PARDISO use half the memory now


From Intel MKL 10.2 version onwards, the memory footprint for PARDISO (*) is significantly reduced for both in-core and out-of-core symmetric matrix solvers. PARDISO out-of-core mode in MKL 10.2 uses less then a half of memory used by MKL 10.1 (**) , for real symmetric, complex Hermitian and complex symmetric matrices. The improvement of PARDISO memory consumption will be useful for many customers who deal with large and very large problems.

The memory footprint optimization is done for the reordering stage of the PARDISO solver. This explains why the reduction effect is bigger for the out-of-core mode vs. the in-core mode.  In the in-core mode the factorization stage is responsible for the major part of the memory footprint. Therefore the improvement of the memory consumption at the reordering stage demonstrates a limited reduction of the integral memory footprint.  While for the out-of-core mode the reordering stage is the most memory consuming and its optimization results in the bigger improvement of the integral footprint.

Below charts present the comparison of PARDISO memory consumption for in-core (Fig.1) and out-of-core (Fig.2)  modes for MKL 10.1 and 10.2.  The data on the charts are for MMD (the minimum degree algorithm) reordering type. For METIS (nested dissection algorithm) reordering the improvements are similar.

The charts show that the improvements for the in-core solver are in the range of 30-50% and for the out-of-core solver are in the range of 10-50%

Fig.1 Decrease memory for PARDISO In-Core mode

fig1.JPG

Fig.2 Decrease memory for PARDISO Out-Of-Core mode.
fig2.JPG

These charts represent results for all supported matrix types. These are:

Number of matrix type

matrix type

1

real and structurally symmetric matrix

2

real and symmetric positive definite matrix

-2

real and symmetric indefinite matrix

3

complex and structurally symmetric matrix

4

complex and Hermitian positive definite matrix

-4

complex and Hermitian indefinite matrix

6

complex and symmetric matrix

 

Note:

* PARDISO - Parallel Direct Sparse Solver.
Intel® MKL PARDISO is compliant with the 3.2 release of PARDISO freely distributed by the University of Basel which can be obtained at http://www.pardiso-project.or

 **  It requires to provide more precisely info about MKL versions contain these improvements: Actually the reducing memory use in the PARDISO was introduced originally since MKL v.10.1 Update1.
Please refer to the Release Notes for the MKL 10.1 Update1, we can find the following information about that:
New in Intel® MKL 10.1 Update 1:
    ................
    Bug fixes and other improvements:
          ...................          
    Reduced memory use in the PARDISO out-of-core (OOC) solver by 30 - 50% and in-core by 10 - 50%

 Therefore, all latest MKL versions ( MKL 10.1 update2, 3 and the modest version 10.2 ) contain such improvements for both ( in-core and ooc ) modes.

For more complete information about compiler optimizations, see our Optimization Notice.

Comments

For more info about correlation between Intel® Compiler Pro and MKL, please refer to the KB article “Which version of IPP, Intel MKL and TBB is installed by the Intel® Compiler Professional Edition?” by the following link: http://software.intel.com/en-us/articles/which-version-of-ipp--mkl--tbb-is-installed-with-intel-compiler-professional-edition/
--Gennady