Context-aware and stateless deep learning autotuning framework
Systems and methods are provided for improving autotuning procedures using stateless processing with a remote key-value store. For example, the system can implement a task launcher, a scheduler, and an agent to launch, schedule, and execute decomposed autotuning stages, respectively. The scheduling policy implemented by the scheduler may perform operations beyond a simple scheduling policy (e.g., a FIFO-based scheduling policy), which produces a high queuing delay. Compared to the traditional systems, by leveraging autotuning specific domain knowledge, queueing delay is reduced and resource utilization is improved.
1. An autotuning computer system for performing stateless autotuning tasks with respect to a machine learning (ML) model, the autotuning computer system comprising:
a memory storing a scheduler circuit, wherein the scheduler circuit operates in accordance with machine executable instructions; and
one or more processors configured to access the memory and execute the machine executable instructions stored to:
receive, from the scheduler circuit, a scheduling request for an autotuning task, wherein the scheduling request comprises an autotuning stage, an autotuning task ID, and parameters;
receive a set of states of the autotuning task from a key-value store by using the autotuning task ID as a key, wherein the set of states of the autotuning task comprise a current stage and an average runtime per stage;
based at least on the average runtime per stage, determine a hardware resource, the hardware resource configured to generate at least one of the states of the autotuning task through execution of the autotuning task by the hardware resource;
load the states of the autotuning task at the hardware resource;
execute, by the hardware resource, the current stage with the loaded states, which generates new states to be used for executing a next stage; and
store the generated new states as values in the key-value store by using the autotuning task ID as the key, wherein decoupling of the states from computation to the key-value store forms the autotuning computer system as stateless.
2. The autotuning computer system of claim 1 , wherein the autotuning computer system is configured to optimize inference performance of the machine learning (ML) model for a particular hardware configuration.
3. The autotuning computer system of claim 1 , wherein the states of the autotuning task is externalized to the key-value store and sub-procedures load the states from the key-value store when the sub-procedures are scheduled.
4. The autotuning computer system of claim 3 , wherein the sub-procedures are executed on different resources simultaneously.
5. The autotuning computer system of claim 1 , wherein the scheduler circuit utilizes shortest job first (SJF) to prioritize the current stage requiring short completion time over the next stage with longer completion time.
6. The autotuning computer system of claim 1 , wherein a Multi-Process Service (MPS) is implemented to isolate computational resources.
7. The autotuning computer system of claim 1 , wherein the hardware resource comprises a set of loop tiles, and the states of the autotuning task configure the set of loop tiles to execute the autotuning task based on one or more of ordering, caching, and loop unrolling.
8. A computer-implemented method for performing stateless autotuning tasks with respect to a machine learning (ML) model, the method comprising:
receiving, from a scheduler circuit, a scheduling request for an autotuning task, wherein the scheduling request comprises an autotuning stage, an autotuning task ID, and parameters;
receiving a set of states of the autotuning task, wherein the set of states of the autotuning task comprise a current stage and an average runtime per stage;
based at least on the average runtime per stage, determining a hardware resource, the hardware resource configured to generate at least one of the states of the autotuning task through execution of the autotuning task by the hardware resource;
loading the states of the autotuning task at the hardware resource;
executing the current stage with the loaded states, which generates new states to be used for executing a next stage; and
storing the generated new states as values in a key-value store by using the autotuning task ID as a key, wherein decoupling of the states from computation to the key-value store forms a computer system corresponding with the computer-implemented method as stateless.
9. The computer-implemented method of claim 8 , wherein inference performance of the machine learning (ML) model is optimized for a particular hardware configuration.
10. The computer-implemented method of claim 8 , wherein the states of the autotuning task are externalized to the key-value store and sub-procedures load the states from the key-value store when the sub-procedures are scheduled.
11. The computer-implemented method of claim 10 , wherein the sub-procedures are executed on different resources simultaneously.
12. The computer-implemented method of claim 8 , further comprising:
receiving, by the scheduler circuit, a registration request with autotuning options.
13. The computer-implemented method of claim 8 , wherein the scheduler circuit utilizes shortest job first (SJF) to prioritize the current stage requiring short completion time over the next stage with longer completion time.
14. The computer-implemented method of claim 8 , wherein a Multi-Process Service (MPS) is implemented to isolate computational resources.
15. A non-transitory computer-readable storage medium storing a plurality of instructions executable by one or more processors, the plurality of instructions when executed by the one or more processors cause the one or more processors to:
receive, from a scheduler circuit, a scheduling request for an autotuning task, wherein the scheduling request comprises an autotuning stage, an autotuning task ID, and parameters;
receive a set of states of the autotuning task, wherein the set of states of the autotuning task comprise a current stage and an average runtime per stage;
based at least on the average runtime per stage, determine a hardware resource, the hardware resource configured to generate at least one of the states of the autotuning task through execution of the autotuning task by the hardware resource;
load the states of the autotuning task at the hardware resource;
execute the current stage with the loaded states, which generates new states to be used for executing a next stage; and
store the generated new states as values in a key-value store by using the autotuning task ID as a key, wherein decoupling of the states from computation to the key-value store forms an autotuning computer system as stateless.
16. The computer-readable storage medium of claim 15 , wherein inference performance of a machine learning (ML) model implemented by the autotuning computer system is optimized for a particular hardware configuration.
17. The computer-readable storage medium of claim 15 , wherein the states of the autotuning task is externalized to the key-value store and sub-procedures load the states from the key-value store when the sub-procedures are scheduled.
18. The computer-readable storage medium of claim 17 , wherein the sub-procedures are executed on different resources simultaneously.
19. The computer-readable storage medium of claim 15 , the plurality of instructions further cause the one or more processors to:
receive, by the scheduler circuit, a registration request with autotuning options.
20. The computer-readable storage medium of claim 15 , wherein the scheduler circuit utilizes shortest job first (SJF) to prioritize the current stage requiring short completion time over the next stage with longer completion time.