Intel® Math Kernel Library 11.3 Update 4 Developer Guide
Certain situations can cause conflicts in the execution environment that make the use of threads in Intel MKL problematic. This section briefly discusses why these problems exist and how to avoid them.
If your program is parallelized by other means than Intel® OpenMP* run-time library (RTL) and Intel TBB RTL, several calls to Intel MKL may operate in a multithreaded mode at the same time and result in slow performance due to overuse of machine resources.
The following table considers several cases where the conflicts may arise and provides recommendations depending on your threading model:
|
Threading model |
Discussion |
|---|---|
|
You parallelize the program using the technology other than Intel OpenMP and Intel TBB (for example: pthreads on Linux* OS). |
If more than one thread calls Intel MKL, and the function being called is threaded, it may be important that you turn off Intel MKL threading. Set the number of threads to one by any of the available means (see Techniques to Set the Number of Threads). |
|
You parallelize the program using OpenMP directives and/or pragmas and compile the program using a non-Intel compiler. |
To avoid simultaneous activities of multiple threading RTLs, link the program against the Intel MKL threading library that matches the compiler you use (see Linking Examples on how to do this). If this is not possible, use Intel MKL in the sequential mode. To do this, you should link with the appropriate threading library: libmkl_sequential.a or libmkl_sequential.so (see Appendix C: Directory Structure in Detail). |
|
You thread the program using Intel TBB threading technology and compile the program using a non-Intel compiler. |
To avoid simultaneous activities of multiple threading RTLs, link the program against the Intel MKL TBB threading library and Intel TBB RTL if it matches the compiler you use. If this is not possible, use Intel MKL in the sequential mode. To do this, link with the appropriate threading library: libmkl_sequential.a or libmkl_sequential.so (see Appendix C: Directory Structure in Detail). |
|
You run multiple programs calling Intel MKL on a multiprocessor system, for example, a program parallelized using a message-passing interface (MPI). |
The threading RTLs from different programs you run may place a large number of threads on the same processor on the system and therefore overuse the machine resources. In this case, one of the solutions is to set the number of threads to one by any of the available means (see Techniques to Set the Number of Threads). Section Intel® Optimized MP LINPACK Benchmark for Clusters discusses another solution for a Hybrid (OpenMP* + MPI) mode. |
Using the mkl_set_num_threads and mkl_domain_set_num_threads functions to control parallelism of Intel MKL from parallel user threads may result in a race condition that impacts the performance of the application because these functions operate on internal control variables that are global, that is, apply to all threads. For example, if parallel user threads call these functions to set different numbers of threads for the same function domain, the number of threads actually set is unpredictable. To avoid this kind of data races, use the mkl_set_num_threads_local function (see the "Support Functions" section in the Intel MKL Developer Reference for the function description).
|
Intel's compilers may or may not optimize to the same degree for non-Intel microprocessors for optimizations that are not unique to Intel microprocessors. These optimizations include SSE2, SSE3, and SSSE3 instruction sets and other optimizations. Intel does not guarantee the availability, functionality, or effectiveness of any optimization on microprocessors not manufactured by Intel. Microprocessor-dependent optimizations in this product are intended for use with Intel microprocessors. Certain optimizations not specific to Intel microarchitecture are reserved for Intel microprocessors. Please refer to the applicable product User and Reference Guides for more information regarding the specific instruction sets covered by this notice. Notice revision #20110804 |