By Topic

Low-Cost Tuning of Two-Step Algorithms for Scheduling Mixed-Parallel Applications onto Homogeneous Clusters

Sign In

Cookies must be enabled to login.After enabling cookies , please use refresh or reload or ctrl+f5 on the browser for the login options.

Formats Non-Member Member
$33 $13
Learn how you can qualify for the best price for this item!
Become an IEEE Member or Subscribe to
IEEE Xplore for exclusive pricing!
close button

puzzle piece

IEEE membership options for an individual and IEEE Xplore subscriptions for an organization offer the most affordable access to essential journal articles, conference papers, standards, eBooks, and eLearning courses.

Learn more about:

IEEE membership

IEEE Xplore subscriptions

1 Author(s)
Sascha Hunold ; Int. Comput. Sci. Inst., Berkeley, CA, USA

Due to the strong increase of processing units available to the end user, expressing parallelism of an algorithm is a major challenge for many researchers. Parallel applications are often expressed using a task-parallel model (task graphs), in which tasks can be executed concurrently unless they share a dependency. If these tasks can also be executed in a data-parallel fashion, e.g., by using MPI or OpenMP, then we call it a mixed-parallel programming model. Mixed-parallel applications are often modeled as directed a cyclic graphs (DAGs), where nodes represent the tasks and edges represent data dependencies. To execute a mixed-parallel application efficiently, a good scheduling strategy is required to map the tasks to the available processors. Several algorithms for the scheduling of mixed-parallel applications onto a homogeneous cluster have been proposed. MCPA (Modified CPA) has been shown to lead to efficient schedules. In the allocation phase, MCPA considers the total number of processors allocated to all potentially concurrently running tasks as well as the number of processors in the cluster. In this article, it is shown how MCPA can be extended to obtain a more balanced workload in situations where concurrently running tasks differ significantly in the number of operations. We also show how the allocation procedure can be tuned in order to deal not only with regular DAGs (FFT), but also with irregular ones. We also investigate the question whether additional optimizations of the mapping procedure, such as packing of allocations or backfilling, can reduce the make span of the schedules.

Published in:

Cluster, Cloud and Grid Computing (CCGrid), 2010 10th IEEE/ACM International Conference on

Date of Conference:

17-20 May 2010