A closer look at coscheduling approaches for a network of workstations

Design and Implementation of a Contention-Aware Coscheduling Strategy on Multi-Programmed Heterogeneous Clusters

“…Periodic Boost (PB) [12], [19] is an alternative coscheduling scheme to avoid expensive interrupt costs. In PB, the receiver is busy-waiting in send and receive operations, as with DCS; however, rather than raising an interrupt for each incoming message, a kernel thread periodically examines message queues of each process in a round-robin way and boosts the priority of one of the processes with unconsumed (or pending) messages.…”

“…Two main strategies for coordinating individual local schedulers have been proposed: gang scheduling (GS) [8], [14] and communication-driven coscheduling (CDC) [2]- [5], [11], [12], [15], [19]. GS uses explicit global synchronization to schedule all the processes of a job simultaneously.…”

“…While GS is efficient for fine-grained parallel jobs, it can suffer from resource fragmentation when jobs are not well-balanced and/or do not make use of all nodes [17]. Recently, a full appreciation of these practical limitations of GS has led to a myriad of CDC schemes such as DCS [15] SB [5], PB [12], [19], and HYBRID [4]. These schemes use inherent communication events of parallel processes to approximately guide the local schedulers toward coscheduled execution.…”

ClusterSchedSim: A Unifying Simulation Framework for Cluster Scheduling Strategies

“…• Dynamic Co-Scheduling Schemes Dynamic co-scheduling schemes are proposed to further boost the system utilization [5,1,21,15]. These schemes allocate multiple tasks (from different jobs) to a node, and leave the temporal scheduling of that node to its local CPU scheduler.…”

Buffered coscheduling: a new methodology for multitasking parallel jobs on distributed systems

“…In this paper, we present a new methodology that conjugates the positive aspects of explicit and implicit coscheduling using three techniques: communication buffering to amortize communication overhead (a technique similar to periodic boost [11]); strobing to globally exchange information at regular intervals; and non-blocking, one-sided communication to decouple communication and synchronization. By leveraging these techniques, we can perform effective optimizations based on the status of the parallel machine rather than on the limited knowledge available locally to each processor.…”