I don't know what hardware we'd choose, but I wholeheartdly agree on the need for some better hardware than what we have now. During afternoon (french time), it becomes slow as hell, painful to see Wikipedia crawl like that... -.- Ryo

On Apr 6, 2004, at 14:36, Luc Van Oostenryck wrote: We have an intermittent problem where one of the servers will have an unusually large number of running processes for a minute or two, which will slow down response time. By unusual I mean like 50-100 processes fighting for CPU time; normally there are just a handful. The cause of this isn't entirely clear yet. My pure speculation is that the squids might be distributing traffic poorly; for instance the NFS server occasionally freezes up for a couple seconds, which might cause the squids to give up on all of the servers, then start sending all traffic to the first one that responds. Maybe? Maybe not. Just speculation which I haven't gotten a chance to investigate. -- brion vibber (brion @ pobox.com)

Timwi wrote: It would be useful to have MRTG or similar graphs available for _all_ the servers in the cluster, showing for each system: * disk I/O seeks/second * disk I/O bytes/second * network interfaces bytes/second * CPU load * memory usage * swap load This, in conjunction with the existing real-time performance graphs, would enable "many eyes" to inspect the system, and form hypotheses about where the bottlenecks are. -- Neil