The mini mock data challenge was run from Monday-Wednesday morning. Some results We were running reconstruction jobs for about 12 hours with reasonable efficiency. a.. Cpu time per job was down by considerable amount compared to MDC2 (known before) typical runs took .5-.45 hour per 200Mb i.e. .06-0.1 Mb/sec per job i.e. 12*.1 == 1.2 Mb/sec input. Details of the analysis are summarized in an spreadsheet found on the computing Web pages b.. 255 jobs were run - 24 10% had stage failed. This is thought to be related to communication losses that Tom Throwe have seen between crs server and crs nodes. c.. 28 jobs had a stale NFS file handle. This clearly appeared because the BramReco job opens the root output files on the nfs mounted disks. The proper way is to open to a local file, and then copy the histogram files at the end of the job. d.. Due to the problems the load on the 12 farm nodes was below 2 for much of the time. e.. In the early set too, the BramReco job was writing the event number information out on a per event basis. This too lead to a degradation of performance. f.. Data were only transferred for a short while from the BRAHMS DAQ to RCF. (Monday evening). HPSS returned errors. g.. Since Phobos did not take part in this MDC Bruce will like to repeat this likely by end of February. h.. Star achieved DAQ -> HPSS sustained rates of 18Mb/sec over the 2 days. ------------------------------ Flemming Videbaek Physics Department Brookhaven National Laboratory tlf: 516-344-4106 fax: 516-344-1334 videbaek@bnl.gov
This archive was generated by hypermail 2b29 : Thu Feb 03 2000 - 16:58:41 EST