The result memory location may be derived from the source memory location.
[0112] The HA management node 110 and / or the processing circuit 304 and / or the processing unit 301 and / or the sending unit 320 are configured to send an indication of a source memory location for storing source data to the HA interfacing node 130 .
[0113] The HA management node 110 and / or the processing circuit 304 and / or the processing unit 301 and / or the selection unit 330 are configured to select at least one HA device 120 , 121 , 122 from the plurality of HA devices 120 - 122 .
[0114] The HA management node 110 and / or the processing circuit 304 and / or the processing unit 301 and / or the receiving unit 310 are configured to receive a code block to be accelerated from the HA interface connection node 130 .
[0115] The HA management node 110 and / or the processing circuit 304 and / or the processing unit 301 and / or the sending unit 320 are configured to send an indication of the source memory location and a set of acceleration instructions related to the code block to the one HA device 120 , 121 , 122 .
[0116] The HA management node 110 and / or the processing circuit 304 and / or the processing unit 301 and / or the receiving unit 310 are configured to receive, from the HA devices 120 , 121 , 122 , an indication of a result memory location indicating result data generated by processing source data.
[0117] The HA management node 110 and / or the processing circuit 304 and / or the processing unit 301 and / or the sending unit 320 are configured to send an indication of the completed hardware acceleration to the HA interfacing node 130 .
[0118] As used herein, the term "node" or "network node" may refer to one or more physical entities, such as devices, equipment, computers, servers, etc. This may mean that the embodiments herein may be implemented in one physical entity. Alternatively, the embodiments herein may be implemented in multiple physical entities (such as an arrangement including the one or more physical entities), that is, the embodiments may be implemented in a distributed manner, such as on a cloud system that may include a collection of server machines. In the case of a cloud system, the term "node" may refer to a virtual machine, such as a container, a virtual runtime environment, etc. A virtual machine may be assembled from hardware resources such as memory, processing, network, and storage resources that may reside in different physical machines (e.g., in different computers).
[0119] As used herein, the term "unit" may refer to one or more functional units, each of which may be implemented as one or more hardware units and / or one or more software units and / or combined software / hardware units in a node. In some examples, a unit may represent a functional unit implemented as software and / or hardware of a node.
[0120] As used herein, the term "computer program carrier", "program carrier" or "carrier" may refer to one of an electronic signal, an optical signal, a radio signal, and a computer-readable medium. In some examples, the computer program carrier may exclude transitory propagating signals such as electronic, optical and / or radio signals. Thus, in these examples, the computer program carrier may be a non-transitory carrier, such as a non-transitory computer-readable medium.
[0121] As used herein, the term "processing unit" may include one or more hardware units, one or more software units, or a combination thereof. Any such unit, whether hardware, software, or a combined hardware-software unit, may be a determining component, an estimating component, a capturing component, an associating component, a comparing component, an identifying component, a selecting component, a receiving component, a sending component, etc. as disclosed herein. As an example, the expression "component" may be a unit corresponding to the units listed above in conjunction with the figure.
[0122] As used herein, the term "software unit" may refer to a software application, a dynamic link library (DLL), a software component, a software object, an object according to the Component Object Model (COM), a software function, a software engine, an executable binary software file, etc.
[0123] The term "processing unit" or "processing circuit" herein may encompass a processing unit including, for example, one or more processors, an application specific integrated circuit (ASIC), a field programmable gate array (FPGA), etc. A processing circuit, etc. may include one or more processor cores.
[0124] As used herein, the expression “configured to / for” may mean that the processing circuit is configured to (such as adapted to or operable to) perform one or more actions described herein by means of software configuration and / or hardware configuration.
[0125] As used herein, the term "action" may refer to an action, step, operation, response, reaction, activity, etc. It should be noted that, if applicable, an action herein may be split into two or more sub-actions. In addition, also if applicable, it should be noted that two or more actions described herein may be combined into a single action.
[0126] As used herein, the term "memory" may refer to a hard disk, a magnetic storage medium, a portable computer floppy disk or disk, flash memory, random access memory (RAM), etc. In addition, the term "memory" may refer to an internal register memory of a processor, etc.
[0127] As used herein, the term "computer-readable medium" may be a Universal Serial Bus (USB) memory, a Digital Versatile Disk (DVD), a Blu-ray Disc, a software unit received as a data stream, a flash memory, a hard drive, a memory card (such as MemoryStick), a MultiMedia Card (MMC), a Secure Digital (SD) card, etc. One or more of the above examples of computer-readable media may be provided as one or more computer program products.
[0128] As used herein, the term "computer readable code unit" may be the text of a computer program, an entire binary file representing the computer program in a compiled format or a portion thereof, or anything in between.
