Edge domain-specific accelerator virtualization and scheduling
Presented herein are embodiments to implement a temporal queueing system with class-based fair queuing and dynamic resource allocation based on a novel look-ahead capability to manage various models and workloads for utilization/efficiency improvements. Embodiments may be implemented to allocate accelerator resources based on platform-defined timeslots, and therefore significantly increase the ability of workloads to access hardware accelerator resources. Training and inference may be supported with flexible preemption and the ability to support run-to-completion for training tasks while still supporting non-run-to-completion for inference tasks. Embodiments may be implemented by an edge software operation platform through virtual accelerators to allow emulation of different types of hardware accelerators and to map to the hardware accelerators with hardware-specific procedures managed by an edge orchestrator and an edge endpoint. Accordingly, embodiments of the present disclosure reduce the requirements for the workload to manage platform capacity and hardware.
1 . A processor-implemented method for edge domain-specific accelerator (DSA) virtualization and scheduling comprising:
configuring, by an edge orchestrator (EO), a virtual accelerator to virtualize one or more DSAs in an edge endpoint;
associating the virtual accelerator with an application to be executed at the edge endpoint; and
assigning, by the EO, one or more resource utilization parameters to the virtual accelerator for executing the application at the edge endpoint using allocated DSA resources in a timeslot scheduled using a time division queuing.
2 . The processor-implemented method of claim 1 wherein the virtual accelerator virtualizes the one or more DSAs into a single resource pool at the edge endpoint.
3 . The processor-implemented method of claim 1 wherein the timeslot is allocated in the time division queuing with elastic dynamic allocation.
4 . The processor-implemented method of claim 3 wherein the time division queuing is realized as a class-based weighted fair queuing (CBWFQ) with strict priority queuing (SPQ).
5 . The processor-implemented method of claim 1 wherein the one or more resource utilization parameters comprise one or more of:
a minimum resource utilization required for task execution;
a maximum streaming multiprocessor (SM) utilization limit; and
a mean resource utilization as a target average utilization.
6 . The processor-implemented method of claim 1 wherein the timeslot is obtained based at least on a mean normalized accelerator unit (NAU) from a resource normalization framework.
7 . The processor-implemented method of claim 1 wherein the allocated DSA resources are determined based on one or more of:
a priority specified by a customer through a service plan/manifest;
a power consumption specific to edge deployment;
a cost for cloud domains;
an accelerator resource requirement estimated from an application resource uncertainty estimation process;
parameters related to streaming multiprocessor (SM)/logic block (LB) execution in real-time;
a task/job category as run-to-completion (RTC) or non-RTC (NRTC); and
a task/job category as preemptable or non-preemptable.
8 . A processor-implemented method for edge domain-specific accelerator (DSA) virtualization and scheduling comprising:
given an application to be executed in an edge endpoint, the application being associated with a virtual accelerator that is configured to virtualize one or more physical accelerators in the edge endpoint into a single resource pool:
implementing, by a queuing scheduler, a temporal queuing with time slicing to allocate the virtual accelerator a timeslot;
during the allocated timeslot, loading one or more models and workloads for the application into a memory of the physical accelerators for application execution; and
executing the application in the allocated timeslot using the one or more models and workloads.
9 . The processor-implemented method of claim 8 wherein the timeslot is allocated using a time division queuing with elastic dynamic allocation.
10 . The processor-implemented method of claim 9 wherein the time division queuing is realized as a class-based weighted fair queuing (CBWFQ) with strict priority queuing (SPQ).
11 . The processor-implemented method of claim 10 wherein timeslot allocation for a class is managed dynamically based on one or more of:
a category of job/task of the class as run-to-completion (RTC) or non-RTC (NRTC); and
a category of job/task of the class as preemptable or non-preemptable.
12 . The processor-implemented method of claim 8 wherein the allocated timeslot is obtained based at least on a mean normalized accelerator unit (NAU) from a resource normalization framework.
13 . The processor-implemented method of claim 8 further comprising:
removing the one or more models from the memory by an end of the allocated timeslot such that the physical accelerators are ready for executing another application at a next timeslot.
14 . The processor-implemented method of claim 8 further comprising:
responsive to the virtual accelerator not being able to submit all data to be completed during the allocated timeslot, queuing the workload at the virtual accelerator until a next timeslot cycle.
15 . A non-transitory computer-readable medium or media comprising one or more sequences of instructions which, when executed by at least one processor, cause steps to be performed comprising:
configuring a virtual accelerator to virtualize one or more domain-specific accelerators (DSAs) in an edge endpoint;
associating the virtual accelerator with an application to be executed at the edge endpoint; and
assigning one or more resource utilization parameters to the virtual accelerator for executing the application at the edge endpoint using allocated DSA resources in a timeslot scheduled using a time division queuing.
16 . The non-transitory computer-readable medium or media of claim 15 wherein the one or more DSAs are virtualized into a single resource pool at the edge endpoint.
17 . The non-transitory computer-readable medium or media of claim 15 wherein the timeslot is allocated in the time division queuing with elastic dynamic allocation.
18 . The non-transitory computer-readable medium or media of claim 17 wherein the time division queuing is realized as a class-based weighted fair queuing (CBWFQ) with strict priority queuing (SPQ).
19 . The non-transitory computer-readable medium or media of claim 15 wherein timeslot allocation for the application is based at least on a category of the application as a run-to-completion (RTC) task, which is generally non-preemptable, or a non-RTC (NRTC) task, which is generally preemptable.
20 . The non-transitory computer-readable medium or media of claim 15 wherein the timeslot is obtained based at least on a mean normalized accelerator unit (NAU) from a resource normalization framework.