On Thursday 29 March 2012 12:34:10 Jeff Darcy wrote: > On 03/29/2012 03:30 PM, Harry Mangalam wrote: > > I'm doing some perf tests on a small gluster filesystem - 5 > > bricks on 4 servers, all single-homed on the private net. > > > > I've spawned up to 70 simultaneous jobs on our cluster nodes > > writing files of various sizes from /dev/zero to the gluster fs > > to see what the effect on the aggregate bandwith and the data is > > slightly unbelievable in that it seems to exceed the theoretical > > max of the network. (I used /dev/zero instead of /dev/urandom > > since /dev/urandom couldn't generate data fast enough. > > > > The 35,000 files of the right size do hit the filesystem (of > > course they're all zero's) but the speed at which they transfer > > exceeds (by quite a bit) the theoretical max of a 1 Gb network. > > > > Does gluster (or anything else) do transparent compression? What > > else would explain this oddity? > > How do you define "theoretical max of a 1Gb network"? If it's a > switched network, the actual maximum throughput depends on the > capabilities of the switch but is likely to be far in excess of > 1Gb/s. Could that be it? Could you give more detail about the > actual traffic patterns and results? > > BTW, this is my favorite message title ever. Thanks for that. :) So something good has come out of this.. :) You make a good point - I'm more of a biologist gone to to the dark side than a formally trained network guy. The cluster is in fact a distributed one - some of the cluster workers are in a remote building VLAN'ed into the main group, and the rest are on the main cluster network. From my understanding of the cluster networking, there are a number of 24- and 48- port stacked switches providing ports to the worker nodes, as well as to the gluster servers. My assumption was that at some point there would be a 1Gb bottleneck, but if the packets were switched all the way thru, there would be a theoretical max related the number of gluster-server Gb ports (4). So until I approach 4Gb/s, I guess I wouldn't necessarily see this bottleneck. Is that correct? Harry -- Harry Mangalam - Research Computing, OIT, Rm 225 MSTB, UC Irvine [ZOT 2225] / 92697 Google Voice Multiplexer: (949) 478-4487 415 South Circle View Dr, Irvine, CA, 92697 [shipping] MSTB Lat/Long: (33.642025,-117.844414) (paste into Google Maps) -- -------------- next part -------------- An HTML attachment was scrubbed... URL: <http://gluster.org/pipermail/gluster-users/attachments/20120329/bac21aee/attachment.htm>
