DETAILED ACTION
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 § 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 35 U.S.C. 112 (pre-AIA ), 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(s) 1-14 is/are rejected under 35 U.S.C. 112(b) or 35 U.S.C. 112 (pre-AIA ), 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 applications subject to pre-AIA 35 U.S.C. 112, the applicant), regards as the invention.
Claim 1 (similarly claim 8) recite: “directed acyclic graphic (DAG)”. The examiner is unclear if “DAG” should be “directed acyclic graph” or “directed acyclic graphic”. The specification discloses both as DAG. Therefore, the examiner is unclear if the word “graphic” should interpreted differently from “graph”.
Claim 3 (similarly claim 10) recites the limitation "the multiple functions in the first stage". There is insufficient antecedent basis for this limitation in the claim. The examiner is unclear what multiple function, “the multiple functions in the first stage” is referring to.
Claim 4 (similarly claim 11) recites the limitation "the stages". There is insufficient antecedent basis for this limitation in the claim. The examiner is unclear what stages of the plurality of stages, “the stages” are referring to.
Claims 2-7 and 9-14 are rejected based on rejection of its corresponding dependent claim.
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 (i.e., changing from AIA to pre-AIA ) 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, 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.
Claim(s) 1-14 is/are rejected under 35 U.S.C. 103 as being unpatentable over Fine-Grained Performance and Cost Modeling and Optimization for Faas Applications (Changyuan Lin, IEEE 10/14/2022) (hereafter Changyuan).
As per claim 1, Changyuan teaches:
A method, comprising:
receiving a directed acyclic graphic (DAG) for an application, the DAG comprising a plurality of stages arranged in a series, each stage comprising at least one function; ([Page 180], The proposed analytical model leverages graph algorithms to process structures in the serverless application and covert the workflow into a weighted directed acyclic graph (DAG). Then, the analytical model calculates the weighted average response time of paths in DAG and the average number of invocations of functions, from which derive the output of the model, namely the average values of end-to-end response time and cost. [Page 183 - 184], When multiple functions are chained together and executed in a row, and each function takes the output of the previous function as its input, these functions are constructed as a sequence. Fig. 2 – Fig.6] )
profiling the DAG with a plurality of computer resource allocations to generate an end-to-end (E2E) latency model; ([Page 187], Before any DFS phase, the FaaS application should go through a profiling phase to obtain the firing logs for the given application and configuration. During the profiling phase, the performance and cost of the application are profiled using the modeling algorithm proposed in Section 4. Since a more accurate profile can improve the effectiveness of the optimization algorithm, the number of iterations K should be large enough in the profiling phase, which is 10000 by default in our setting. The algorithm obtains the end-to-end response time, cost, and exit status of the application during the profiling phase. Also, for each transition fired before the execution of the application is complete, the algorithm collects its last firing time, firing delay, and incurred cost into the firing logs each time the transition fires. The firing logs will be used in the DFS phase to obtain the response time and incurred cost of functions and structures for identifying bottlenecks. [Page 191-192], Also, the CSPN model can give accurate distributions of the response time and cost. Thus, the performance and cost modeling and optimization algorithms are a significant improvement over those proposed in our previous work… Later, they introduced SimFaaS [45], which is a performance simulator designed specifically for FaaS platforms to allow single-function performance simulations. Manner et al. [46] proposed a local simulation approach to find the optimal configuration for individual serverless functions. Eismann et al. [47], [48] presented frameworks to predict the response time of serverless functions, cost of serverless workflows, and the optimal memory size of functions using machine learning models.)
