Scott Atchley
2006-Nov-10 08:57 UTC
[Lustre-discuss] Updated Lustre on MX-10G performance results
Hi all, Using the last code submitted to CFS, I reran the Lustre echo (diskless) tests using MXLND on MX-10G. The updated page is at: https://mail.clusterfs.com/wikis/lustre/MX-10G and it has a link to the original results. As before, MXLND can nearly saturate the wire using a single client for read and write. Using 3 clients slightly increases the read and write performance, but there is not much room for an increase to begin with. ;-) The big change is the drop in CPU usage. MXLND uses a single thread to handle completions and it can be set to block or poll on completions. I tested blocking, polling 5,000 times before blocking and polling 10,000 times before blocking. The blocking performance for metadata was better than the polling results. For reading and writing 1 MB messages, polling added 0.25% throughput but required much more CPU. Scott -- Scott Atchley Myricom Inc. http://www.myri.com
Scott Atchley
2006-Nov-10 09:24 UTC
[Lustre-discuss] Re: [Lustre-devel] Updated Lustre on MX-10G performance results
And I failed to update the Getattr 3 client performance. It is now updated. Scott On Nov 10, 2006, at 10:57 AM, Scott Atchley wrote:> Hi all, > > Using the last code submitted to CFS, I reran the Lustre echo > (diskless) tests using MXLND on MX-10G. The updated page is at: > > https://mail.clusterfs.com/wikis/lustre/MX-10G > > and it has a link to the original results. > > As before, MXLND can nearly saturate the wire using a single client > for read and write. Using 3 clients slightly increases the read and > write performance, but there is not much room for an increase to > begin with. ;-) > > The big change is the drop in CPU usage. MXLND uses a single thread > to handle completions and it can be set to block or poll on > completions. I tested blocking, polling 5,000 times before blocking > and polling 10,000 times before blocking. The blocking performance > for metadata was better than the polling results. For reading and > writing 1 MB messages, polling added 0.25% throughput but required > much more CPU. > > Scott > > -- > Scott Atchley > Myricom Inc. > http://www.myri.com > > > _______________________________________________ > Lustre-devel mailing list > Lustre-devel@clusterfs.com > https://mail.clusterfs.com/mailman/listinfo/lustre-devel >