[0129] As used herein, the expressions "transmit" and "send" are considered interchangeable. These expressions include transmissions by broadcast, unicast, multicast, etc. In this context, a transmission by broadcast can be received and decoded by any authorized device within range. In the case of unicast, one specifically addressed device can receive and decode the transmission. In the case of multicast, a group of specifically addressed devices can receive and decode the transmission.
[0130] As used herein, the terms "quantity" and / or "value" may be any kind of number, such as binary, real, imaginary or rational numbers, etc. In addition, "quantity" and / or "value" may be one or more characters, such as letters or strings of letters. "Quantity" and / or "value" may also be represented by a bit string (i.e., 0 and / or 1).
[0131] As used herein, unless otherwise apparent from the context, the terms "first", "second", "third", etc. may be used only to distinguish features, devices, elements, units, etc. from each other.
[0132] As used herein, the term "subsequent action" may refer to an action that is performed after a previous action, while additional actions may or may not be performed before the one action, but after the previous action.
[0133] As used herein, the term "a set of" may refer to one or more of something. For example, according to embodiments herein, a set of devices may refer to one or more devices, a set of parameters may refer to one or more parameters, and the like.
[0134] As used herein, the expression "in some embodiments" has been used to indicate that features of the described embodiment can be combined with any other embodiments disclosed herein.
[0135] Although various embodiments have been described, many different alterations, modifications and the like will become apparent to those skilled in the art. Therefore, the described embodiments are not intended to limit the scope of the present disclosure.
Claims
1. A method for managing a request for hardware acceleration "HA" by means of at least one hardware acceleration device (120, 121, 122) among a plurality of hardware acceleration devices (120-122) managed by the hardware acceleration management node (110), the method comprising: receiving (A020) a request for hardware acceleration of processing source data from a HA interface connection node (130), wherein the source data is managed by an application that calls the HA interface connection node (130) to initiate the hardware acceleration of processing the source data, sending (A030) to the HA interface connection node (130) an indication of a source memory location for storing the source data, selecting (A040) at least one hardware acceleration device (120, 121, 122) from the plurality of hardware acceleration devices (120-122), receiving (A070) a code block to be accelerated from the HA interface connection node (130), sending (A080) the indication of the source memory location and a set of acceleration instructions related to the code block to the at least one hardware acceleration device (120, 121, 122), receiving (A110) from the at least one hardware acceleration device (120, 121, 122) an indication of a result memory location indicating result data generated by processing the source data, and sending (A120) an indication of the completed hardware acceleration to the HA interface connection node (130), At least one selected hardware acceleration device is transparent to the application. 2 . The method of claim 1 , wherein the request includes the indication of the result memory location.
3. The method of any of the preceding claims, wherein the result memory location is derived from the source memory location.
4. A hardware acceleration management node (110) configured to manage a request for hardware acceleration "HA" by means of at least one hardware acceleration device (120, 121, 122) among a plurality of hardware acceleration devices (120-122) managed by the hardware acceleration management node (110), wherein the hardware acceleration management node (110) is configured to: receiving a request for hardware acceleration of processing source data from a HA interface connection node (130), wherein the source data is managed by an application that calls the HA interface connection node (130) to initiate the hardware acceleration of processing the source data, sending an indication of a source memory location for storing the source data to the HA interface connection node (130), selecting at least one hardware acceleration device (120, 121, 122) from the plurality of hardware acceleration devices (120-122), receiving a code block to be accelerated from the HA interface connection node (130), sending the indication of the source memory location and a set of acceleration instructions related to the code block to the at least one hardware acceleration device (120, 121, 122), receiving from the at least one hardware acceleration device (120, 121, 122) an indication of a result memory location indicating result data generated by processing the source data, and sending an indication of completed hardware acceleration to the HA interface connection node (130), At least one selected hardware acceleration device is transparent to the application.
5. The hardware acceleration management node (110) of claim 4, wherein the request includes the indication of the result memory location.
6. The hardware acceleration management node (110) according to any of claims 4-5, wherein the result memory location is derived from the source memory location.
7. A computer readable medium having stored thereon a computer program (303) comprising computer readable code means which, when executed on a network node (110), causes the network node (110) to perform the method according to any one of claims 1-3.
8. A computer program product having stored thereon a computer program (303) comprising computer readable code means which, when executed on a network node (110), causes the network node (110) to perform the method according to any one of claims 1 to 3.
Citation Information
Patent Citations
Pipelining hardware accelerators to computer systems
US8250578B2
Acceleration management node, acceleration node, client and method
CN105357258A
Virtualization of hardware accelerator
CN105579959A
Data processing system having a hardware acceleration plane and a software plane
CN107750357A
Apparatus and method for selectable hardware accelerators in a data driven architecture
US20040257370A1
