Hectiling: An Integration of Fine and Coarse-Grained Load-Balancing Strategies

Abstract –– General–purpose programmers have come to expect a high degree of portability among widely varying architectures. Advances in run–time systems for parallel programs have been proposed in order to harness available resources as efficiently as possible. Simultaneously, advances in algorithmic ways of dynamically balancing computational load have been proposed in order to respond to variations in actual performance and therefore in runtime. The primary mechanism for harnessing idle resources effectively, task migration, can be used alongside the primary mechanism for dynamic load balancing, data redistribution. Besides the fact that the two methods can be used simultaneously to spur further increases in performance, the run–time information–gathering infrastructure necessary to detect and use idle resources can also benefit dynamically load–balanced applications. This paper describes an architecture for and preliminary implementation of a system that combines data–parallel load–balancing with task–parallel load–balancing. Performance test results are included as well.– General–purpose programmers have come to expect a high degree of portability among widely varying architectures. Advances in run–time systems for parallel programs have been proposed in order to harness available resources as efficiently as possible. Simultaneously, advances in algorithmic ways of dynamically balancing computational load have been proposed in order to respond to variations in actual performance and therefore in runtime. The primary mechanism for harnessing idle resources effectively, task migration, can be used alongside the primary mechanism for dynamic load balancing, data redistribution. Besides the fact that the two methods can be used simultaneously to spur further increases in performance, the run–time information–gathering infrastructure necessary to detect and use idle resources can also benefit dynamically load–balanced applications. This paper describes an architecture for and preliminary implementation of a system that combines data–parallel load–balancing with task–parallel load–balancing. Performance test results are included as well. 1: Introduction and Survey of Existing Systems One of the responsibilities of a parallel program and/or run–time system is that of load–balancing. Individual processors may vary in performance, external workload, or data distribution, and so methods to maintain an even distribution of work are usually needed to obtain good performance and speedup. One can consider a parallel program as consisting of p threads of execution and q data partitions. In general, p does not have to equal q (although this is usually the case). To maintain a balanced load, the threads can be moved, the data in the partitions can be redistributed, or the two (matched pairs of threads and data) can be moved together. These three cases can be called “thread migration”, “data migration”, and “task migration”, respectively. This paper considers the network–of–workstation environment, and in the NOW environment there is an additional reason to migrate work. It is often desirable to return a workstation back to its “owner” and release the CPU and memory resources. Thus “release of resources” is an additional reason to provide and use task migration. This is an important distinction, as task–migration–based load–balancing systems can perform task migrations for reasons unrelated to load–balancing, and data–migration and thread–migration–based systems may also provide task migration services. The three styles of migration (and of programming) are characterized below.