Le 04/02/10 20:26, Tonmaus a écrit :
Yes, we mixed them among controllers and pools.Hi again, thanks for the answer. Another thing that came to my mind is that you mentioned that you mixed the disks among the controllers. Does that mean you mixed them as well among pools? Unsurprisingly, the WD20EADS is slower than the Hitachi that is a fixed 7200 rpm drive. I wonder what impact that would have if you use them as vdevs of the same pool.Cheers, Tonmaus We've done something that's not recommended : a 15 disk raidz3 pool. Disks are as follows : c3 (LSI SAS) has : - 1x 64 GB Intel X25E - 3 x 2TB WD20EADS - 4 x 2TB Hitachi c2 (LSI SAS) has : - 4 x 2TB WD20EADS - 4 x 2TB Hitachi c5 (motherboard ICH10 if I remember well) has : - 1x160GB 2,5'' WD - DVD All the 2TB drivers are in the raidz3 zpool named tank (we've been very innovative here ;-). X25E is sliced in 20GB for the system, 1GB for ZIL for tank, the rest as cache for tank. The 2,5'' 160GB WD was not initially part of the setup since we were planning to slice the 2TB drives in 32GB for the system (mirrored accross all drives) and the rest for the big zpool, while the X25E was just there for the ZIL and the cache, but two things we've read on lists and forums made us change our minds : - the disk write cache is disabled when you're not using the whole drive - some reports on this list about X25E loosing up to 256 cache flushes in case of power failures. So we bought this 160GB disk (it was really the last thing that could fit in the chassis) and sliced it in the same way as the X25E. The system and the ZIL are mirrored between the X25E and the WD160. We do not use the WD160 for the cache : we thought it would be better to save IOPS on this disk for the ZIL mirror. I don't know wether it's a good idea to mirror the ZIL on such a disk but we prefer having slower setup and not loose that much cache flushes on power failure. Regarding the perfs obtained by using only Hitachi disks, I can't tell, I haven't tested it, and can't do it right now as the system is in preproduction testing. Also, I should have mentionned in my previous post that some WD20EADS (the 32SB0) have shorter reponse times (as reported by iostat). They're even "faster" than the Hitachi : I've seen them quite a few times in the range 0.3 to 1.5 ms, which seems far to short for this kind of drives. I suspect they're sort of dropping flush requests. Add to it that 2 out of 3 failed WD20EADS were 32SB0 and you get the picture... Note they might also be hybrid drives with some flash memory which allows quick acknoledgment of writes, but I think we would have heard of such a feature on this list. Arnaud |
_______________________________________________ zfs-discuss mailing list zfs-discuss@opensolaris.org http://mail.opensolaris.org/mailman/listinfo/zfs-discuss
