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 Rejections - 35 USC § 102
The following is a quotation of the appropriate paragraphs of 35 U.S.C. 102 that form the basis for the rejections under this section made in this Office action:
A person shall be entitled to a patent unless –
(a)(1) the claimed invention was patented, described in a printed publication, or in public use, on sale, or otherwise available to the public before the effective filing date of the claimed invention.
(a)(2) the claimed invention was described in a patent issued under section 151, or in an application for patent published or deemed published under section 122(b), in which the patent or application, as the case may be, names another inventor and was effectively filed before the effective filing date of the claimed invention.
Claim(s) 1, 4, 9, 10, 13, and 16 is/are rejected under 35 U.S.C. 102(a)(2) as being anticipated by CHOI et al. Pub. No. US 2021/0390405 A1 (hereafter Choi).
The applied reference has a common applicant and/or inventor(s) with the instant application. Based upon the earlier effectively filed date of the reference, it constitutes prior art under 35 U.S.C. 102(a)(2). This rejection under 35 U.S.C. 102(a)(2) might be overcome by: (1) a showing under 37 CFR 1.130(a) that the subject matter disclosed in the reference was obtained directly or indirectly from the inventor or a joint inventor of this application and is thus not prior art in accordance with 35 U.S.C. 102(b)(2)(A); (2) a showing under 37 CFR 1.130(b) of a prior public disclosure under 35 U.S.C. 102(b)(2)(B) if the same invention is not being claimed; or (3) a statement pursuant to 35 U.S.C. 102(b)(2)(C) establishing that, not later than the effective filing date of the claimed invention, the subject matter disclosed in the reference and the claimed invention were either owned by the same person or subject to an obligation of assignment to the same person or subject to a joint research agreement.
Regarding claim 1, Choi teaches “A training system comprising: a job proxy configured to partition a training job corresponding to a neural network model into a plurality of microservices respectively executed by a plurality of logical workers ([0012] teaches a job proxy that receives a training request corresponding to a neural network model, and partitions a training job into microservices; [0032] teaches executing microservices through a plurality of virtual workers); and a scheduler configured to schedule the plurality of microservices to a plurality of processing units ([0012] teaches a scheduler that dynamically schedules microservices to be executed by processing units), respectively, wherein the plurality of microservices include a plurality of first microservices executed by a first logical worker among the plurality of logical workers and a plurality of second microservices executed by a second logical worker among the plurality of logical workers ([0029-0036] teaches executing a first and second task on a first and second virtual worker respectively), and the scheduler is configured to schedule the plurality of first microservices and the plurality of second microservices to any one processing unit among the plurality of processing units based on an availability status of the plurality of processing units ([0012] teaches the scheduler dynamically scheduling the first and second microservices to be executed by processing units; [0028] teaches scheduling microservices to be executed by an available processing unit among the processing units).”
Regarding claim 13, it is similar to claim 1 and is rejected for the same reasons. Claim 13 is directed towards “An operating method ([0013])”
Examiner suggests that claims 1, 13 be amended to include the main novelties of the invention and/or to elaborate on what the availability status entails within the claim.
Regarding claim 4, Choi teaches “The training system of claim 1, wherein the plurality of microservices each include a function that processes a plurality of minibatches obtained by partitioning training data for the training job ([0024] teaches that a microservice may be a computation unit, that is, a function in a microservice system such that the microservice may be the function, and the microservice processes a mini-batch from the training data), and the scheduler is configured to schedule minibatch processing of any one of the plurality of first microservices and minibatch processing of any one of the plurality of second microservices to be executed in multiple phases in the any one processing unit ([0047-0052] teaches microservices being subdivided into microservices with regard to a mini-batch, indicating an ‘F’ or ‘B’ for a forward or backward pass respectively. [0072] teaches the scheduler scheduling a forward/backward pass of a mini-batch; the scheduler may dynamically schedule a microservice which aggregates parameters corresponding to the mini-batch processed at each virtual worker. [0012] teaches scheduling a first and second microservice).”
