[Beowulf] C-States and Latency

Christopher Samuel samuel at unimelb.edu.au
Wed Feb 26 16:28:16 PST 2014

Hash: SHA1

On 27/02/14 05:16, Kevin Van Workum wrote:

> Yes, this is the recommendation I'm wondering about. Can you
> describe in more detail the performance problems you saw? Any
> particular application? Measurables?

This was when we were bringing up our new Sandy Bridge cluster in the
middle of last year and were doing the acceptance testing.

Part of that was running HPL for long periods to weed out bad hardware
and check we were getting the performance you'd expect.

Changing those states meant going from 15-20% efficiency to around 90%
efficiency (reported by the binary, statically linked, Intel HPL build).

Sorry I've not been able to reply to the offist queries, swamped with
some IB work here to replace a dodgy switch.

Summary:  You are lost in a maze of twisty IB cables, all allegedly
alike, but not really.

All the best,
- -- 
 Christopher Samuel        Senior Systems Administrator
 VLSCI - Victorian Life Sciences Computation Initiative
 Email: samuel at unimelb.edu.au Phone: +61 (0)3 903 55545
 http://www.vlsci.org.au/      http://twitter.com/vlsci

Version: GnuPG v1.4.14 (GNU/Linux)
Comment: Using GnuPG with Thunderbird - http://www.enigmail.net/


More information about the Beowulf mailing list