generating, based on the E2E latency model, an execution plan comprising optimized computer resource allocations and timing information, the optimized computer resource allocations associated with functions in the DAG and the timing information associated with the stages in the DAG; and ([Page 187], Before any DFS phase, the FaaS application should go through a profiling phase to obtain the firing logs for the given application and configuration. During the profiling phase, the performance and cost of the application are profiled using the modeling algorithm proposed in Section 4. Since a more accurate profile can improve the effectiveness of the optimization algorithm, the number of iterations K should be large enough in the profiling phase, which is 10000 by default in our setting. The algorithm obtains the end-to-end response time, cost, and exit status of the application during the profiling phase. Also, for each transition fired before the execution of the application is complete, the algorithm collects its last firing time, firing delay, and incurred cost into the firing logs each time the transition fires. The firing logs will be used in the DFS phase to obtain the response time and incurred cost of functions and structures for identifying bottlenecks. [Page 191-192], Also, the CSPN model can give accurate distributions of the response time and cost. Thus, the performance and cost modeling and optimization algorithms are a significant improvement over those proposed in our previous work… Later, they introduced SimFaaS [45], which is a performance simulator designed specifically for FaaS platforms to allow single-function performance simulations. Manner et al. [46] proposed a local simulation approach to find the optimal configuration for individual serverless functions. Eismann et al. [47], [48] presented frameworks to predict the response time of serverless functions, cost of serverless workflows, and the optimal memory size of functions using machine learning models. [Page 190], The performance and cost optimization algorithm was first evaluated on a serverless application with six functions and four types of structures, which is illustrated in Fig. 10b. We first profiled the six functions with different viable memory sizes. As response time becomes insensitive with large memory sizes, memory is allocated in larger increments when the memory size is larger while profiling functions. Specifically, the viable memory sizes of functions range from 128 MB to 1,024 MB in 64 MB increments, from 1024 MB to 2,048 MB in 128 MB increments, from 2,048 MB to 4,096 MB in 256 MB increments, and from 4,096 MB in 10,240 MB in 512 MB increments, resulting in 43 viable memory…)
cause an FaaS infrastructure to:
execute a function in a first stage of the DAG model on a first virtual machine, the first virtual machine having a computer resource allocation based on the execution plan;
initialize, a time determined specified by the execution plan, a second virtual machine for a function in a second stage in the DAG model, the second virtual machine having computer resources allocated according to one of the optimized computer resource allocations; and ([Fig. 2-6] [Page 180], Most FaaS solutions are based on lightweight virtualization techniques that provide ephemeral isolated sandboxes, such as containers, unikernels, and Firecracker VMs, which have low operating overhead, fast startup, and scaling speeds [1], [2], [3]. Therefore, serverless architecture can provide cloud-native applications with significant performance, scalability, and availability boosts over traditional monolithic architecture. [Page 191-192], Also, the CSPN model can give accurate distributions of the response time and cost. Thus, the performance and cost modeling and optimization algorithms are a significant improvement over those proposed in our previous work… Later, they introduced SimFaaS [45], which is a performance simulator designed specifically for FaaS platforms to allow single-function performance simulations. Manner et al. [46] proposed a local simulation approach to find the optimal configuration for individual serverless functions. Eismann et al. [47], [48] presented frameworks to predict the response time of serverless functions, cost of serverless workflows, and the optimal memory size of functions using machine learning models. [Page 190], The performance and cost optimization algorithm was first evaluated on a serverless application with six functions and four types of structures, which is illustrated in Fig. 10b. We first profiled the six functions with different viable memory sizes. As response time becomes insensitive with large memory sizes, memory is allocated in larger increments when the memory size is larger while profiling functions. Specifically, the viable memory sizes of functions range from 128 MB to 1,024 MB in 64 MB increments, from 1024 MB to 2,048 MB in 128 MB increments, from 2,048 MB to 4,096 MB in 256 MB increments, and from 4,096 MB in 10,240 MB in 512 MB increments, resulting in 43 viable memory…)
