Notice of Pre-AIA or AIA Status
The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA .
Claim Objections
Claims 4-6 are objected to because of the following informalities:
Claims 4 and 5 disclose “wherein the dividing the service task into computing tasks and distributing the computing tasks to the target accelerator… and not being occupied comprises:”. However, parent Claim 1 recites this step as distributing to a target accelerator “which is idle.” Because the dependent claims are referencing the step from the parent claim, it is suggested to amend “not being occupied” to “which is idle” to ensure consistent terminology and proper antecedent basis.
Claim 6 discloses “or the direct-connection accelerator” (line 12). This should be amended to “direct-connected accelerator” for consistency with the terminology used throughout the claims.
Claim Rejections - 35 USC § 112
The following is a quotation of 35 U.S.C. 112(b):
(B) CONCLUSION.—The specification shall conclude with one or more claims particularly pointing out and distinctly claiming the subject matter which the inventor or a joint inventor regards as the invention.
The following is a quotation of pre-AIA 35 U.S.C. 112, second paragraph:
The specification shall conclude with one or more claims particularly pointing out and distinctly claiming the subject matter which the applicant regards as his invention.
Claim 22 is rejected under 35 U.S.C. 112(b) or pre-AIA 35 U.S.C. 112, second paragraph, as being indefinite for failing to particularly point out and distinctly claim the subject matter which the inventor or a joint inventor, or for pre-AIA the applicant regards as the invention.
Claim 22 discloses “and/or the two accelerators have the same application computing logic.” The use of “and/or” is alternative language that renders the exact boundaries and scope of the claim unclear as it simultaneously claims multiple different alternative combinations. For the purpose of examination, the claim is interpreted as requiring at least one of the listed conditions.
Claim Rejections - 35 USC § 103
In the event the determination of the status of the application as subject to AIA 35 U.S.C. 102 and 103 (or as subject to pre-AIA 35 U.S.C. 102 and 103) is incorrect, any correction of the statutory basis for the rejection will not be considered a new ground of rejection if the prior art relied upon, and the rationale supporting the rejection, would be the same under either status.
The following is a quotation of 35 U.S.C. 103 which forms the basis for all obviousness rejections set forth in this Office action:
A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102 of this title, if the differences between the claimed invention and the prior art are such that the claimed invention as a whole would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to which the claimed invention pertains. Patentability shall not be negated by the manner in which the invention was made
Claims 1, 2, 4, 6, 18, and 22-24 are rejected under 35 U.S.C. 103 as being unpatentable over Kundu (US 2022/0276914, hereinafter Kundu) in view of Thyagaturu (US 2022/0075655, hereinafter Thyagaturu).
Regarding claim 1, Kundu discloses
A distributed computing method, applied to a controller of a distributed accelerator cluster, the distributed computing method comprising (fig. 1-53): (paragraph [0068]: A logical device distributes workloads to a plurality of accelerators; paragraph [0076]: application 302 uses a logical device 304 as an interface to a plurality of accelerators; Note: Kundu’s “logical device” corresponds to the claimed “controller,” and the “plurality of accelerators” corresponds to the “distributed accelerator cluster”);
acquiring information of accelerators in the distributed accelerator cluster; (paragraph [0082]: logical device 610 examines workflow 604 and determines, based at least in part on knowledge of available acceleration resources and acceleration profiles; Note: Kundu’s teaching of acquiring knowledge of available acceleration resources and their acceleration profiles corresponds to acquiring information of the accelerators);
establishing an accelerator direct-connection pair according to the information of the accelerators; and (paragraph [0079]: logical device 404 utilizes physical communication link between first accelerator 408 and second accelerator 412; Note: Kundu’s logical device utilizing a physical communication link to pair a first and second accelerator together based on their acceleration profiles/information corresponds to establishing an accelerator direct-connection pair);
in response to receiving a service task, dividing the service task into computing tasks and distributing the computing tasks to a target accelerator which is idle and has an application computing logic matching a type of the computing tasks (paragraph [0086]: logical device 710 distributes workflow 704 to first accelerator 714 and second accelerator 720 by sending workflow 704 to first accelerator 714 and second accelerator 720, and causing first accelerator 714 to perform part of workflow 716…; paragraph [0078]: logical device 404 determines which accelerator to use for each workload based on an acceleration profile associated with each workload; Note: Kundu’s “workflow” corresponds to the claimed “service task,” and dividing the workflow into parts/workloads corresponds to dividing into “computing tasks.” Furthermore, determining which accelerator to use based on the “acceleration profile” corresponds to distributing the tasks to a target accelerator having an “application computing logic” matching the type of computing tasks), whereby the target accelerator executes the computing tasks and shunts the computing tasks to a direct-connected accelerator or to an indirect-connected accelerator via the controller … (paragraph [0079]: first accelerator 408 processes a portion of them and passes intermediate results and any remaining workloads to second accelerator 412 via direct communication mechanism 414; Note: Kundu’s first accelerator executing its portion of the workload and passing the remaining workloads to the second accelerator via the direct communication mechanism corresponds to the target accelerator executing and shunting the computing tasks to a direct-connected accelerator),
wherein the accelerator direct-connection pair comprises two accelerators that are directly connected to each other, and (paragraph [0079]: direct communication mechanism 414 can be a direct memory access (“DMA”), a communication bus, a serial link, or shared region in memory… that does not transition through a central intermediate processor such as a CPU; Note: Kundu’s direct communication mechanism that bypasses the CPU/central processor corresponds to the two accelerators being directly connected to each other), and
wherein at least one of: the two accelerators have the same application computing logic, or at least one of the two accelerators is a first accelerator that supports a computer express link protocol and has an extended memory. (paragraph [0225]: one or more components of system 1600 are interconnected using compute express link (CXL) interconnects; paragraph [0274]: an entire virtual/effective memory space (sometimes referred to as an effective address space) is thereby distributed across each of processor memories 2001-2002 and GPU memories 2020-2023, allowing any processor or GPU to access any physical memory with a virtual address mapped to that memory; paragraph [0276]: GPU-attached memory 2020-2023 to be mapped as part of system memory, and accessed using shared virtual memory (SVM) technology; Note: Kundu’s CXL interconnects correspond to the claimed “computer express link protocol.” Furthermore, Kundu’s “GPU-attached memory” that is mapped as shared virtual memory across the cluster so that any GPU can access it corresponds structurally and functionally to the claimed “extended memory”).
