Tech Articles

MSC Nastran Parallel Information


Tech Articles ID    KB8022087
Status:    Published
Published date:    04/24/2015
Updated:    06/29/2015
Reported In:   MSC Nastran
 

Abstract (Question)

Summary

  • Shared Memory Parallel ( SMP ) is using multiple processors on the same computer. This option improves performance at a low level ( like a Matrix Multiply operation ).
  • Distributed Memory Parallel ( DMP ) is using multiple processors on either the same or multiple computers. This option improves performance at a higher analysis level ( like breaking up a frequency range ).
  • Parallel performance improvement is highly model dependent. Understanding expectations on your job is important.
  • In general, SMP should be limited to the number of cores per socket. DMP should be limited to the number of sockets or total number of sockets of all nodes.

Description (Answer)

Details

Consider the image below for DMP:

 

Some SOL's like 101 may have a fair amount of commnication between processors. For these SOL, running DMP on a single node is best. However, for SOL's with little communcation, SOL 108 for example, optimum performance may be found using multiple nodes. The total network traffic may be found at the bottom of the F04 file. e.g.

 *** MPI STATISTICS FOR DISTRIBUTED NASTRAN ***

    FROM     MESSAGES  TOTAL (MB)  MAX BYTE  MIN BYTE  AVG BYTE        FROM     MESSAGES  TOTAL (MB)  MAX BYTE  MIN BYTE  AVG BYTE
 ----------  --------  ----------  --------  --------  --------     ----------  --------  ----------  --------  --------  --------
   1 -> 2       13694   87738.5181186039168         8   6718307       2 -> 1        7526    4342.320 200299008         8    605002

 TOTAL NUMBER OF MESSAGES SENT =           21220
 TOTAL AMOUNT OF DATA XFR (MB) =       92080.838
 AVERAGE MESSAGE LENGTH        =         4550130.

Infiniband may help on DMP analysis with multiple nodes on SOLs with a lot of communication. MSC has limited testing with Infiniband but would expect 10 to 30% improvement in DMP performance with multiple nodes using Infiniband depending on the SOL. Infiniband can be requested with IntelMPI with: symbol=I_MPI_DEVICE=rdssm

Below are examples of parallel with different SOLs.

    SOL 101 SMP Parallel
    SOL 101 DMP Parallel
    SOL 103 SMP Parallel
    SOL 103 DMP Parallel
    SOL 108 DMP Parallel
    SOL 400 SMP Parallel
    SOL 400 DMP Parallel
    Fatigue

Caveats

  • DMP is limited to SOL 101, 103, 108, 110, 111, 112, and 200
  • ACMS has a limitation of DMP=8
  • The EIGRL entry needs F2 for Frequency Domain or ACMS. For DOF Domain or Geometric Domain ND may be specified.
  • The Iterative solver requires an ITER entry with PRECOND=J and SMETHOD=ITER_ID
  • SOL 110 and 112 support DMP for:
    • ACMS w/ "DOMAINSOLVER ACMS"
    • Matrix Domain w/ "DOMAINSOLVER MODES (PART=DOF)"
    • Frequecy Domain w/ "DOMAINSOLVER MODES (PART=FREQ)"
  • SOL 200.
    • For ANAL=FREQ, same as SOL 108
    • For ANAL=MFREQ, same as SOL 111

 


Rate this Page
Rate the quality of this article from 5 (high) to 1 (low):
1
2
3
4
5
 
How can we improve this?