I agree with Jim, we need some numbers to help. I would recommend also looking not just at 'iostat' but also 'fsstat' to get a better idea of what the IO load is like on an op basis.
Some questions and suggestions come to mind: 1) Have you disabled atime on the dataset(s)? (zfs set atime=off pool) 2) How much physical memory do you have? 3) How do successive runs look? 4) Did you enable compression? On thing to consider regarding ZFS reads is that there is prefetch involved. If you have a good amount of physical memory for caching this can really help, and ZFS's aggresive caching (ZFS ARC) will make read IO over time very responsive, however in the short term (such as single run sequential benchmarks) will look slower than more simplistic filesystems such as UFS. If you want to put this to the test, consider disabling prefetch and trying again. See http://www.solarisinternals.com/wiki/index.php/ZFS_Evil_Tuning_Guide I assume your write caching in the controller, which is why writes always look speedy. As a side note, I'd suggest using a more compressive benchmark which emulates real-world workloads such as FileBench or IOmeter, as opposed to dd's unrealistic sequential benchmark. (And if you must, use /dev/null not /dev/zero.) benr. roland wrote: > Hello, > > i have a very weird problem on a FSC RX300 Server machine (LSI Raid > controller) > > i first came across this when trying Nexenta-OS, but the problem remains with > nv_109 and nv_110. > > the problem is: > Writing to a ZFS volume is ok, but read performance sucks big. (slower by a > factor of 5-10, maybe even more) > > if i format the same volume/slice with ufs, i don`t see this issue. > > the disks are hardware raid1 and raid5 volumes and i see the same behaviour > on both. while reading from the disks (dd if=/zfs/path/to/file of=/dev/zero) > , using mpstat i don`t see anything weird - there just isn`t "happening" > anything -no large time spent in usr/sys and no large values for wait-I/O, no > excessive high interrupts or context switching.... > > so, why is it slow and where does solaris spend it`s time when reading from > disk? > > hints ? > > maybe scheduling/timer issue ? > > i don`t have a clue where to look at..... > > regards > roland > _______________________________________________ perf-discuss mailing list perf-discuss@opensolaris.org