Kundu does not explicitly disclose when the target accelerator is in a computing overload state. Thyagaturu discloses shifting workloads to alternative accelerators when the target accelerator is in a computing overload state (fig. 1-6, paragraph [0040]: during run-time the static values may not provide the optimal result, especially when the accelerator device is highly loaded. In this case, an alternative device with a lower load and the least cost among available accelerators meeting the requirements should be selected; paragraph [0042]: Avg_load_threshold = [a, b]; Note: Thyagaturu’s teaching of an accelerator device becoming “highly loaded” and exceeding an “Avg_load_threshold” corresponds to the target accelerator being in a “computing overload state,” which triggers the selection of an alternative device to offload/shunt the tasks). It would have been obvious to a person of ordinary skill in the art before the effective filing date of the claimed invention to modify the distributed accelerator framework of Kundu by incorporating the load-based accelerator selection and offloading of Thyagaturu. The motivation would have been to provide the optimal result during run-time and avoid processing bottlenecks when an initially selected accelerator device becomes highly loaded (Thyagaturu, paragraph [0040]), thereby ensuring that if the first accelerator reaches a computing overload state, it dynamically utilizes Kundu’s direct-connection link to shunt pending tasks to the second accelerator to maintain optimal parallel processing efficiency.
Regarding claim 2, Kundu discloses
the target accelerator shunting the computing tasks to the direct-connected accelerator when the target accelerator is in the computing overload state comprises: at least one of the target accelerator occupying the extended memory of the direct-connected accelerator or shunting the computing tasks to the direct-connected accelerator for execution when the target accelerator is in the computing overload state; and … (paragraph [0079]: first accelerator 408 processes a portion of them and passes intermediate results and any remaining workloads to second accelerator 412 via direct communication mechanism 414; Thyagaturu, paragraph [0040]: an alternative device with a lower load and the least cost among available accelerators meeting the requirements should be selected; Note: Because the claim requires “at least one of” A or B, teaching B is sufficient. Kundu’s teaching of the first accelerator passing remaining workloads to the second accelerator via the direct communication mechanism, triggered by Thyagaturu’s high load/overload state, corresponds to shunting the computing tasks to the direct-connected accelerator for execution. Furthermore, Kundu also teaches the alternative “OR” condition of occupying the extended memory in paragraph [0276] by allowing the GPU to access GPU-attached memory mapped as shared virtual memory).
Kundu does not explicitly disclose the target accelerator shunting the computing tasks to the indirect-connected accelerator via the controller when the target accelerator is in the computing overload state comprises: at least one of the target accelerator occupying an extended memory of the indirect-connected accelerator or shunting the computing tasks to the indirect-connected accelerator for execution via the controller when the target accelerator is in the computing overload state. Thyagaturu discloses the target accelerator shunting the computing tasks to the indirect-connected accelerator via the controller when the target accelerator is in the computing overload state comprises: at least one of the target accelerator occupying an extended memory of the indirect-connected accelerator or shunting the computing tasks to the indirect-connected accelerator for execution via the controller when the target accelerator is in the computing overload state (paragraph [0049]: the ACPI BIOS will return information including the device ID for the accelerator to the application… using the ACPI_Response message; Thyagaturu, paragraphs [0057]-[0058]: the operating system moves execution of application 222’s process and threads to core 228 on Socket 1… application 222 offloads the workload portion to accelerator 212… the operating system moves execution… to core 230 on Socket 2… application 222 offloads the workload portion to accelerator 214; Note: Thyagaturu teaches that when an accelerator is highly loaded/overloaded, the system (via the ACPI BIOS/OS, which acts as the controller) selects a new accelerator device that is located on a different socket (e.g., moving from Socket 0 to Socket 1 to Socket 2). An accelerator on a different socket connected via socket-to-socket interconnects rather than a direct GPU-to-GPU link corresponds to an “indirect-connected accelerator.” Offloading the workload to this new accelerator via the ACPI BIOS/OS corresponds to shunting the computing tasks to the indirect-connected accelerator for execution via the controller). It would have been obvious to a person of ordinary skill in the art before the effective filing date of the claimed invention to modify the distributed accelerator framework of Kundu by incorporating the load-based accelerator selection and offloading of Thyagaturu. The motivation would have been to provide the optimal result during run-time and avoid processing bottlenecks when an initially selected accelerator device becomes highly loaded (Thyagaturu, paragraph [0040]), thereby ensuring that if the first accelerator reaches a computing overload state, it dynamically utilizes Kundu’s direct-connection link to shunt pending tasks to the second accelerator to maintain optimal parallel processing efficiency.
