I have maintenance windows on Wednesday & Thursday to make changes to the throughput cluster here and I could use some advice. There are 130 system, 126 of them are compute nodes, connected via Gb ethernet with the head node and fileserver node connected via 10Gb ethernet. At the moment, my configuration from last summer, which was done in a hurry, is running really, really slow. My hope is to upgrade to 1.4.0, but I'll settle for a release candidate if I have to. Here is the rough plan for how I'm going to configure it. Server: XFS filesystem -->> io-thread -->> tcp Client: brick -->> readahead -->> writebehind Should I be considering readahead/writebehind on the server end? How about threads on the client? Any other translators I should consider? I'm hoping to spend much of Thursday doing performance testing. What benchmarks do y'all recommend? I was planning on trying iozone, but could easily be convinced to go with something else. The plan is to script it to run automatically with a number of different settings for readahead & writebehind. Any particular words of wisdom? Thanks for the help! -- Matt
