Increasing SpMV Energy Efficiency Through Compression: A study of how format, input and platform properties affect the energy efficiency of Compressed Sparse eXtended
This work is a continuation and augmentation of previous energy studies of
Compressed Sparse eXtended (CSX), a framework for efficiently executing Sparse
Matrix-Vector Multiplication (SpMV).
CSX was developed by the CSLab at the National Technical University of Athens
(NTUA), and utilizes compression to overcome a significant memory bottleneck
inherent in SpMV, thus increasing performance and energy efficiency of its
SpMV is notorious within scientific computing for its low performance. However,
the problem is unavoidable, as SpMV can be found within several scientific
applications. In this work, CSX is tested as the SpMV kernel in a framework
implementing the Conjugate Gradient Method (CG), an iterative algorithm for
solving specific linear algebra problems. CSX is also evaluated against
Compressed Sparse Row (CSR), a storage scheme widely used when executing SpMV.
This work augments existing studies by evaluating properties in the formats
themselves, in the matrices used as input and in the target platform to gain
knowledge on how to maximize the benefits of CSX, as well as for what cases
CSX does not prove beneficial. The work also compares the performance of
SpMV-execution on a stand-alone server known as the CARD-server to similar
execution on the Vilje supercomputer. This is done to evaluate how the
differences between these two machines affect the results.
Based on the results, it is shown that CSX should be used for matrices larger
than the Last Level Cache (LLC) of the target machine and for matrices with high
degrees of clustering in their values. The best energy efficiency trade-offs are
found at eight threads on dual socket configurations, and this is shown to be
related to the amount of physical cores per CPU. Similarly, frequency
throttling is shown to increase the energy efficiency of the execution only at
high numbers of threads and at the cost of performance.
Overall, CSX is shown to obtain higher energy efficiency than CSR for
SpMV-execution, given a suitable problem and run configuration. Thus, it is
confirmed that CSX can be used to decrease the energy consumption of SpMV
Place, publisher, year, edition, pages
Institutt for datateknikk og informasjonsvitenskap , 2013. , 103 p.
IdentifiersURN: urn:nbn:no:ntnu:diva-22977Local ID: ntnudaim:9692OAI: oai:DiVA.org:ntnu-22977DiVA: diva2:655611
Natvig, Lasse, Professor