Regarding claim 4, Kundu discloses
wherein the dividing the service task into computing tasks and distributing the computing tasks to the target accelerator having the application computing logic matching the type of the corresponding computing tasks … comprises: selecting the target accelerator according to at least one of a direct-connection relationship and extended memory occupation, dividing the service task into computing tasks and distributing the computing tasks to the target accelerator (paragraph [0078]: logical device 404 determines which accelerator to use for each workload based on an acceleration profile associated with each workload; paragraph [0079]: logical device 404 utilizes physical communication link between first accelerator 408 and second accelerator 412 so that intermediate results produced by first accelerator 408 can pass directly to second accelerator 412 via direct communication mechanism 414; Note: Kundu teaches dividing and distributing tasks based on matching logic/acceleration profiles, and selecting the target accelerator to receive the divided tasks because it shares a physical communication link/direct-connection relationship with the first accelerator).
Kundu does not explicitly disclose the target accelerator … not being occupied. Thyagaturu discloses distributing the computing tasks to the target accelerator not being occupied (paragraph [0040]: when initial load of the accelerator after a boot-up is zero… an alternative device with a lower load and the least cost among available accelerators meeting the requirements should be selected; Note: Thyagaturu teaches ensuring the selected accelerator has a zero load or lower load, which corresponds to the target accelerator not being occupied). It would have been obvious to a person of ordinary skill in the art before the effective filing date of the claimed invention to modify the method of Kundu by incorporating the load-aware accelerator selection of Thyagaturu. The motivation would have been because during run-time the static values may not provide the optimal result, especially when the accelerator device is highly loaded, and therefore to provide the optimal result, an alternative device with a lower load and the least cost among available accelerators meeting the requirements should be selected (Thyagaturu, paragraph [0040]).
Regarding claim 6, Kundu discloses
wherein the target accelerator executing the computing tasks and shunting the computing tasks to the direct-connected accelerator or to the indirect-connected accelerator via the controller when the target accelerator is in the computing overload state comprises: the target accelerator shunting the computing tasks to the direct-connected accelerator for the target accelerator in a case that the target accelerator is in the computing overload state and the direct-connected accelerator for the target accelerator satisfies a condition that it… has the same application computing logic as that of the target accelerator; (paragraph [0079]: first accelerator 408 processes a portion of them and passes intermediate results and any remaining workloads to second accelerator 412 via direct communication mechanism 414; paragraph [0078]: logical device 404 determines which accelerator to use for each workload based on an acceleration profile associated with each workload; Note: Kundu teaches shunting workloads to a direct-connected accelerator that has matching logic/acceleration profiles).
Kundu does not explicitly disclose the condition that the direct-connected accelerator is in an idle state … assigning the indirect-connected accelerator to the target accelerator as an accelerator to be shunted, whereby the target accelerator shunts the computing tasks to the accelerator to be shunted, in the case that the target accelerator is in the computing overload state and no direct-connected accelerator is provided for the target accelerator or the direct-connection accelerator for the target accelerator is in a non-idle state or the direct-connected accelerator for the target accelerator has different application computing logic from that of the target accelerator. Thyagaturu discloses the condition that the direct-connected accelerator is in an idle state … assigning the indirect-connected accelerator to the target accelerator as an accelerator to be shunted, whereby the target accelerator shunts the computing tasks to the accelerator to be shunted, in the case that the target accelerator is in the computing overload state and no direct-connected accelerator is provided for the target accelerator or the direct-connection accelerator for the target accelerator is in a non-idle state or the direct-connected accelerator for the target accelerator has different application computing logic from that of the target accelerator (paragraph [0040]: when initial load of the accelerator after a boot-up is zero… an alternative device with a lower load and the least cost among available accelerators meeting the requirements should be selected; paragraphs [0057]-[0058]: the operating system moves execution of application 222’s process and threads to core 228 on Socket 1… the operating system moves execution… to core 230 on Socket 2; Note: Thyagaturu teaches evaluating accelerators based on their current load to find one with a zero/lower load, which corresponds to checking for an “idle state” versus a “non-idle state.” Furthermore, by teaching the selection of an alternative device based on the “least cost among available accelerators meeting the requirements,” Thyagaturu teaches a fallback mechanism: if a lowest-cost direct-connected accelerator is unavailable, non-idle, or does not meet the logic requirements, the system falls back to assigning a higher-cost, indirect-connected accelerator across a different socket). It would have been obvious to a person of ordinary skill in the art before the effective filing date of the claimed invention to modify the distributed accelerator framework of Kundu by incorporating the conditional load-and-cost-based fallback assignment of Thyagaturu. The motivation would have been because during run-time the static values may not provide the optimal result, especially when the accelerator device is highly loaded, and therefore to ensure continuous processing, an alternative device with a lower load and the least cost among available accelerators meeting the requirements should be selected (Thyagaturu, paragraph [0040]).
