상세정보
학술대회
Improvement of Worker Scaling-based Scheduling Algorithm to Efficiently Respond to Explosion of Micro Function Service Requests
- 발행일
- 202102
- 출처
- International Conference on Electronics, Information and Communication (ICEIC) 2021, pp.598-601
- 협약과제
- 20JS1800, 초병렬 프로세서 기반 고집적 컴퓨팅 노드 및 시스템 개발, 한우종
- 초록
- We present improved worker scheduler algorithm and architecture to minimize the scheduling cost and function service response time of FaaS platform. We try to solve the problem of performance degradation of previous scaling-based worker scheduling algorithm due to the explosion of the function invocation requests and the blocking-based communication channel thread with the computation node. To this end, two algorithms including Load-based function worker demand prediction algorithm and adoption of asynchronous communication channel integration and non-blocking allocation method have been developed. Through the performance evaluation of proposed worker scheduling algorithm on FaaS worker scheduler, the proposed worker scheduler achieves speedups as high as 10.1 and 16.8 compared to the baseline Bare-metal Computing nodes with the CPU and GPU devices using the in-house TSL Inferencing application.
- KSP 제안 키워드
- Allocation method, Asynchronous communication, Bare Metal, Demand prediction, Performance evaluation, Scheduling algorithm, Scheduling cost, Service requests, Service response time, channel integration, communication channel