Questions and Answers for the Intel® Math Kernel Library Webinar on November 11, 2010

Submit New Article

November 14, 2010 8:00 PM PST


If you missed our webinar "Get Ready for Intel® Math Kernel Library 10.3 — A Component of Intel® Composer XE 2011" presented on Nov 11, 2010, please download a recording of the webinar as well as a PDF file of the slides. Below are listed some of the questions and answers that were brought up during the presentation.


Questions on the new releases, version numbers, and upgrade

Q: what's the latest mkl version? we have Intel compiler 11.1 which includes mkl. Does this mean the mkl version is 11.1?
A: The version of Intel MKL that is included in the Intel Compiler 11.1 is Intel MKL 10.2 (and updates of this compiler contain updates of Intel MKL). If you’d like to know which version of Intel MKL you’re using, you can check the mklsupport* file in the doc or Documentation directory. Another place to look is one of the following knowledgebase articles:

Q: You mentioned that Intel Composer XE contains the C++ compiler. Why is there a separate C++ Composer XE product then?
A: There is an Intel® C++ Composer XE 2011 as well as an Intel® Fortran Composer XE 2011. These are available for those that need only one of the two compilers in Intel® Composer XE 2011. See the buy or renew products page for a full list.

Q: Intel® Parallel Studio XE 2011 - your first slide did not mention Mac OS* X. Why?
A: Intel® Parallel Studio XE 2011 is a suite of tools some of which are not available for Mac OS* X. The compiler and library support continues through the Intel® C++ Composer XE 2001 and Intel® Fortran Composer XE 2001 for Mac OS* X products. See: http://software.intel.com/en-us/articles/intel-sdp-products/.

Q: I recently purchased Intel Fortran Compiler 11.1, is the new Intel MKL 10.3 included as an upgrade?
A: Yes. Anyone with a current (unexpired) license for Intel MKL or professional editions of the Intel Compilers can obtain the new version of those tools (including Intel MKL 10.3).


Performance

Q: In which areas does ATLAS beat MKL?
A: We do regular performance benchmarking against ATLAS. We are not aware of places where ATLAS has better performance. If you know of any, please let us know.

Q: Is the hybrid improvement (MPI+OpenMP*) mainly for 3D FFTs?
A: Earlier we had implemented hybrid parallelism for 3D FFTs and now we have introduced it (MPI + OpenMP*) on cluster 1D complex transforms too. Most of the improvement is for vector lengths which are a multiple of the number of MPI processes

Q: Do you have a performance comparison with 64-bit GotoBLAS? I observed that GotoBLAS is faster in some case but a bit unstable.
A: No we don't have any performance comparisons.

Q: You talked about the MKL 10.3 performance on the new 6 core system. What about the performance on the existing Nehelem system?
A: You can find more performance information on the Intel MKL site under the 'resources' tab.

Q: I am testing PARDISO in my laptop and comparing with DSS sparse solver. They solve the matrix in the same amount of time. Am I doing something wrong?
A: This would be expected. DSS sits on top of Pardiso. The main value of DSS is that it provides a simplified interface.


Questions on Intel MKL features

Q: C interface to LAPACK - what kind of overhead with using it? I have been okay with the current usage, confusing but okay. Should I move over?
A: New kernels are steadily being introduced to eliminate the performance and memory overhead of transposition required for row-major data structures. Let us know if you have questions about a particular LAPACK function in a particular version of Intel MKL.

Q: Random number generator - should I expect to see differences in numbers generated (everything else constant including CPU) in a 32-bit build and a 64-bit build?
A: There are some cases where random generators differ between 32- and 64-bit builds due to different accuracy level of the VML functions used. More info to follow...

Q: Can you explain disadvantages of linking the OpenMP run-time library statically?
A: If your application or plug-in will be used in an environment where any other application or plug-in might also be threaded using OpenMP* then the OpenMP run-time may stall when it finds another statically linked run-time library already initialized.

Usage and tips

Q: Are there any common mistakes that new users make that frequently lead to undefined behaviour?
A: There are many ways that we see users make mistakes, but they are not so easy to enumerate. Here are some general categories that the problems can fall into: linking problems (use the link line advisor or user’s guide), improper data layout (e.g., for FFTs or PARDISO—see the reference manual), or simple API misunderstandings (see the reference manual)

Q: Is the icc -mkl=[parallel,sequential] link method recommended?
A: Yes! Take a look at the "Using MKL in Intel® Compiler - mkl, Qmkl options" article for more information.

Q: Is there some support for Eclipse IDE?
A: There is some integration of our documentation into the IDE as well as documentation in the product on how to setup Eclipse for use of Intel MKL. Take a look at “Programming with Intel® Math Kernel Library in the Eclipse* Integrated Development Environment (IDE)” section of the User's Guide for Linux*.

Plans for the future

Q: What new library functionality can users expect to appear in Intel MKL in coming years?
A: A package of Eigensolvers, a cluster version of PARDISO, extensions to VSL, etc.

Q: In the roadmap, do you have any implementations on GPU?
A: We have no current plans to extend GPU support.



Do you need more help?


This article applies to: Intel® Math Kernel Library Knowledge Base