Regarding claim 18, Kundu discloses
A distributed computing method, applied to a target accelerator in a distributed accelerator cluster, the distributed computing method comprising (fig. 1-53):
receiving and executing computing tasks divided and assigned by a controller of the distributed cluster according to a service task (paragraph [0068]: A logical device distributes workloads to a plurality of accelerators; paragraph [0086]: logical device 710 distributes workflow 704 to first accelerator 714 and second accelerator 720 by sending workflow 704 to first accelerator 714… and causing first accelerator 714 to perform part of workflow 716; Note: Kundu’s first accelerator receiving and performing workloads distributed by the logical device corresponds to the target accelerator receiving and executing computing tasks divided and assigned by a controller);
shunting the computing tasks to a direct-connected accelerator or to an indirect-connected accelerator via the controller … (paragraph [0079]: first accelerator 408 processes a portion of them and passes intermediate results and any remaining workloads to second accelerator 412 via direct communication mechanism 414; Note: Kundu’s first accelerator passing remaining workloads to the second accelerator via the direct communication mechanism corresponds to shunting the computing tasks to a direct-connected accelerator),
wherein the target accelerator is an accelerator that is in an idle state and has an application computing logic matching the type of the computing tasks (paragraph [0069]: logical device dequeues workloads from a queue and dispatches them to an appropriate accelerator as resources become available; paragraph [0078]: logical device 404 determines which accelerator to use for each workload based on an acceleration profile associated with each workload; Note: Kundu’s teaching of dispatching workloads to an appropriate accelerator “as resources become available” based on their “acceleration profile” corresponds to the target accelerator being in an idle state and having an application computing logic matching the type of the computing tasks);
the distributed accelerator cluster comprises a pre-established accelerator direct-connection pair, wherein the accelerator direct-connection pair comprises two accelerators that are directly connected to each other (paragraph [0079]: direct communication mechanism 414 can be a direct memory access (“DMA”), a communication bus, a serial link, or shared region in memory… that does not transition through a central intermediate processor such as a CPU; Note: Kundu’s direct communication mechanism that bypasses the CPU/central processor corresponds to the two accelerators being directly connected to each other), and
at least one of: the two accelerators have the same application computing logic, or at least one of the two accelerators is a first accelerator that supports a computer express link protocol and has an extended memory (paragraph [0225]: one or more components of system 1600 are interconnected using compute express link (CXL) interconnects; paragraph [0274]: an entire virtual/effective memory space… is thereby distributed across each of processor memories 2001-2002 and GPU memories 2020-2023; paragraph [0276]: GPU-attached memory 2020-2023 to be mapped as part of system memory, and accessed using shared virtual memory (SVM) technology; Note: Kundu’s CXL interconnects correspond to the claimed “computer express link protocol.” Furthermore, Kundu’s “GPU-attached memory” that is mapped as shared virtual memory across the cluster corresponds to the claimed “extended memory”).
Kundu does not explicitly disclose shunting the computing tasks when the target accelerator is in a computing overload state. Thyagaturu discloses shifting workloads to alternative accelerators when the target accelerator is in a computing overload state (fig. 1-6, paragraph [0040]: during run-time the static values may not provide the optimal result, especially when the accelerator device is highly loaded. In this case, an alternative device with a lower load and the least cost among available accelerators meeting the requirements should be selected; paragraph [0042]: Avg_load_threshold; Note: Thyagaturu’s teaching of an accelerator device becoming “highly loaded” and exceeding an “Avg_load_threshold” corresponds to the target accelerator being in a “computing overload state,” which triggers the selection of an alternative device to offload/shunt the tasks). It would have been obvious to a person of ordinary skill in the art before the effective filing date of the claimed invention to modify the distributed accelerator framework of Kundu by incorporating the load-based accelerator selection and offloading of Thyagaturu. The motivation would have been to provide the optimal result during run-time and avoid processing bottlenecks when an initially selected accelerator device becomes highly loaded (Thyagaturu, paragraph [0040]), thereby ensuring that if the first accelerator reaches a computing overload state, it dynamically utilizes Kundu’s direct-connection link to shunt pending tasks to the second accelerator to maintain optimal parallel processing efficiency.
