thanks for sharing your experience.
I am running peer credits value 63 as it was for mlx4.
I did not notice problems with these settings anyway to saturate the EDR
bandwith writing and reading using Lustre I have to do it from 2
different clients,
so that I can reach up to 9GB/s per OSS. Otherwise from a single client
I cannot reach more than 5GB/s. it is not clear to me why.
On 11/8/18 4:58 AM, Martin Hecht wrote:
On 11/7/18 9:44 PM, Riccardo Veraldi wrote:
Anyway I Was wondering if something different is needed for mlx5 and
what are the suggested values in that case ?
Anyone has experience with mlx5 LNET performance tunings ?
Hi Riccardo,
We have recently integrated mlx5 nodes into our fabric, and we had to
reduce the values to
peer_credits = 16
concurrent_sends = 16
because mlx5 doesn't support larger values for some reason. The peer_credits
must have the same value in all connected lnets, even across routers (at least
it used to be like this. I believe we are currently running some Lustre 2.5.x
derivates on the server side, and newer versions on the various clients).
kind regards,
Martin
_______________________________________________
lustre-discuss mailing list
[email protected]
http://lists.lustre.org/listinfo.cgi/lustre-discuss-lustre.org
_______________________________________________
lustre-discuss mailing list
[email protected]
http://lists.lustre.org/listinfo.cgi/lustre-discuss-lustre.org