Regarding claim 16, it is similar to claim 4 and is rejected for the same reasons.
Regarding claim 9, Choi teaches “The training system of claim 1, wherein the plurality of microservices includes a computation function that computes respective weights for a plurality of minibatches ([0038, 0046-0047] teaches first type microservices being works in which the plurality of virtual workers calculate weights, and may in regards to minibatches for each microservice divided into microservices), and an aggregation function ([0040] teaches a second-type microservices that aggregates the calculated weights) that computes a global parameter obtained by aggregating the respective weights for the plurality of minibatches ([0059-0062] teaches calculating parameters and aggregating the parameters and storing them; a parameter globally shared the training process may also be stored such that a value of a parameter of the neural network is calculated whenever each mini-batch is processed by virtual workers (e.g. calculating weights) and then the values are aggregated and stored, wherein the parameters may be globally shared)”.
Regarding claim 10, Choi teaches “The training system of claim 9, wherein a first computation function of the plurality of first microservices and a second computation function of the plurality of second microservices are sequentially executed in a same iteration ([0033-0036] teaches a sequential execution order within a virtual worker (e.g. M1-1 -> Agg1 -> M1-2 -> Agg2 …). [0041] teaches that a first-partial microservice is executed for calculating weights and then aggregate the first weights by executing the second-type microservice such that it suggests iterative execution. [0061-0062] teaches synchronization policies applied after each mini-batch is processed such that it suggests an iterative step of synchronization after mini-batch processing)”.
Claim Rejections - 35 USC § 103
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, 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 2, 11, 12, 14 and 20 are rejected under 35 U.S.C. 102(a)(2) as anticipated by or, in the alternative, under 35 U.S.C. 103 as obvious over Choi as used above in claims 1 and 13.
The applied reference has a common applicant and/or inventor(s) with the instant application. Based upon the earlier effectively filed date of the reference, it constitutes prior art under 35 U.S.C. 102(a)(2).
This rejection under 35 U.S.C. 103 might be overcome by: (1) a showing under 37 CFR 1.130(a) that the subject matter disclosed in the reference was obtained directly or indirectly from the inventor or a joint inventor of this application and is thus not prior art in accordance with 35 U.S.C.102(b)(2)(A); (2) a showing under 37 CFR 1.130(b) of a prior public disclosure under 35 U.S.C. 102(b)(2)(B); or (3) a statement pursuant to 35 U.S.C. 102(b)(2)(C) establishing that, not later than the effective filing date of the claimed invention, the subject matter disclosed and the claimed invention were either owned by the same person or subject to an obligation of assignment to the same person or subject to a joint research agreement. See generally MPEP § 717.02.
Regarding claim 2, Choi teaches “The training system of claim 1, wherein the scheduler is further configured to sequentially schedule the plurality of first microservices and the plurality of second microservices to the any one processing unit ([0026] teaches adding microservices to a queue of the scheduler, and the that the scheduler may dynamically schedule the microservices to be executed by a cluster of processing units; [0105] teaches the scheduling of microservices such that the second microservices are executed immediately after the execution of other microservices such that it suggests a sequential scheduling of microservices to be executed in a sequential order)...”.
Choi does not explicitly teach that the microservices are scheduled in accordance with a number of available processing units being less than a number of logical workers.
Choi, however, implies the limitation “in accordance with a number of available processing units being less than a number of the plurality of logical workers” by teaching that virtual workers are not bound to a dedicated GPU as they are decoupled, and that they may be executed without the allocation of a dedicated GPU such that they are just dynamically scheduled to be executed by an available processing unit in a heterogenous cluster of processing units [0026-0032]. Additionally, Choi teaches that multiple training jobs share the same cluster [0069] and that the scheduler may allocate available resources based on resource availability and cluster conditions [0069-0074, 0106]. A PHOSITA would have understood that when the number of available processing units is less than the number of virtual workers, the plurality of virtual workers will not be able to execute tasks in parallel within the GPU cluster. Instead, the scheduler would have to schedule the microservices of different virtual workers sequentially on the same available processing unit as they become available, allowing for continued execution of training jobs while sharing limited GPU resources (as implied by [0105]).