Regarding claim 22, Kundu discloses
A distributed computing system comprises (fig. 1-53) a distributed accelerator cluster and a controller (paragraph [0068]: A logical device distributes workloads to a plurality of accelerators; paragraph [0076]: application 302 uses a logical device 304 as an interface to a plurality of accelerators; Note: Kundu’s “logical device” corresponds to the claimed “controller,” and the “plurality of accelerators” corresponds to the “distributed accelerator cluster”),
wherein the controller is configured to acquire information of accelerators in the distributed accelerator cluster (paragraph [0082]: logical device 610 examines workflow 604 and determines, based at least in part on knowledge of available acceleration resources and acceleration profiles; Note: Kundu’s teaching of acquiring knowledge of available acceleration resources and their acceleration profiles corresponds to acquiring information of the accelerators);
establish an accelerator direct-connection pair according to the information of the accelerators (paragraph [0079]: logical device 404 utilizes physical communication link between first accelerator 408 and second accelerator 412; Note: Kundu’s logical device utilizing a physical communication link to pair a first and second accelerator together based on their acceleration profiles/information corresponds to establishing an accelerator direct-connection pair);
and in response to receiving a service task, divide the service task into computing tasks and distribute the computing tasks to an idle target accelerator having an application computing logic matching the type of the corresponding computing tasks (paragraph [0086]: logical device 710 distributes workflow 704 to first accelerator 714 and second accelerator 720 by sending workflow 704 to first accelerator 714… and causing first accelerator 714 to perform part of workflow 716; paragraph [0069]: logical device dequeues workloads from a queue and dispatches them to an appropriate accelerator as resources become available; paragraph [0078]: logical device 404 determines which accelerator to use for each workload based on an acceleration profile associated with each workload; Note: Kundu’s “workflow” corresponds to the claimed “service task,” and dividing the workflow into parts/workloads corresponds to dividing into “computing tasks.” Dispatching as resources become available based on the acceleration profile corresponds to distributing to an idle target accelerator having matching application computing logic),
whereby the target accelerator executes the computing tasks and shunts the computing tasks to a direct-connected accelerator or to an indirect-connected accelerator via the controller … (paragraph [0079]: first accelerator 408 processes a portion of them and passes intermediate results and any remaining workloads to second accelerator 412 via direct communication mechanism 414; Note: Kundu’s first accelerator executing its portion of the workload and passing the remaining workloads to the second accelerator via the direct communication mechanism corresponds to the target accelerator executing and shunting the computing tasks to a direct-connected accelerator),
wherein the accelerator direct-connection pair comprises two accelerators that are directly connected to each other (paragraph [0079]: direct communication mechanism 414 can be a direct memory access (“DMA”), a communication bus, a serial link, or shared region in memory… that does not transition through a central intermediate processor such as a CPU; Note: Kundu’s direct communication mechanism that bypasses the CPU/central processor corresponds to the two accelerators being directly connected to each other), and
at least one of the two accelerators is a first accelerator that supports a computer express link protocol and has an extended memory, and/or the two accelerators have the same application computing logic (paragraph [0225]: one or more components of system 1600 are interconnected using compute express link (CXL) interconnects; paragraph [0274]: an entire virtual/effective memory space… is thereby distributed across each of processor memories 2001-2002 and GPU memories 2020-2023; paragraph [0276]: GPU-attached memory 2020-2023 to be mapped as part of system memory, and accessed using shared virtual memory (SVM) technology; Note: Kundu’s CXL interconnects correspond to the claimed “computer express link protocol.” Furthermore, Kundu’s “GPU-attached memory” that is mapped as shared virtual memory across the cluster corresponds to the claimed “extended memory”).
Kundu does not explicitly disclose the controller being configured such that the target accelerator shunts the computing tasks when the target accelerator is in a computing overload state. Thyagaturu discloses shifting workloads to alternative accelerators when the target accelerator is in a computing overload state (fig. 1-6, paragraph [0040]: during run-time the static values may not provide the optimal result, especially when the accelerator device is highly loaded. In this case, an alternative device with a lower load and the least cost among available accelerators meeting the requirements should be selected; paragraph [0042]: Avg_load_threshold = [a, b]; Note: Thyagaturu’s teaching of an accelerator device becoming “highly loaded” and exceeding an “Avg_load_threshold” corresponds to the target accelerator being in a “computing overload state,” which triggers the selection of an alternative device to offload/shunt the tasks). It would have been obvious to a person of ordinary skill in the art before the effective filing date of the claimed invention to modify the distributed accelerator system of Kundu by incorporating the load-based accelerator selection and offloading of Thyagaturu. The motivation would have been to provide the optimal result during run-time and avoid processing bottlenecks when an initially selected accelerator device becomes highly loaded (Thyagaturu, paragraph [0040]), thereby ensuring that if the first accelerator reaches a computing overload state, the system dynamically utilizes Kundu’s direct-connection link to shunt pending tasks to the second accelerator to maintain optimal parallel processing efficiency.
Regarding claim 23, Kundu discloses A distributed computing device comprisies a storage device, configured to store computer-readable instructions therein; and a processor, configured to execute the computer-readable instructions (paragraph [0225]: system 1600 includes one or more processors… and memory; paragraph [0274]: processor memories 2001-2002), the computer-readable instructions, when executed by the processor, implementing the distributed computing method according to claim 1 (See the rejection for claim 1).
Regarding claim 24, Kundu discloses A non-transitory computer-readable storage medium, having computer-readable instructions stored therein (paragraph [0225]: system 1600 includes one or more processors… and memory; paragraph [0274]: processor memories 2001-2002), wherein the computer-readable instructions, when executed by a processor, implement the distributed computing method according to claim 1 (See the rejection for claim 1).
