Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Add MPI process stride for tasks exclusively owned by a component #4859

Merged
merged 1 commit into from
May 4, 2022

Conversation

amametjanov
Copy link
Member

@amametjanov amametjanov commented Mar 29, 2022

This lets a component run exclusively on an MPI task: e.g. ATM on 6 MPI tasks for 6 GPUs on a Summit node.

[NML] - due to new excl_stride namelist variables

@amametjanov amametjanov added BFB PR leaves answers BFB Performance labels Mar 29, 2022
@amametjanov
Copy link
Member Author

amametjanov commented Mar 29, 2022

A single-node test:

$ ./pelayout
Comp  NTASKS  NTHRDS  ROOTPE PSTRIDE
CPL :     42/     1;      0      1
ATM :      6/     1;      1      7
LND :      7/     1;     21      1
ICE :     35/     1;      0      1
OCN :      7/     1;     35      1
ROF :      7/     1;     28      1
GLC :      1/     1;      0      1
WAV :      1/     1;      0      1
IAC :      1/     1;      0      1
ESP :      1/     1;      0      1

    <entry id="EXCL_STRIDE">
      <type>integer</type>
      <values>
        <value compclass="ATM">7</value>
        <value compclass="CPL">0</value>
        <value compclass="OCN">0</value>
        <value compclass="WAV">0</value>
        <value compclass="GLC">0</value>
        <value compclass="ICE">0</value>
        <value compclass="ROF">0</value>
        <value compclass="LND">0</value>
        <value compclass="ESP">0</value>
        <value compclass="IAC">0</value>
      </values>
      <desc>Stride of MPI tasks owned exclusively by a component. If 0, exclusivity is disabled.</desc>
    </entry>

$ ./xmlchange INFO_MPROF=2

$ zgrep "comps" run/cpl.log.1956863.220328-232850.gz 
 memory_write: model date =   00010101       0 memory =    1105.25 MB (highwater)        443.81 MB (usage)  (pe=    0 comps= cpl ICE GLC WAV IAC ESP)

$ zgrep "comps.*ATM" run/e3sm.log.1956863.220328-232850.gz 
1: 22:  memory_write: model date =   00010101       0 memory =     955.31 MB (highwater)        333.88 MB (usage)  (pe=   22 comps= cpl ATM)
1: 15:  memory_write: model date =   00010101       0 memory =     953.69 MB (highwater)        333.44 MB (usage)  (pe=   15 comps= cpl ATM)
1: 8:  memory_write: model date =   00010101       0 memory =     961.00 MB (highwater)        338.69 MB (usage)  (pe=    8 comps= cpl ATM)
1: 1:  memory_write: model date =   00010101       0 memory =     974.44 MB (highwater)        339.62 MB (usage)  (pe=    1 comps= cpl ATM)
1: 29:  memory_write: model date =   00010101       0 memory =     950.81 MB (highwater)        329.44 MB (usage)  (pe=   29 comps= cpl ATM)
1: 36:  memory_write: model date =   00010101       0 memory =     949.06 MB (highwater)        330.25 MB (usage)  (pe=   36 comps= cpl ATM)

$ zgrep "comps.*LND" run/e3sm.log.1956863.220328-232850.gz 
1: 21:  memory_write: model date =   00010101       0 memory =    1010.06 MB (highwater)        377.81 MB (usage)  (pe=   21 comps= cpl LND ICE)
1: 27:  memory_write: model date =   00010101       0 memory =     969.38 MB (highwater)        352.31 MB (usage)  (pe=   27 comps= cpl LND ICE)
1: 23:  memory_write: model date =   00010101       0 memory =     973.31 MB (highwater)        356.44 MB (usage)  (pe=   23 comps= cpl LND ICE)
1: 26:  memory_write: model date =   00010101       0 memory =     974.88 MB (highwater)        355.75 MB (usage)  (pe=   26 comps= cpl LND ICE)
1: 25:  memory_write: model date =   00010101       0 memory =     971.50 MB (highwater)        352.06 MB (usage)  (pe=   25 comps= cpl LND ICE)
1: 24:  memory_write: model date =   00010101       0 memory =     974.12 MB (highwater)        356.31 MB (usage)  (pe=   24 comps= cpl LND ICE)

$ zgrep "comps.*OCN" run/e3sm.log.1956863.220328-232850.gz 
1: 35:  memory_write: model date =   00010101       0 memory =     976.38 MB (highwater)        347.31 MB (usage)  (pe=   35 comps= cpl OCN)
1: 37:  memory_write: model date =   00010101       0 memory =     968.88 MB (highwater)        346.19 MB (usage)  (pe=   37 comps= cpl OCN)
1: 41:  memory_write: model date =   00010101       0 memory =     972.75 MB (highwater)        349.81 MB (usage)  (pe=   41 comps= cpl OCN)
1: 38:  memory_write: model date =   00010101       0 memory =     986.12 MB (highwater)        364.94 MB (usage)  (pe=   38 comps= cpl OCN)
1: 40:  memory_write: model date =   00010101       0 memory =    1023.88 MB (highwater)        400.88 MB (usage)  (pe=   40 comps= cpl OCN)
1: 39:  memory_write: model date =   00010101       0 memory =    1012.75 MB (highwater)        391.38 MB (usage)  (pe=   39 comps= cpl OCN)

@rljacob rljacob requested a review from jonbob April 4, 2022 17:36
@rljacob rljacob self-assigned this Apr 4, 2022
@rljacob rljacob requested review from philipwjones and removed request for jonbob April 4, 2022 17:39
Copy link
Contributor

@philipwjones philipwjones left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This looks good to me. Approve by visual inspection.

rljacob added a commit that referenced this pull request May 3, 2022
#4859)

Add MPI process stride for tasks exclusively owned by a component
This lets a component run exclusively on an MPI task: e.g. ATM on 6 MPI tasks
for 6 GPUs on a Summit node.

[BFB]
@rljacob
Copy link
Member

rljacob commented May 3, 2022

merged to next

@amametjanov amametjanov added the NML label May 4, 2022
@amametjanov amametjanov merged commit 00e5c34 into master May 4, 2022
@amametjanov amametjanov deleted the azamat/driver/add-exclusive-process-stride branch May 4, 2022 19:36
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
BFB PR leaves answers BFB NML Performance
Projects
None yet
Development

Successfully merging this pull request may close these issues.

None yet

3 participants