execute the function in the second stage on the second virtual machine after completion of the function in the first stage. ([Fig. 2-6] [Page 180], Most FaaS solutions are based on lightweight virtualization techniques that provide ephemeral isolated sandboxes, such as containers, unikernels, and Firecracker VMs, which have low operating overhead, fast startup, and scaling speeds [1], [2], [3]. Therefore, serverless architecture can provide cloud-native applications with significant performance, scalability, and availability boosts over traditional monolithic architecture. [Page 183], The most common method is to generate a sequence based on the execution logs or trace files of a serverless application deployed in the production environment, which record the sequence of actual events that happen within the application, the frequency distribution of invocations for each function, and the real inputs of functions. [Page 184], Each choice has multiple branches, which are sequences composed of functions and structures. Only one branch (sequence) in the choice will receive the input and be executed depending on the satisfied condition, which can be described using probabilities defined by business logic and derived from execution logs or empirical knowledge. Fig. 4 demonstrates the rules for modeling the choice structure with CSPNs. Two conditions are equivalent to two state transitions between the places that represent the start state of the branches and the following action in different branches, whose input and output are one token. Namely, we have Iðt2Þ ¼ Iðt3Þ ¼ 0½p2 7! 1, Oðt2Þ ¼ 0½p3 7! 1, and Oðt3Þ ¼ 0½p5 7! 1. As it may take some time to verify conditions, t2 and t3 may have a probabilistic firing delay.)
Although Changyuan silently discloses stages are executed via virtual machines by stating most FaaS solutions are based on containers, VMs. [Page 180].
Changyuan does not explicitly discloses execute a function in a first stage of the DAG model on a first virtual machine, the first virtual machine having a computer resource allocation based on the execution plan; initialize, a time determined specified by the execution plan, a second virtual machine for a function in a second stage in the DAG model, the second virtual machine having computer resources allocated according to one of the optimized computer resource allocations.
Haghighat teaches execute a function in a first stage of the DAG model on a first virtual machine, the first virtual machine having a computer resource allocation based on the execution plan; initialize, a time determined specified by the execution plan, a second virtual machine for a function in a second stage in the DAG model, the second virtual machine having computer resources allocated according to one of the optimized computer resource allocations.([Paragraph 5], As shown in FIG. 2A, the serverless service platform 203 has a serverless services manager 203a to receive the serverless function code 201, and store the serverless function code 201 to a storage, and schedule the execution of the serverless function code 201 using required computer resources such as containers. [Paragraph 45], FIG. 15B is a flowchart of memory allocation for containers and functions of a FaaS platform according to an embodiment; [Paragraph 154], Containers often run within virtual machines (VMs) for security and isolation. [Paragraph 407], Furthermore, the invoker and batch balancer 1606 may identify latency constraints of particular functions associated with function requests 1604 and distribute the function requests 1604 to meet the latency constraints. A latency constraint may correspond to a requested time by which the function should complete execution. For example, shorter latency functions may be grouped together when the latency constraints permit and ordered such that all of the latency constraints for every function in the group are satisfied. T )
Haghighat also teaches stages ([Paragraph 1032], Turning now to FIG. 41C, according to an exemplary embodiment, in block 4120 of method 4100, resource characteristics may be associated with executed functions to generate demand fingerprints of the executed functions at each stage of execution. In block 4130, an orchestrator (not shown), for example, may generate detailed reports on the usage of different resources at multiple stages of execution of the functions. In block 4140, the execution of functions may be scheduled, and resources may be allocated to the functions based on the generated reports. [Paragraph 422], In some embodiments, some of the batchable functions may be executed simultaneously, and others of the batchable functions may be executed serially. In some embodiments, the container 1614a may execute a first group of the batchable functions, for example, a maximum number of batchable functions that may be supported by the resources of the container 1614a. After the first group complete execution, a second group of the batchable functions may begin execution, for example a maximum number of batchable functions that may be supported by the resources of the container 1614a. Thus, the invoker and batch balancer 1606 may schedule a hybrid serial and parallel execution of the batchable functions in the container 1614a based on security protocols and resource availability of the container 1614a relative to resource requirements of the functions. )
It would have been obvious to a person with ordinary skill in the art, before the effective filing date of the invention, to combine the teachings of Changyuan wherein a directed acyclic graphic (DAG) is received for an application with plurality of stages with function(s), DAG is profiled to improve effectiveness based on end-to-end latency model, execution plan is generated with optimized resource allocations and timing information, the plurality of stages and corresponding functions are executed via virtual machine(s), containers and unikernels, into teachings of Haghighat wherein virtual machine(s) is/are used to execute containers, because this would enhance the teachings of Changyuan wherein by utilizing virtual machine(s) to execute stages of functions, it provides added isolation/security. [Haghighat paragraph 154, 407].