Claims 13-16 are rejected under 35 U.S.C. 103 as being unpatentable over Kundu (US 2022/0276914, hereinafter Kundu) in view of Thyagaturu (US 2022/0075655, hereinafter Thyagaturu) as applied to claim 1, and further in view of Luitjens (US 20230039180, hereinafter Luitjens).
Regarding claim 13, Kundu discloses
wherein the establishing the accelerator direct-connection pair comprises: establishing the accelerator direct-connection pair by applying … communication protocol; and the target accelerator shunting the computing tasks to the direct-connected accelerator comprises: the target accelerator shunting the computing tasks to the direct-connected accelerator for the target accelerator on the basis of … transmission link (paragraph [0079]: logical device 404 utilizes physical communication link between first accelerator 408 and second accelerator 412 so that intermediate results produced by first accelerator 408 can pass directly to second accelerator 412 via direct communication mechanism 414; paragraph [0225]: one or more components of system 1600 are interconnected using compute express link (CXL); Note: Kundu teaches establishing a direct physical communication link between accelerators and passing workloads/results directly between them. In the context of GPU/accelerator architectures, direct accelerator-to-accelerator links, such as CXL or PCIe peer-to-peer connections, function as high-speed transmission links that utilize inter-device/inter-kernel communication protocols to allow computing kernels on one accelerator to communicate directly with kernels on another).
Kundu in view of Thyagaturu does not explicitly detail the specific protocol or name the link an inter-kernel high-speed transmission link. Luitjens discloses a computer system comprising a plurality of GPUs where the GPUs are interconnected with high-speed interconnects to execute compute kernels (paragraph [0027]: GPUs 140, CPUs 120, and PCIe switches 130 may be interconnected with high-speed interconnects such as, for example and without limitation, NVLink interfaces 160… and/or PCIe connections 150; paragraph [0023]: a compute kernel includes a routine compiled for high throughput accelerators, such as graphics processing units). Note: Kundu teaches establishing a direct physical communication link between accelerators and passing workloads/results directly between them. Luitjens teaches that accelerators execute compute kernels and communicate over high-speed interconnects. In the context of GPU/accelerator architectures, direct accelerator-to-accelerator links, such as CXL, NVLink, or PCIe peer-to-peer connections, function as high-speed transmission links that utilize inter-device/inter-kernel communication protocols to allow computing kernels on one accelerator to communicate directly with kernels on another.
It would have been obvious to a person of ordinary skill in the art before the effective filing date of the claimed invention to modify the direct communication mechanism of Kundu in view of Thyagaturu by implementing the inter-kernel high-speed interconnects such as NVLink or PCIe taught by Luitjens. The motivation would have been to provide high-speed interconnects (Luitjens, paragraph [0027]) to support high throughput accelerators executing compute kernels (Luitjens, paragraph [0023]), thereby ensuring that the direct passing of workloads and intermediate results between the paired accelerators as taught by Kundu occurs with sufficient bandwidth and speed to avoid processing bottlenecks and maximize overall system throughput.
Regarding claim 14, Kundu discloses
wherein the target accelerator shunting the computing tasks (Kundu, paragraph [0079]: first accelerator 408 processes a portion of them and passes intermediate results and any remaining workloads to second accelerator 412).
Kundu does not explicitly disclose shunting the computing tasks to the indirect-connected accelerator via the controller comprises: receiving a shunting request sent by the target accelerator; determining an accelerator to be shunted on the basis of the shunting request; and sending information of the accelerator to be shunted to the target accelerator, whereby the target accelerator shunts the computing tasks to the accelerator to be shunted.
Thyagaturu discloses shunting the computing tasks to the indirect-connected accelerator via the controller comprises: receiving a shunting request sent by the target accelerator; determining an accelerator to be shunted on the basis of the shunting request; and sending information of the accelerator to be shunted to the target accelerator, whereby the target accelerator shunts the computing tasks to the accelerator to be shunted (paragraph [0040]: when initial load of the accelerator after a boot-up is zero… an alternative device with a lower load and the least cost among available accelerators meeting the requirements should be selected; paragraphs [0057]-[0058]: the operating system moves execution of application 222’s process and threads to core 228 on Socket 1… the operating system moves execution… to core 230 on Socket 2; Note: Thyagaturu teaches the controller/OS receiving load status updates (acting as a shunting request), determining an alternative indirect-connected accelerator across a different socket based on the load, and orchestrating the move of the execution by sending information to shift the workload to that alternative accelerator). It would have been obvious to a person of ordinary skill in the art before the effective filing date of the claimed invention to modify the distributed accelerator framework of Kundu by incorporating the controller-based alternative accelerator assignment of Thyagaturu. The motivation would have been because during run-time the static values may not provide the optimal result, especially when the accelerator device is highly loaded, and therefore to ensure continuous processing, an alternative device with a lower load and the least cost among available accelerators meeting the requirements should be selected (Thyagaturu, paragraph [0040]).
