Optimizing OpenMP Parallelized DGEMM Calls on SGI Altix 3700

Using functions of parallelized mathematical libraries is a common way to accelerate numerical applications. Computer architectures with shared memory characteristics support different approaches for the implementation of such libraries, usually OpenMP or MPI. This paper’s content is based on the performance comparison of DGEMM calls (floating point matrix multiplication, double precision) with different OpenMP parallelized numerical libraries, namely Intel MKL and SGI SCSL, and how they can be optimized. Additionally, we have a look at the memory placement policy and give hints for initializing data. Our attention has been focused on a SGI Altix 3700 Bx2 system using BenchIT [1] as a very convenient performance measurement suite for the examinations. 1 Measurement Environment For a detailed analysis of a system architecture by parameter studies, the choice of a suitable measuring framework is an important decision. To benchmark the DGEMM calls we use BenchIT. This performance measurement suite helps to compare different algorithms, implementations of algorithms, features of the software stack, and hardware details of whole systems. It has been designed to run many microbenchmarks on every POSIX 1.003 compatible system in a very userfriendly way. BenchIT has been developed at the Center for Information Services and High Performance Computing (ZIH) at the Technische Universität Dresden and was previously mentioned at [2–4]. Sources and results are freely available at [1]. 2 The SGI Altix 3700 Bx2 System 2.1 System Architecture The SGI [5] Altix 3700 Bx2 is a ccNUMA shared memory system based on Intel Itanium 2 processors and SGI’s scalable node architecture SN2. In developing this, special attention has been paid to building a highly scalable computer