[petsc-users] Sometimes it's better NOT to parallelize ??? (SLEPc question?)

[Jose E. Roman]

As stated in the documentation, the 'lapack' solver is a wrapper to LAPACK functions. LAPACK is a sequential library, so do not expect speedup for more than one process. The LAPACK wrappers are provided for convenience, as a debugging tool, for small problems. SLEPc is intended for large-scale sparse eigenproblems. We do not provide parallel eigensolvers for dense matrices.

Jose


El 29/07/2011, a las 09:56, John Chludzinski escribió:

> These are the resulting stats from decomposing a 4002x4002 (dense matrices) generalized eigenvalue problem into 2 MPI processes. Note the amount of message traffic: 
> 
>                          Max       Max/Min        Avg      Total
> MPI Messages:         8.011e+03      1.00000   8.011e+03  1.602e+04
> MPI Message Lengths:  2.242e+08      1.00000   2.799e+04  4.485e+08
> 
> Total # of messages: 1.602e+04 with an average message length: 2.799e+0.  
> 
> 
> With the number of MPI set to 1, you get (not surprisingly):
> 
>                          Max       Max/Min        Avg      Total
> MPI Messages:         0.000e+00      0.00000   0.000e+00  0.000e+00
> MPI Message Lengths:  0.000e+00      0.00000   0.000e+00  0.000e+00
> 
> In the end the time require to solve my 4002x4002 eigenvalue problem for 2 MPI processes: 1.821e+03 vs. 1.312e+03 for 1 MPI process.  
> 
> Am I reading this correctly?
> 
> 
> ---John
> 
> 
> Complete stats for 2 MPI process run:
> 
> Using Petsc Release Version 3.1.0, Patch 7, Mon Dec 20 14:26:37 CST 2010
> 
>                          Max       Max/Min        Avg      Total
> Time (sec):           1.821e+03      1.00064   1.820e+03
> Objects:              2.005e+04      1.00000   2.005e+04
> Flops:                1.282e+11      1.00000   1.282e+11  2.564e+11
> Flops/sec:            7.046e+07      1.00064   7.044e+07  1.409e+08
> Memory:               1.286e+09      1.00000              2.571e+09
> MPI Messages:         8.011e+03      1.00000   8.011e+03  1.602e+04
> MPI Message Lengths:  2.242e+08      1.00000   2.799e+04  4.485e+08
> MPI Reductions:       2.412e+04      1.00000
> 
> Flop counting convention: 1 flop = 1 real number operation of type (multiply/divide/add/subtract)
>                             e.g., VecAXPY() for real vectors of length N --> 2N flops
>                             and VecAXPY() for complex vectors of length N --> 8N flops
> 
> Summary of Stages:   ----- Time ------  ----- Flops -----  --- Messages ---  -- Message Lengths --  -- Reductions --
>                         Avg     %Total     Avg     %Total   counts   %Total     Avg         %Total   counts   %Total
>  0:      Main Stage: 1.8203e+03 100.0%  2.5645e+11 100.0%  1.602e+04 100.0%  2.799e+04      100.0%  2.007e+04  83.2%
> 
> 
> Complete stats for 1 MPI process run:
> 
>                          Max       Max/Min        Avg      Total
> Time (sec):           1.312e+03      1.00000   1.312e+03
> Objects:              2.003e+04      1.00000   2.003e+04
> Flops:                2.564e+11      1.00000   2.564e+11  2.564e+11
> Flops/sec:            1.955e+08      1.00000   1.955e+08  1.955e+08
> Memory:               1.029e+09      1.00000              1.029e+09
> MPI Messages:         0.000e+00      0.00000   0.000e+00  0.000e+00
> MPI Message Lengths:  0.000e+00      0.00000   0.000e+00  0.000e+00
> MPI Reductions:       2.404e+04      1.00000
> 
> Flop counting convention: 1 flop = 1 real number operation of type (multiply/divide/add/subtract)
>                             e.g., VecAXPY() for real vectors of length N --> 2N flops
>                             and VecAXPY() for complex vectors of length N --> 8N flops
> 
> Summary of Stages:   ----- Time ------  ----- Flops -----  --- Messages ---  -- Message Lengths --  -- Reductions --
>                         Avg     %Total     Avg     %Total   counts   %Total     Avg         %Total   counts   %Total
>  0:      Main Stage: 1.3119e+03 100.0%  2.5645e+11 100.0%  0.000e+00   0.0%  0.000e+00        0.0%  2.002e+04  83.3%
> 
>