Kundu in view of Thyagaturu does not explicitly disclose the target accelerator sending the tasks via a routing subnetwork. Luitjens discloses a computer system comprising a plurality of GPUs where the GPUs are interconnected via a routing subnetwork (paragraph [0027]: GPUs 140, CPUs 120, and PCIe switches 130 may be interconnected with high-speed interconnects such as, for example and without limitation, NVLink interfaces 160… and/or PCIe connections 150; Note: Luitjens teaches interconnecting accelerators using PCIe switches, which function as a routing subnetwork to route data between devices). It would have been obvious to a person of ordinary skill in the art before the effective filing date of the claimed invention to modify the distributed computing method of Kundu in view of Thyagaturu by implementing the PCIe switches to act as a routing subnetwork as taught by Luitjens. The motivation would have been to provide a scalable architecture where PCIe switches 130 may be interconnected with high-speed interconnects (Luitjens, paragraph [0027]) to allow the host controller to dynamically orchestrate the routing of workloads between indirect-connected accelerators across different sockets without forcing all data to bottleneck through the host CPU, thereby optimizing system throughput and load distribution.
Regarding claim 15, Kundu discloses
wherein the target accelerator shunting the computing tasks to the accelerator to be shunted … comprises: the target accelerator shunting the computing tasks to the accelerator to be shunted … (Kundu, paragraph [0079]: first accelerator 408 processes a portion of them and passes intermediate results and any remaining workloads to second accelerator 412).
Kundu in view of Thyagaturu does not explicitly disclose via a routing subnetwork … via the routing subnetwork on the basis of a remote direct memory access protocol. Luitjens discloses via a routing subnetwork … via the routing subnetwork on the basis of a remote direct memory access protocol (paragraph [0093]: DMA may enable components of PVA to access system memory independently of CPU(s) 706; paragraph [0112]: include their own DMA engines, and that may be used to free CPU(s) 706 from routine data management tasks; Note: In the context of distributed computing where GPUs are interconnected via a routing subnetwork such as PCIe switches, utilizing DMA to transfer data directly between the memories of remote accelerators without CPU intervention constitutes utilizing a remote direct memory access (RDMA) protocol, such as PCIe Peer-to-Peer DMA). It would have been obvious to a person of ordinary skill in the art before the effective filing date of the claimed invention to modify the distributed computing method of Kundu in view of Thyagaturu and Luitjens to execute the shunting of computing tasks across the routing subnetwork on the basis of a remote direct memory access protocol. The motivation would have been to free CPU(s) 706 from routine data management tasks (Luitjens, paragraph [0112]) and enable the accelerators to transfer data independently of CPU(s) (Luitjens, paragraph [0093]), thereby significantly reducing latency and preventing the host CPU from becoming a bottleneck when transferring heavy workloads between indirect-connected accelerators.
Regarding claim 16, Kundu discloses at least one of selecting the candidate shunting accelerator that satisfies the longest idle time or belongs to the first accelerators as the accelerator to be shunted (paragraph [0079]: first accelerator 408 processes a portion of them and passes intermediate results and any remaining workloads to second accelerator 412 via direct communication mechanism 414; paragraph [0225]: one or more components of system 1600 are interconnected using compute express link (CXL) interconnects; paragraph [0276]: GPU-attached memory 2020-2023 to be mapped as part of system memory, and accessed using shared virtual memory (SVM) technology; Note: As established in claim 1, Kundu’s CXL-enabled accelerators correspond to the claimed “first accelerators,” thus Kundu teaches selecting a candidate that belongs to the first accelerators).
Kundu does not explicitly disclose the determining the accelerator to be shunted according to the shunting request comprises: acquiring an accelerator list of the distributed accelerator cluster; determining in the accelerator list information indicating that the accelerators that have the same application computing logics as that of the target accelerator and are idle are candidate shunting accelerators.
Thyagaturu discloses the determining the accelerator to be shunted according to the shunting request comprises: acquiring an accelerator list of the distributed accelerator cluster; determining in the accelerator list information indicating that the accelerators that have the same application computing logics as that of the target accelerator and are idle are candidate shunting accelerators (paragraph [0039]: FIG. 3b shows an exemplary core-to-accelerator cost table 350 containing core-to-accelerator distance cost metrics; paragraph [0040]: when initial load of the accelerator after a boot-up is zero… an alternative device with a lower load and the least cost among available accelerators meeting the requirements should be selected; Note: Thyagaturu teaches acquiring a cost table which acts as an accelerator list, and evaluating available accelerators that “meet the requirements” with a “zero” load, which corresponds to determining candidate accelerators that have the same application computing logic and are idle). It would have been obvious to a person of ordinary skill in the art before the effective filing date of the claimed invention to modify the distributed computing method of Kundu by incorporating the accelerator list and selection of Thyagaturu. The motivation would have been to ensure that when an accelerator device is highly loaded, an alternative device with a lower load among available accelerators meeting the requirements is selected (Thyagaturu, paragraph [0040]), thereby dynamically shifting workloads to idle CXL-enabled accelerators (first accelerators) within the cluster to maintain optimal parallel processing efficiency.