It would have been obvious to a person of ordinary skill in the art before the effective
filing date to combine the teachings of Choi to show that the sequential scheduling of microservices may be in accordance with a number of available processing units being less than a number of virtual workers. A person having ordinary skill in the art would have been motivated to adopt this strategy to improve fairness among multiple training jobs while permitting newly arriving microservices to execute without requiring preemption of currently executing jobs [0072-0073] as well as enabling scalable execution of a greater number of virtual workers than available GPUs while reducing the need for dedicated hardware allocation [0121].
Regarding claim 14, it is similar to claim 2 and is rejected for the same reasons.
Regarding claim 11, Choi does not explicitly teach that global parameters are read and transferred. Choi implies such that it teaches the limitation “The training system of claim 9, wherein a first computation function of the plurality of first microservices reads the global parameter and transfers the global parameter to a second computation function of the plurality of second microservices ([0051] teaches of microservice updates, updating parameters so as to be used by any other virtual worker. [0060] teaches storing a globally shared parameter in a database, and that parameters may be shared between virtual workers. [0061-0062] teaches synchronization policies for parameters and aggregating parameters based on the policies for subsequent computation by a generated microservice. A PHOSITA would have understood that the microservice performing the subsequent work depending on the policy would be a first microservice that implicitly reads the parameter in order to carry out the aggregation work, and then recalling that global parameters are shared among virtual workers, it is implied that the aggregation of parameters effectively transfer/share the parameter to other virtual workers to carry out the next step of executing the second-type microservices. Also see Fig. 1B)”.
It would have been obvious to a person of ordinary skill in the art before the effective
filing date to combine the teachings of Choi to show that microservices may share/transfer global parameters. A person having ordinary skill in the art would have been motivated to adopt this strategy to maintain a consistent global model across multiple virtual workers, allowing subsequent computation functions to continue training using synchronized parameters.
Regarding claim 12, Choi teaches “The training system of claim 1, wherein the plurality of microservices includes a plurality of third microservices executed by a third logical worker among the plurality of logical workers ([0029] teaches executing a third task by a plurality of virtual workers, one of which being virtual worker 3) and a plurality of fourth microservices executed by a fourth logical worker among the plurality of logical workers ([0024-0026] teaches of a fourth microservice, though not explicitly used in the embodiments, it may be scheduled as to be executed by a cluster as seen in [0025-0026], and [0030] discloses that the number of virtual workers are not limited to 3, therefore if the fourth were implemented in embodiments, it would implicitly include a fourth virtual worker to execute the fourth task), the scheduler schedules the plurality of third microservices and the plurality of fourth microservices to another processing unit among the plurality of processing units ([0028] teaches scheduling microservices to be executed by an available processing unit among the processing units such that the third and fourth may be scheduled to a different processing unit after the first and second, as the processing unit that they were scheduled on may become unavailable), and the plurality of third microservices are executed in parallel with the plurality of first microservices ([0029] teaches that the first, second, and third task may be executed in parallel).”
It would have been obvious to a person of ordinary skill in the art before the effective
filing date to combine the teachings of Choi to show that more microservices may be executed by a plurality of virtual workers, the microservices deployed on a different processing unit. A person having ordinary skill in the art would have been motivated to adopt this strategy in order to increase the throughput of distributed neural network training by allowing multiple portions of the trained workload be processed concurrently.
Regarding claim 20, it is similar to claim 12 and is rejected for the same reasons.
Claims 3 and 15 is/are rejected under 35 U.S.C. 103 as being unpatentable over Choi as used in claims 1 and 13 above, and in further view of Adams et al. Pub. No. US 2024/0143301 A1 (hereafter Adams).
Regarding claim 3, Choi does not explicitly teach scheduling microservices to a same container.
