Analyzing the Performance of Allocation Strategies Based on Space-Filling Curves

Future exascale supercomputers will be composed of thousands of nodes. In those massive systems, the search for physically close nodes will become essential to deliver an optimal environment to execute parallel applications. Schedulers manage those resources, shared by many users and jobs, searching for partitions in which jobs will run. Significant effort has been devoted to develop allocation strategies that maximize system utilization, while providing partitions that are adequate for the communication demands of applications. In this paper we evaluate a class of strategies based on space-filling curves (SFCs) that search for partitions in which nodes are physically close, compared to other alternatives that relax this requirement (e.g. non-contiguous), or make it even more strict (e.g. contiguous). Several metrics are used to assess the quality of an allocation strategy, some based on system utilization, some others measuring the quality of the resulting partitions. Contiguous allocators suffer from severe degradation in terms of system utilization, while non-contiguous allocators provide inadequate partitions. Somewhere in the middle, SFC allocators offer good system utilization while using quite compact partitions. The final metric to decide which allocator is the best depend on the severity of the slowdown suffered by applications when running in non-optimal partitions.

[1]  Bill Nitzberg,et al.  Noncontiguous Processor Allocation Algorithms for Mesh-Connected Multicomputers , 1997, IEEE Trans. Parallel Distributed Syst..

[2]  Ibm Blue,et al.  Overview of the IBM Blue Gene/P Project , 2008, IBM J. Res. Dev..

[3]  Javier Navaridas,et al.  Effects of Job and Task Placement on Parallel Scientific Applications Performance , 2009, 2009 17th Euromicro International Conference on Parallel, Distributed and Network-based Processing.

[4]  Esther M. Arkin,et al.  Processor allocation on Cplant: achieving general processor locality using one-dimensional allocation strategies , 2002 .

[5]  M. Jette,et al.  Simple Linux Utility for Resource Management , 2009 .

[6]  José Antonio Lozano,et al.  Optimization-based mapping framework for parallel applications , 2011, J. Parallel Distributed Comput..

[7]  Rolf Niedermeier,et al.  On Multi-dimensional Hilbert Indexings , 1998, COCOON.

[8]  David P. Bunde,et al.  Local search to improve coordinate-based task mapping , 2016, Parallel Comput..

[9]  David P. Bunde,et al.  Faster high-quality processor allocation. , 2010 .

[10]  Cyriel Minkenberg,et al.  Quiet Neighborhoods: Key to Protect Job Performance Predictability , 2015, 2015 IEEE International Parallel and Distributed Processing Symposium.

[11]  Jarek Nabrzyski,et al.  Topology-Aware Scheduling on Blue Waters with Proactive Queue Scanning and Migration-Based Job Placement , 2015, JSSPP.

[12]  Uwe Schwiegelshohn,et al.  Parallel Job Scheduling - A Status Report , 2004, JSSPP.

[13]  Christopher R. Johnson,et al.  A Tie-Breaking Strategy for Processor Allocation in Meshes , 2010, 2010 39th International Conference on Parallel Processing Workshops.

[14]  Bill Nitzberg,et al.  Non-contiguous processor allocation algorithms for distributed memory multicomputers , 1994, Proceedings of Supercomputing '94.

[15]  D. Hilbert.  Ueber die stetige Abbildung einer Line auf ein Flächenstück , 1891 .

[16]  Philip Heidelberger,et al.  The IBM Blue Gene/Q interconnection network and message unit , 2011, 2011 International Conference for High Performance Computing, Networking, Storage and Analysis (SC).

[17]  José Antonio Lozano,et al.  Strategies to Map Parallel Applications onto Meshes , 2010, DCAI.

[18]  Javier Navaridas,et al.  Effects of Job and Task Placement on the Performance of Parallel Scientific Applications , 2008 .

[19]  José Antonio Lozano,et al.  Locality-aware policies to improve job scheduling on 3D tori , 2014, The Journal of Supercomputing.

[20]  V. Lo,et al.  Contiguous and Non-contiguous Processor Allocation , 1995 .

[21]  José Antonio Lozano,et al.  Application-aware metrics for partition selection in cube-shaped topologies , 2014, Parallel Comput..

[22]  Peter J. H. King,et al.  Using Space-Filling Curves for Multi-dimensional Indexing , 2000, BNCOD.

[23]  José Antonio Lozano,et al.  A fast implementation of the first fit contiguous partitioning strategy for cubic topologies , 2014, Concurr. Comput. Pract. Exp..

[24]  Rolf Niedermeier,et al.  On Multidimensional Curves with Hilbert Property , 2000, Theory of Computing Systems.

[25]  D. Hilbert.  Über die stetige Abbildung einer Linie auf ein Flächenstück , 1935 .