Claim 17 is rejected under 35 U.S.C. 103 as being unpatentable over Kundu (US 2022/0276914, hereinafter Kundu) in view of Thyagaturu (US 2022/0075655, hereinafter Thyagaturu) as applied to claim 1, and further in view of Dubal et al. (US 2011/0142064, hereinafter Dual).
Regarding claim 17, Kundu does not explicitly disclose the target accelerator being in the computing overload state comprises: the target accelerator recording a full occupation timestamp when the local memory is fully occupied for the first time, and querying a local memory occupation state every query cycle; and determining the local memory to be in the computing overload state in the case that the local memory is still fully occupied for a continuous preset cycle.
Thyagaturu discloses the target accelerator being in the computing overload state comprises: the target accelerator … querying a local memory occupation state every query cycle (paragraph [0040]: during run-time the static values may not provide the optimal result, especially when the accelerator device is highly loaded. In this case, an alternative device with a lower load and the least cost among available accelerators meeting the requirements should be selected; paragraph [0042]: Avg_load_threshold = [a, b]; paragraph [0054]: periodically check MSR registers; Note: Thyagaturu teaches querying the state of the system periodically, i.e., every query cycle). It would have been obvious to a person of ordinary skill in the art before the effective filing date of the claimed invention to modify the distributed accelerator framework of Kundu by incorporating the load-based accelerator selection and offloading of Thyagaturu. The motivation would have been to provide the optimal result during run-time and avoid processing bottlenecks when an initially selected accelerator device becomes highly loaded (Thyagaturu, paragraph [0040]), thereby ensuring that if the first accelerator reaches a computing overload state, it dynamically utilizes Kundu’s direct-connection link to shunt pending tasks to the second accelerator to maintain optimal parallel processing efficiency.
Dubal discloses recording a full occupation timestamp when the local memory is fully occupied for the first time… and determining the local memory to be in the computing overload state in the case that the local memory is still fully occupied for a continuous preset cycle (paragraph [0034]: In order to prevent “false alarms” of an overloaded queue, circuitry 310 may include a time-out mechanism (e.g., timer circuitry…) that is permitted to lapse before any message is sent to the OS… the high threshold may momentarily be exceeded… Before circuitry 310 generates a message, circuitry 310 may wait until a predetermined time-out period has lapsed. Note: Dubal’s teaching of a high threshold being exceeded corresponds to the memory being fully occupied for the first time. The “time-out mechanism/timer circuitry” corresponds to recording a timestamp. Waiting until a “predetermined time-out period has lapsed” before generating the overload message corresponds to determining the overload state only if it remains fully occupied for a continuous preset cycle). It would have been obvious to a person of ordinary skill in the art before the effective filing date of the claimed invention to modify the periodic load monitoring of Kundu in view of Thyagaturu by incorporating the time-out overload detection mechanism of Dubal. The motivation would have been to prevent false alarms of an overloaded memory/queue in a bursty environment (Dubal, paragraph [0034]), thereby ensuring that the system only incurs the overhead of shunting computing tasks to an alternative accelerator when the target accelerator is experiencing a true, sustained computing overload state rather than a momentary spike.
Allowable Subject Matter
Claims 3, 5, and 7-11 are objected to as being dependent upon a rejected base claim, but would be allowable if rewritten in independent form including all of the limitations of the base claim and any intervening claims.
Conclusion
The prior art made of record and not relied upon is considered pertinent to applicant’s disclosure.
Currid et al. (US 2024/0143408) discloses systems and methods for migrating workloads across distributed hardware components including direct accelerator-to-accelerator links and indirect connections based on path performance and hardware compatibility (see, e.g., paragraphs [0067], [0076]-[0080], [0084]-[0087], [0118]). This teaching is highly relevant to the disclosure of utilizing pre-established accelerator direct-connection pairs and dynamically shunting computing tasks between target accelerators to manage computing workloads.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to SISLEY N. KIM whose telephone number is (571)270-7832. The examiner can normally be reached M-F 11:30AM -7:30PM.
Examiner interviews are available via telephone, in-person, and video conferencing using a USPTO supplied web-based collaboration tool. To schedule an interview, applicant is encouraged to use the USPTO Automated Interview Request (AIR) at http://www.uspto.gov/interviewpractice.
If attempts to reach the examiner by telephone are unsuccessful, the examiner’s supervisor, April Y. Blair can be reached on (571)270-1014. The fax phone number for the organization where this application or proceeding is assigned is 571-273-8300. The fax phone number for the organization where this application or proceeding is assigned is 571-273-8300.
Information regarding the status of published or unpublished applications may be obtained from Patent Center. Unpublished application information in Patent Center is available to registered users. To file and manage patent submissions in Patent Center, visit: https://patentcenter.uspto.gov. Visit https://www.uspto.gov/patents/apply/patent-center for more information about Patent Center and https://www.uspto.gov/patents/docx for information about filing in DOCX format. For additional questions, contact the Electronic Business Center (EBC) at 866-217-9197 (toll-free). If you would like assistance from a USPTO Customer Service Representative, call 800-786-9199 (IN USA OR CANADA) or 571-272-1000.
/SISLEY N KIM/Primary Examiner, Art Unit 2196 10/26/2026