On Thu, May 31, 2007 at 01:38:46PM -0700, David Brown wrote: > Okay this is kinda weird but here you go.
Heh, yeah, I'm having a bit of trouble making sense of the data > 1) Setup N equal number of data servers and clients. > 2) From each client choose a server to write to and determine the path > of least resistance to that server, try to make the data from this > client to this server as fast as possible. For pvfs this means run the > client and server on the same box. > 3) Simultaneously write a large chunk of data from all the clients to > their individually selected server from above. Running the clients and servers on the same machine might actually hurt your performance a bit. Since PVFS doesn't have a native quadrics method, maybe you save a lot of overhead skipping the tcp-over-quadrics stuff. > The objective is to increase N to large numbers to see what performance > hits are taken by the file system. It might be easier to see patterns if you held servers constant and increased the number of clients, or held clients constant while varying numbers of servers. To visualize both you'd end up with a 3-d plot... > Okay Question: > The data shows some interesting things as it jumps to the higher numbers > theres a step increase in the time it takes to dd a 10Gb file from > 45-50s then jumps up to 65-70s on some nodes and they are groupped > across the node set as well. Any clues as to what could cause this? I'm looking at 255-test.csv. That's 256 nodes (acting as servers and clients), each client running dd to write 10 GB to a single server? I don't know why that workload would take about a minute for up to 64 clients, speed up for 65-141 clients, and then go back to being slower for the rest of the runs, except for a cluster of fast runs at 173-183 clients. Since you've got things set up so each client talks to a single server locally, we shouldn't be seeing network contention or switch wierdness. Since you have a single client talking to a single server, the access pattern from each client should look pretty regular to pvfs2-server. The tight bimodal distribution of results suggests... I don't know... fortunate placement of files on the storage device for some runs? http://www.coker.com.au/bonnie++/zcav ==rob -- Rob Latham Mathematics and Computer Science Division A215 0178 EA2D B059 8CDF Argonne National Lab, IL USA B29D F333 664A 4280 315B _______________________________________________ Pvfs2-users mailing list [email protected] http://www.beowulf-underground.org/mailman/listinfo/pvfs2-users