As per claim 2, rejection of claim 1 is incorporated:
Changyuan teaches wherein generating, based on the E2E latency model, the execution plan further comprises: generating bundling information associating multiple functions in a stage with a computer resource allocation for a virtual machine that will execute all of the multiple functions of the stage in parallel. ([Fig. 3-6] [Page 183 – 184], Parallels introduce parallelism (fan-out pattern) into serverless applications. The parallel is a structure that allows functions (actions) in multiple sequences to be executed in parallel. [Page 192], Manner et al. [46] proposed a local simulation approach to find the optimal configuration for individual serverless functions. Eismann et al. [47], [48] presented frameworks to predict the response time of serverless functions, cost of serverless workflows, and the optimal memory size of functions using machine learning models. They focused on predicting the optimal memory size for each individual function without considering the application-level constraints (i.e., end-to-end response time and cost).)
Haghighat teaches all of the multiple functions of the stage in parallel. ([Paragraph 2], FaaS may therefore be considered as a step in the evolution of cloud computing. Sometimes also referred to as “Serverless Computing,” FaaS may enable software developers to write highly-scalable code but without the cost, time, and expense required to provision or otherwise pre-define the hardware or application software resources that execution of the code will involve. FaaS may also enable Cloud Service Providers (CSPs) to increase resource usage due to better allocation (e.g., bin packing). [Paragraph 422], In some embodiments, the container 1614a may execute a first group of the batchable functions, for example, a maximum number of batchable functions that may be supported by the resources of the container 1614a. [Paragraph 412], . Parallel execution of multiple batched functions in a single container may further reduce latency.)
As per claim 3, rejection of claim 2 is incorporated:
Haghighat teaches executing, based on the bundling information, the multiple functions in the first stage in parallel. ([Paragraph 2], FaaS may therefore be considered as a step in the evolution of cloud computing. Sometimes also referred to as “Serverless Computing,” FaaS may enable software developers to write highly-scalable code but without the cost, time, and expense required to provision or otherwise pre-define the hardware or application software resources that execution of the code will involve. FaaS may also enable Cloud Service Providers (CSPs) to increase resource usage due to better allocation (e.g., bin packing). [Paragraph 422], In some embodiments, the container 1614a may execute a first group of the batchable functions, for example, a maximum number of batchable functions that may be supported by the resources of the container 1614a. [Paragraph 412], . Parallel execution of multiple batched functions in a single container may further reduce latency.)
As per claim 4, rejection of claim 1 is incorporated:
Changyuan teaches wherein profiling the DAG with a plurality of computer resource allocations to generate an end-to-end (E2E) latency model comprises:
generating a plurality of latency distributions for each function in the DAG, the latency distributions representing initialization and execution time of each function in the DAG executed using the computer resource allocations, respectively; and
generating, based on the latency distributions for each function, an end-to-end latency model which models the end-to-end latency for executing the stages of the DAG model using the computer resource allocations. ([Page 181], We design a performance and cost modeling algorithm based on CSPNs sampling, which enables the fine-grained estimation of the distribution of the response time, cost, and exit status of FaaS applications. [Page 185], Therefore, the initial marking of CSPNs modeling FaaS applications should be m0 ¼ 0½pS 7! 1, and the age of the token should be zero in most cases, namely UðpSÞ ¼ f0g, where pS is the start state. In case there is initial latency (e.g., initialization latency/scheduling overhead incurred by the orchestration service), In this section, we propose the algorithm for performance and cost modeling of FaaS applications, which could obtain the distribution of the end-to-end response time and cost by simulating CSPNs and sampling delayed transitions. Then, we evaluate the efficiency of the proposed algorithms.)