Adams teaches that containers may contain multiple microservices such that it teaches the limitation “The training system of claim 1, wherein the scheduler is configured to schedule the plurality of first microservices and the plurality of second microservices to a same container ([0049] teaches that each of the containers may contain multiple microservices such that a first and second microservice may be part of the same container)”.
It would have been obvious to a person of ordinary skill in the art before the effective
filing date to combine the teachings of Adams to the invention of Choi to modify the scheduler of Choi to perform the scheduling of microservices to the same container. A person having ordinary skill in the art would have been motivated to adopt this strategy to improve various aspects of the containerized cloud architecture such as reliability, security, agility, and efficiency (Adams [0049] as well as reducing latency between cooperating microservices and improving the locality of execution thereby improving the overall performance of the distributed training system.
Regarding claim 15, it is similar to claim 3 and is rejected for the same reasons.
Claims 5, 6, and 17 is/are rejected under 35 U.S.C. 103 as being unpatentable over Choi as used in claims 1 and 13 above, and in further view of CAPES et al. Pub. No. US 2020/0159589 A1 (hereafter Capes).
Regarding claim 5, Choi teaches “The training system of claim 1, further comprising a resource manager ([0074] teaches a resource manager)”.
Choi does not explicitly teach the resource manager allocating processing units in units of 2n corresponding to a job.
Capes teaches a number of processing units assigned for a job being a power of 2 such that it teaches the limitation “wherein the resource manager is configured to allocate the plurality of processing units in units of 2n to the training job corresponding to the neural network model (wherein n is 0 or a natural number). ([0059-0063] teaches a number of processing units provisionally assigned for a job, that is where fj(gi) for j = 1 and i is a power of two for the number of processors assigned to the job)”.
It would have been obvious to a person of ordinary skill in the art before the effective
filing date to combine the teachings of Capes to the invention of Choi in order to configure the resource manager to allocate 2n processing units to a training job. A person having ordinary skill in the art would have been motivated to adopt this strategy to improve training efficiency of distributed training (Capes [0049]) and to avoid getting stuck at local optima (Capes [0082]).
Regarding claim 17, it is similar to claim 5 and is rejected for the same reasons.
Regarding claim 6, the combination teaches “The training system of claim 1, further comprising a resource manager (Choi [0074]), … , the resource manager is configured to allocate the plurality of processing units in units of any one of divisors of 2k, and when the training job includes … logical workers, the resource manager is configured to allocate the plurality of processing units in units of any one of divisors of 2k-1 or divisors of 2k, except for 1 and 2k (wherein k is a natural number). (Capes [0059-0063] teaches allocating processing units in powers of 2 such that it is a divisor of 2k)”
The combination does not explicitly teach of including 2k logical workers corresponding to a job.
Capes teaches allocating in powers of two such that it teaches the limitation “wherein, when the training job corresponding to the neural network model includes 2k logical workers ([0015] teaches a doubling heuristic when assigning a job to a processing unit and wherein jobs are two or more, it would then require allowing two or more processing units to each job of the plurality of jobs such that there may be a worker for each job)”.
It would have been obvious to a person of ordinary skill in the art before the effective
filing date to combine the teachings of Capes to the invention of Choi in order to modify the number of virtual workers deployed for a job according to the doubling heuristic. A person having ordinary skill in the art would have been motivated to adopt this strategy to improve training time (Capes [0015]) and to avoid getting stuck at local optima, improving simulation times (Capes [0082]).
Claims 8 and 19 is/are rejected under 35 U.S.C. 103 as being unpatentable over Choi as used in claims 1 and 13 above, and Capes as used in claim 5 above, and in further view of HU et al. Pub. No. US 2024/0086249 A1 (hereafter Hu).
Hu has priority to application No. PCT/CN21/96924 filed on 05/28/2021.
Regarding claim 8, the combination teaches “The training system of claim 1, further comprising a resource manager (Choi [0074]) configured to respectively allocate the plurality of processing units to the plurality of training jobs corresponding to a plurality of neural network models (Capes [0059-0063])…”
The combination teaches executing queued jobs in multiple phases, however does not explicitly teach of reallocating processing units for queued job delays.
