Effective distributed scheduling of parallel workloads
Proceedings of the 1996 ACM SIGMETRICS international conference on Measurement and modeling of computer systems
Demand-Based Coscheduling of Parallel Jobs on Multiprogrammed Multiprocessors
IPPS '95 Proceedings of the Workshop on Job Scheduling Strategies for Parallel Processing
Overhead Analysis of Preemptive Gang Scheduling
IPPS/SPDP '98 Proceedings of the Workshop on Job Scheduling Strategies for Parallel Processing
Dynamic Coscheduling on Workstation Clusters
IPPS/SPDP '98 Proceedings of the Workshop on Job Scheduling Strategies for Parallel Processing
Implementing MPI under AP/Linux
MPIDC '96 Proceedings of the Second MPI Developers Conference
An Adaptive Submesh Allocation Strategy for Two-Dimensional Mesh Connected Systems
ICPP '93 Proceedings of the 1993 International Conference on Parallel Processing - Volume 02
Hi-index | 0.00 |
This paper presents a parallel process scheduling method for the AP/Linux parallel operating system. This method relies on 2 schedulings; local scheduling on each processor and global scheduling which is called moderate co-scheduling. Moderate co-scheduling schedules simultaneously parallel processes on each processor by controlling priorities of parallel processes. This method differs from gang scheduling in that it does not promise the running of a parallel process on all processors at the same time. Moderate co-scheduling only suggests a suitable current process to the local scheduling. However, this is good solution for fine and coarse grain parallel processes, because Moderate co-scheduling tells the timing to schedule simultaneously for fine grain parallel processes (tightly-coupled processes on each processor, which requires quick and frequent communication), and local scheduling can yield CPU time when coarse grain parallel processes (loosely-coupled processes on each processor, which cause long wait and less frequent communication) must wait for long time. The method is implemented using AP1000+ special hardware. We call the implementation "Internal synchronization" which uses the synchronized clock. The co-scheduling skew of the implementation was about 2% in the period of moderate co-scheduling was 200ms.