Intel® MKL 11.3 Beta (released in April 2015) contains significant performance and scalability improvements for the direct sparse solver (a.k.a. Intel MKL PARDISO), on SMP systems. These improvements particularly benefit the Intel Xeon Phi coprocessors and Intel Xeon processors with large core counts. As an example, the chart below shows a 1.7x to 2.5x speedup of Intel MKL 11.3 Beta over Intel MKL 11.2, when using the PARDISO to solve various sparse matrices on an Intel Xeon Phi coprocessor with 61 cores. In this example, PARDISO runs natively on the coprocessor, meaning the host CPU is not involved in the computation. A total number of 240 threads (60 cores, 4 threads each) are used during the execution.
The improvements are applied to all three phases of using PARDISO to solve a linear system, namely, reordering, factorization, and solving. Users do not need to turn any special knob (e.g. special environment settings or service function calls) to get the improved performance. It is available out of the box, as long as you use Intel MKL 11.3 Beta or later.
There is a tip to get the best possible scalability, though. In addition to the usual guidelines of getting good performance from MKL functions, as documented here and here, we recommend PARDISO users to set KMP_AFFINITY to "scatter" when the number of threads is less than or equal to 60. And, set KMP_AFFINITY to "compact" when the number of threads is more than 60.
Limitations: We are working hard to bring this improvement into all functionality areas of PARDISO. But at present (as of April 2015), you may not see the same extent of performance improvement if you use one of these PARDISO features: Schur complement; diagonal pivot control; user defined permutation; reordering with nested levels; and the Conditional Numeric Reproducibility (CNR) feature.