> Let me look into this and I'll respond back (probably > next week sometime). > > -Peter
OK, well I didn't quite make it back in a week (bear in mind Sun was closed last week), but here's what I've got so far. I attempted to reproduce this on a much slower system than what you are using on the theory that if the performance degradation is related to the driver implementation it should scale down with a smaller system. So far I have been unable to reproduce the problem. My test system is an Ultra 60 with 512MB connected to four Seagate ST39102 harddrives in a JBOD with parallel SCSI. Test results for 512 byte sequential reads, 16 outstanding I/Os: Opensolaris/Nevada on S10: ~14200 S10 latest: ~14200 sd S9 latest on S10: ~14100 (Version 1.424) sd S9 GA on S10: 13900 (Version 1.397) IOPS are rounded to the nearest 100, and CPU utilization was 99% on all runs. I realize my config is quite different from yours, but if this was a simple case of "code bloat" between S9 and S10, I should have seen the same issue. Since your V880 has four CPU's and my Ultra 60 only has one, this could be related to synchronization -- I don't have a larger Sparc system to test with right now unfortunately. Note that the S9 version of sd you are testing (1.397) is actually from the S9 GA release (circa 2002). sd has been patched several times since then (27 times going by the version number) and it would be interesting to know what your results are using the latest version of the S9 sd. What CPU utilization did iometer show when you tested? Are you using Sun FC HBA's or third party? If your system is at 100% cpu then we should be able to use dtrace to isolate the performance change. -Peter This message posted from opensolaris.org _______________________________________________ perf-discuss mailing list perf-discuss@opensolaris.org