Hu teaches an elastic training model such that it teaches the limitation “wherein the resource manager is configured to allocate a processing unit to each of the plurality of training jobs and reallocate the processing unit to a queuing training job ([0066] teaches a job queue and a resource allocator allocating based on ETC of jobs received) when a queuing time of a queuing training job is greater than an expected increase time when one or more training jobs are executed in multiple phases, in accordance with presence of the queuing training job stored in a queue. ([0101-0105] teaches dynamically allocating nodes to a job once a job has completed execution, the resource allocator determines whether to allocate the node to another ongoing job or to a queued job. Jobs waiting in the queue during the generation of the allocation sequence will not be allocated a node as such it is still within an expected time frame of job execution. However, if the allocator determines that an expected ongoing job completes within a current time horizon, it would generate a sequence to allocate to the job waiting at the beginning of the queue as that job would have waited the longest amount of time such that it may be greater than the expected time, so that it must be satisfied according to the policy)”.
It would have been obvious to a person of ordinary skill in the art before the effective
filing date to combine the teachings of Hu to the combination of Choi and Capes in order to dynamically allocate resources to training jobs in a queue based on estimated completion time. A person having ordinary skill in the art would have been motivated to adopt this strategy in order to improve efficient utilization of computing resources, speed up the overall training time, and reduce queueing delay (Hu [0010]).
Regarding claim 19, it is similar to claim 8 and is rejected for the same reasons.
Claims 7 and 18 is/are rejected under 35 U.S.C. 103 as being unpatentable over Choi and Capes as used in claims 1, 5, 13, and 17 above, and in further view of Acharya et al. Pat. No. US 6,502,062 B1 (hereafter Acharya).
Regarding claim 7, the combination teaches “The training system of claim 1, further comprising a resource manager (Choi [0074] teaches resource manager) configured to respectively allocate the plurality of processing units to the plurality of training jobs corresponding to a plurality of neural network models (Capes [0059-0063] teaches allocating processing units to a job). … and allocate a remaining processing unit to a training job to which less processing units than required processing units are allocated in accordance with presence of the remaining processing unit in the cluster (Choi [0073-0074] teaches that when resources remain in a cluster, the resource manager may adjust the number of virtual workers used based on a resource state in order to efficiently use many resources remaining on the cloud such that scheduling virtual workers would require processing unit resources to be used).”
The combination does not explicitly teach of allocating processing units to a shortest remaining service time job.
Acharya teaches shortest remaining processing time such that it teaches the limitations “…wherein the resource manager is further configured to allocate a processing unit present in a cluster in an order from a training job having a shortest remaining service time ([Col. 2, lines 29-41] teaches using the Shortest Remaining Processing Time algorithm which minimizes the time it takes to process uncompleted jobs by scheduling shortest remaining job times in a continuous stream)”.
It would have been obvious to a person of ordinary skill in the art before the effective
filing date to combine the teachings of Acharya to the combination of Choi and Capes to modify the resource allocation policy to allocate available processing units according to a shortest remaining processing time (SRPT) scheduling policy. A person having ordinary skill in the art would have been motivated to adopt this strategy in order to improve overall resource utilization, reducing average job completion time, and improving responsiveness in a shared computing environment while retaining the dynamic allocation of remaining cluster resources.
Regarding claim 18, it is similar to claim 7 and is rejected for the same reasons.
Conclusion
Any inquiry concerning this communication or earlier communications from the examiner should be directed to BRANDON A NGUYEN whose telephone number is (571)272-6074. The examiner can normally be reached Mon-Fri (10am-6pm).
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, Aimee Li can be reached at (571) 272-4169. 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.
/BRANDON NGUYEN/Examiner, Art Unit 2195
/PIERRE VITAL/Supervisory Patent Examiner, Art Unit 2198
Prosecution Insights