Distributed job scheduling in SCI local-area multiprocessors

Local Area Multiprocessors (LAMP) is a network of personal workstations with distributed shared physical memory provided by high performance technologies such as SCI. LAMP is more tightly coupled than the traditional local area networks (LAN) but is more loosely coupled than the bus based multiprocessors. The paper presents a distributed scheduling algorithm which exploits the distributed shared memory in SCI-LAMP to schedule the idle remote processors among the requesting workstations it considers fairness by allocating remote processing capacity to the requesting workstations based on their priorities according to the decay-usage scheduling approach. The performance of the algorithm in scheduling both sequential and parallel jobs is evaluated by simulation. It is found that the higher priority nodes achieve faster job response times and higher speedups than that of the lower priority nodes. Lower scheduling overhead allows finer granularity of remote processors sharing than in LAN.