Enabling persistent memory for serverless applications
A scheduling platform for scheduling serverless application tasks in persistent memory (PMEM) is provided. A profiler receives application requests from processes of serverless applications. The profiler categorizes the processes as persistent or non-persistent based on the application requests. A read/write batcher creates batches of the persistent requests including the read requests and write requests and assigns the batches to persistent memory banks. A scheduler creates a schedule of the batches to the persistent memory banks in a manner enabling optimization of job completion time.
1. A system comprising:
a processor; and
a non-transitory storage medium storing instructions executable on the processor to:
receive application requests including read and write requests from processes of a serverless application;
profile the processes of the serverless application as one of persistent processes that issue input/output (I/O) requests to persistent memory;
create batches of the application requests including read batches of the read requests and write batches of the write requests for the persistent processes;
identify a number of excess read batches based on comparing a quantity of the read batches to a quantity of the write batches;
schedule the excess read batches first to the persistent memory before a remainder of the read batches and the write batches, the remainder of the read batches excluding the excess read batches; and
schedule the remainder of the read batches and the write batches to the persistent memory based on an objective function relating to a completion time.
2. The system of claim 1 , wherein the scheduling of the read requests and the write requests from the persistent processes is performed by a first scheduler, and the instructions are executable on the processor to:
profile further processes of the serverless application as non-persistent processes that issue I/O requests to non-persistent memory; and
pass further requests of the non-persistent processes to a second scheduler different from the first scheduler.
3. The system of claim 1 , wherein the instructions are executable on the processor to:
profile the processes of the serverless application as the persistent processes based on determining that a quantity of write requests in a write batch of write requests exceeds a quantity threshold.
4. The system of claim 1 , wherein the number of excess read batches is a quantity of the read batches that exceeds the quantity of the write batches.
5. The system of claim 1 , wherein the instructions are executable on the processor to:
profile the processes of the serverless application as the persistent processes based on examining kernel system calls in the system.
6. The system of claim 1 , wherein the scheduling of the remainder of the read batches and the write batches is according to a time constraint specifying that processing times consumed for the remainder of the read batches and the write batches are proportional to read and write demands of the serverless application.
7. The system of claim 1 , wherein the instructions are executable on the processor to:
implement a fairness algorithm balancing time consumed on the read requests and the write requests for the persistent processes.
8. The system of claim 1 , wherein the objective function seeks to reduce bus switching latency due to switching between different persistent memories.
9. The system of claim 1 , wherein the objective function is further based on a read-to-write latency and a write-to-read latency.
10. The system of claim 1 , wherein the instructions are executable on the processor to:
profile the processes of the serverless application as the persistent processes based on an I/O workload requirement of the serverless application.
11. A method comprising:
receiving, at a system comprising a hardware processor, application requests from processes of a serverless application, the application requests including read requests and write requests;
profiling, by the system, the processes of the serverless application as persistent processes that issue input/output (I/O) requests to persistent memory;
creating, by the system, batches of the application requests including read batches of the read requests and write batches of the write requests for the persistent processes;
identifying, by the system, a number of excess read batches based on comparing a quantity of the read batches to a quantity of the write batches;
scheduling, by the system, the excess read batches first to the persistent memory before a remainder of the read batches and the write batches, the remainder of the read batches excluding the excess read batches; and
scheduling, by the system, the remainder of the read batches and the write batches to the persistent memory based on an objective function relating to a completion time.
12. The method of claim 11 , wherein the scheduling of the read requests and the write requests from the persistent processes is performed by a first scheduler, and the method further comprises:
profiling, by the system, further processes of the serverless application as non-persistent processes that issue I/O requests to non-persistent memory; and
passing, by the system, further requests of the non-persistent processes to a second scheduler different from the first scheduler.
13. The method of claim 11 , wherein the profiling comprises profiling the processes of the serverless application as the persistent processes based on determining that a quantity of write requests in a write batch of write requests exceeds a quantity threshold.
14. The method of claim 11 , wherein the number of excess read batches is a quantity of the read batches that exceeds the quantity of the write batches.
15. The method of claim 11 , wherein the profiling comprises profiling the processes of the serverless application as the persistent processes based on examining kernel system calls in the system.
16. The method of claim 11 , wherein the scheduling of the remainder of the read batches and the write batches is according to a time constraint specifying that times consumed for processing the read batches and the write batches are proportional to read and write demands of the serverless application.
17. The method of claim 11 , further comprising implementing a fairness algorithm balancing time consumed on the read requests and the write requests for the persistent processes.
18. The method of claim 11 , wherein the objective function seeks to reduce bus switching latency due to switching between different persistent memories.
19. A non-transitory computer-readable medium comprising instructions that upon execution cause a system:
receive application requests including read requests and write requests from processes of a serverless application;
profile the processes of the serverless application as persistent processes that issue input/output (I/O) requests to persistent memory;
create batches of the application requests including read batches of the read requests and write batches of the write requests for the persistent processes;
identify a number of excess read batches based on comparing a quantity of the read batches to a quantity of the write batches;
schedule the excess read batches first to the persistent memory before a remainder of the read batches and the write batches, the remainder of the read batches excluding the excess read batches; and
schedule the remainder of the read batches and the write batches to the persistent memory based on an objective function relating to a completion time.
20. The non-transitory computer-readable medium of claim 19 ,
wherein the profiling comprises profiling the processes of the serverless application as the persistent processes based on examining kernel system calls in the system.