Haghighat also teaches ([Paragraph 195], In the illustrated example, the events 700 include time executed, instructions per cycle (IPC), memory bandwidth, cache usage, I/O operations per second (IOPs) and network bandwidth, although other events may also be monitored/collected. [Paragraph 237], Some embodiments may advantageously provide container run ahead speculative execution. Some functions involve long latency/startup, which can slow down execution. Some embodiments may provide a run ahead execution mechanism to fetch data/instruction streams at the processor/core level, and also to reserve and/or reallocate resources. Advantageously, some embodiments may reduce latency/startup for functions that can take advantage of the run ahead capability. [Paragraph 2], Function as a Service (FaaS) is an event-oriented highly-scalable computer code execution model that typically provisions a single purpose application programming interface (API) endpoint on a cloud computing infrastructure to receive and run the code execution requests for a small amount of time. Such code execution requests and/or executions of requested code are variously and commonly referred to as lambdas, functions, actions, and/or run-to-completion procedures.)
As per claim 5, rejection of claim 3 is incorporated:
Changyuan teaches wherein generating the plurality of latency distributions for each function in the DAG further comprises: measuring the latency of executing each function in the DAG with each of the computer resource allocations. ([Page 181], We design a performance and cost modeling algorithm based on CSPNs sampling, which enables the fine-grained estimation of the distribution of the response time, cost, and exit status of FaaS applications. [Page 185], This is also applicable to updating memory sizes and marking optimized functions/structures in a map when optimizing applications, which will be presented in Section 5.2. Therefore, the initial marking of CSPNs modeling FaaS applications should be m0 ¼ 0½pS 7! 1, and the age of the token should be zero in most cases, namely UðpSÞ ¼ f0g, where pS is the start state. In case there is initial latency (e.g., initialization latency/scheduling overhead incurred by the orchestration service), In this section, we propose the algorithm for performance and cost modeling of FaaS applications, which could obtain the distribution of the end-to-end response time and cost by simulating CSPNs and sampling delayed transitions. Then, we evaluate the efficiency of the proposed algorithms.)
As per claim 6, rejection of claim 1 is incorporated:
Changyuan teaches wherein generating, based on the E2E latency model, the execution plan comprising optimized computer resource allocations and timing information further comprises:
determining, based on the E2E latency model, optimum delay times, expressed from the start of the application, to begin virtual machine initialization for each stage of the DAG, wherein the optimum delay times are the highest delay times possible without increasing E2E latency. ([Page 181], We design a performance and cost modeling algorithm based on CSPNs sampling, which enables the fine-grained estimation of the distribution of the response time, cost, and exit status of FaaS applications. [Page 185], This is also applicable to updating memory sizes and marking optimized functions/structures in a map when optimizing applications, which will be presented in Section 5.2. Therefore, the initial marking of CSPNs modeling FaaS applications should be m0 ¼ 0½pS 7! 1, and the age of the token should be zero in most cases, namely UðpSÞ ¼ f0g, where pS is the start state. In case there is initial latency (e.g., initialization latency/scheduling overhead incurred by the orchestration service), In this section, we propose the algorithm for performance and cost modeling of FaaS applications, which could obtain the distribution of the end-to-end response time and cost by simulating CSPNs and sampling delayed transitions. Then, we evaluate the efficiency of the proposed algorithms.)
