Reference :
- Introduction to Parallel Computing, Edition Second by Vipin Kumar et al
- The Australian National University CECS
Embarrassingly Parallel Problems
- 계산은 각각의 프로세서들에 의한 완전하게 독립전인 실행 부분으로 나뉘어질수 있다
- 각각의 어플리케이션은 처치 곤란 병렬적이다
- 데이터의 분포와 모음이 중요 이슈이다 - 비용이 많이 들거나, 중요하다
- 대부분 마스터/슬레이브 접근방식을 사용한다 (p-1 speedup)
Static Task Assignment
Each process needs to execute the procedure after being given the coordinates of the pixels to compute. Slave sends results back in groups which reduces the number of communication stratup times (for each message). — Note that master receives messages in any order.
Dynamic Task Assignment
Ideally, all processes must be busy; but we do not know the speed of each. Dynamic load-balancing can be achieved by a work-pool which get the works supplied when the process is idle. Each processors request a job when it finished its.
Summary
- tasks do not need to communicate
- non-tirvial but. providing input, aseemble the results, load-balance, scheduling, costing…
- Static (low communication) vs dynamic + overdecomp (better load balance)
- Monte carlo or ensumble simulations are a big use of computational power.
- grid computing arose to solve this issue.
