Implicit coscheduling

Abstract

In modern distributed systems, coordinated time-sharing is required for communicating processes to leverage the performance of switch-based networks and low-overhead protocols. Coordinated time-sharing has traditionally been achieved with gang scheduling or explicit coscheduling, implementations of which often suffer from many deficiencies: multiple points of failure, high contextswitch overheads, and poor interaction with client-server, interactive, and I/O-intensive workloads. Implicit coscheduling dynamically coordinates communicating processes across distributed machines without these structural deficiencies. In implicit coscheduling, no communication is required across operating system schedulers; instead, cooperating processes achieve coordination by reacting to implicit information carried by communication existing within the parallel application. The implementation of this approach is simple and allows participating nodes to act autonomously. We introduce two key mechanisms in implicit coscheduling. The first is conditional two-phase waiting, a generalization of traditional two-phase waiting in which spin-time may be increased depending upon events occuring while the process waits. The second is an extension to stride scheduling that provides preemption and is fair to processes that block. To demonstrate that implicit coscheduling performs well, we show results from an extensive set of simulation and implementation experiments. To exercise the conditional two-phase waiting algorithm, we examine three workloads: bulk-synchronous and continuous-communication synthetic applications and application kernels written in the Split-C language. To exercise the local scheduler, we examine competing jobs with different communication characteristics. We demonstrate that our implementation scales well with the number of jobs and workstations and is robust to process placement. Our experiments show that implicit coscheduling is effective and fair for a wide range of workloads; most perform within 30% of an idealized model of gang scheduling.

Dynamic Fractional Resource Scheduling for cluster platforms

“…Flexible coscheduling monitors network activity in order to classify processes and then only uses gang-scheduling for those tasks that seem to require it [46,55]. Implicit co-scheduling is based on the idea of two-phase waiting: processes which are blocking for communication events spin in place for a period of time, usually a little over the time required for two context-switches, before finally blocking and yielding the processor [56,57]. Dynamic co-scheduling boosts the priority level of processes receiving communications, on the theory that this will keep parallel applications loaded across multiple nodes during periods of intense interactive communication [58].…”

Flexible coscheduling: mitigating load imbalance and improving utilization of heterogeneous resources

“…We also compare to gang scheduling (GS) and spin-block (SB). SB is very similar to implicit coscheduling (ICS) [1], and represents an effective way to time-share a machine without global coordination as in gang scheduling. With SB, processes that wait for synchronous communication poll for a given interval, and only if the communication has not completed by this time they block (in contrast, gang and locally-scheduled processes always busy-wait).…”

“…This approach has the appealing advantage that it does not require any changes to existing parallel software, and is therefore able to deal with existing legacy codes. For example, coscheduling algorithms such as Implicit Coscheduling (ICS) [1] can potentially alleviate load imbalance and increase resource utilization. However, ICS is not always able to handle all job types because it cannot rely on global coordination.…”

“…We evaluate FCS by running different mixes of test programs on a 32-node/64-processor cluster using different schedulers. Specifically, we compare the performance achieved by FCS with that of first-come-first-served batch scheduling (FCFS), gang scheduling (GS), local scheduling with busy waiting, and local scheduling with spin blocking (SB), which is very similar to ICS [1] in this context. Different schedulers provide the best performance for different job mixes.…”

“…), as well as to use DC processes to fill in the gaps that F processes create because of their synchronization problems. An F process that waits for pending communication should not block immediately, but rather spin for some time to avoid unnecessary context switch penalties, as in ICS [1].…”