Haghighat also teaches ([Paragraph 282], In selecting the thresholds for launching Bm or Bh (e.g., as respectively the utilization Bs or Bm rises), and in selecting the thresholds for reclaiming Bh or Bm (e.g., as utilization of Bh and Bm respectively declines), some embodiments may take into account service level agreement (SLA) inputs provided dynamically by an AFaaS control service. If no SLA input is provided, then these thresholds may be set heuristically and dynamically, based on the arrival rates for requests for B (e.g., or a moving window average of the arrival rates). [Paragraph 637], FIG. 22 illustrates an exemplary architecture for providing determinism and accuracy in a FaaS environment. In FIG. 22, a computing device, e.g., performance controller 2210, may perform operations “a” and “b” in the paragraph immediately above. The computing device 2210 may also provide history-based resource scheduling, redirect data into appropriate cores for invoking, maximize data sharing to and minimize data movement, bundle functions according to service level agreements. [Paragraph 995], Functions may be maintained locally if latency SLA requires milliseconds (ms) latency. An exemplary embodiment may guarantee staying below some threshold latency and may provide for construction of links that can assure such a threshold latency [Paragraph 260], A conventional FaaS invocation may involve a number of triggers, including some final trigger which invokes a function. After the function is invoked, the work is dispatched to a platform with some containers (e.g., newly started, warmed up, etc.). But there may be too few resources used to support the new invocation. Piled up latencies may delay execution of the function. [Paragraph 325], FIG. 13B illustrates an enhanced scheduling process 1338 for more fair bin packing that may be implemented by the server 1302 of FIG. 13A. In the present example, band 1 and band 2 for each of the function 0 and function 1 may be allocated hardware resources.)
As per claim 7, rejection of claim 1 is incorporated:
Changyuan teaches wherein before executing the function in the first stage of the DAG on the first virtual machine, the method further comprises: allocating storage on the first virtual machine according to the VM size information in the execution plan. ([Page 191-192], Also, the CSPN model can give accurate distributions of the response time and cost. Thus, the performance and cost modeling and optimization algorithms are a significant improvement over those proposed in our previous work… Later, they introduced SimFaaS [45], which is a performance simulator designed specifically for FaaS platforms to allow single-function performance simulations. Manner et al. [46] proposed a local simulation approach to find the optimal configuration for individual serverless functions. Eismann et al. [47], [48] presented frameworks to predict the response time of serverless functions, cost of serverless workflows, and the optimal memory size of functions using machine learning models.)
Haghighat also teaches ([Paragraph 45], FIG. 15B is a flowchart of memory allocation for containers and functions of a FaaS platform according to an embodiment; [Paragraph 154], Containers often run within virtual machines (VMs) for security and isolation. [Paragraph 945], According to an exemplary embodiment of the enhanced FaaS system such as the one shown in FIG. 4, both the size and the usage frequency of a function may be considered, in deciding whether a function should have its data or code in high performance memory tier. In other embodiments, this may be based on usage: keep less used function alive in a lower performance tier (like 3DXP memory) instead of reclaiming it altogether. According to one embodiment of the enhanced FaaS system with tiering in memory, tiering in memory may be made adaptive to both size and usage. That is, if size is large, then, even if a function is frequently used, the function may be put in a far-away tier like in 3DXP. If usage is low, then also the function may be put in a far-away tier. Further, if it is in a farther tier, then the function should be provided a correspondingly longer time to live since it is already consuming less resources.)
As per claims 8-14, these are system claims corresponding to the method claims 1-7. Therefore, rejected based on similar rationale.
Conclusion
The prior art made of record and not relied upon is considered pertinent to applicant's disclosure.
Beare (Pub 20220075660) discloses end-to-end latency of DAG pipeline.
Any inquiry concerning this communication or earlier communications from the examiner should be directed to DONG U KIM whose telephone number is (571)270-1313. The examiner can normally be reached 9:00am - 5:00pm.
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, Bradley Teets can be reached at 5712723338. 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.
/DONG U KIM/Primary Examiner, Art Unit 2197
Prosecution Insights