Tuesday, February 8, 2011

Many-task applications and divisible load theory

Recently, I found a paper about a new kind of applications named 'Many-task applications'. Those applications fill a gap between high performance computing and high throughput computing. Better say, those applications regard with bulk of tasks may be static or dynamic, homogeneous or heterogeneous, loosely coupled or tightly coupled, [1].
Divisible load theory offers another approach to the scheduling problem. DLT, for short, has been widely studied and applied to different distributed and large scale scenarios with important restrictions about the nature of the applications to be scheduled. However, DLT provides a close and optimal solution when some restrictions are observed.
It's important, then, to analyse the nature and characteristics of many-task applications in such a way that DLT approach can be successfully applied under circumvented  circumstances.

[1] Many-task computing for grids and supercomputers. I. Raicu, I. Foster, Y. Zhao, 2008.
[2] Scheduling many-task workloads on supercomputers; dealing with trailing tasks. T. Armstrong, Z. Zhang, D. Katz, M. Wilde, I. Foster, 2010.

No comments: