[Trilinos-Users] Zoltan performance

Biddiscombe, John A. biddisco at cscs.ch
Wed Feb 15 05:01:11 MST 2012

I made some tests of Zoltan partitioning on particle datasets and found something worth reporting here.

Test 1 : strong scalability. I took the same dataset (100 million particles 8GB) and repartitioned it on different numbers of cores,

the graph below shows that performance is great until 2048 cores where it falls off a cliff (or up one in this case).

Tests done on Cray XK6, using trilinos 10.8.5 release. No unusual options selected. RCB method used.

(I haven't given you enough details to diagnose anything, but I can dig out all the settings if this is worth pursuing)

I am going to do tests again and instrument the code to do some profiling of where things have gone wrong, but before I spend a lot of time on that, I'm interested to know if there is any particular reason why 2048 cores might be interesting. Are there any buffer/memory/static variable options that would kick in at this level which simple changes might put off until 4096 8192 etc core counts? Is the data too small for partitioning on this many cores?

[cid:image004.png at 01CCEBE1.E3B487D0]

This second graph is where larger and larger datasets are partitioned, always on 2048 cores - in this case the data is getting larger and larger (64GB for the final point on the graph), but 64GB spread over 2048 cores is just 16MB each so I was surprised to see how badly it runs. I haven' t ruled out a machine related issue, but if these graphs trigger any thoughts, then please let me know so I don't waste time on red herrings.

[cid:image006.png at 01CCEBE1.E3B487D0]

Many Thanks



John Biddiscombe,                            email:biddisco @ cscs.ch


CSCS, Swiss National Supercomputing Centre  | Tel:  +41 (91) 610.82.07

Via Cantonale, 6928 Manno, Switzerland      | Fax:  +41 (91) 610.82.82

-------------- next part --------------
An HTML attachment was scrubbed...
URL: https://software.sandia.gov/pipermail/trilinos-users/attachments/20120215/157f4f0d/attachment.html 
-------------- next part --------------
A non-text attachment was scrubbed...
Name: image004.png
Type: image/png
Size: 16976 bytes
Desc: image004.png
Url : https://software.sandia.gov/pipermail/trilinos-users/attachments/20120215/157f4f0d/attachment.png 
-------------- next part --------------
A non-text attachment was scrubbed...
Name: image006.png
Type: image/png
Size: 15046 bytes
Desc: image006.png
Url : https://software.sandia.gov/pipermail/trilinos-users/attachments/20120215/157f4f0d/attachment-0001.png 

More information about the Trilinos-Users mailing list