ZFS pool
in [Solaris]
|
Prev: jumpstart requests wrong file
Next: Panic strings
From: Canuck57 on 28 Mar 2010 13:01 On 26/03/2010 3:33 PM, Ian Collins wrote: > On 03/27/10 02:55 AM, Michael Laajanen wrote: >> Hi, >> >> A ZFS vdev should be up to 9 disk I have read, what is the disadvantage >> with having 16 disk in one vdev, compared to two vdevs with 8 disk each >> all set to raidz2? > > Two important things will suffer: > > Performance: will suck compared to a stripe of smaller vdevs. > > Reliability: you expose your self to a much greater risk or multiple > drive failures. Resilver times will also be slower, compounding the risk. How would risk of drives be an issue? Say one raidz2 of 16 versus 2 of 8 each? I would really hate to think that if one drive failed, it's mirror is spread across all others as if so it is inferior to traditional RAID 0+1. For any mirror to totally fail, it's opposing mirror must also fail. So spreading the risk across 15 other disks...stupid as any 1 in 15 can fail and it is dead. Where as the probablitities of the exact opposite is 1/15. I dont see the risk. And larger stripes should help also. But admit, I am no ZFS expert. But if doing this with SVM I would raid 0+1 the 16 disks as one on opposing disks assuming I needed all the space as one. -- -------------- Politicians don't provide anything, the tax payers do.
From: Canuck57 on 28 Mar 2010 13:03 On 27/03/2010 1:17 PM, Ian Collins wrote: > On 03/28/10 07:55 AM, Michael Laajanen wrote: >> Hi, >> >> Ian Collins wrote: >>> On 03/27/10 02:55 AM, Michael Laajanen wrote: >>>> Hi, >>>> >>>> A ZFS vdev should be up to 9 disk I have read, what is the disadvantage >>>> with having 16 disk in one vdev, compared to two vdevs with 8 disk each >>>> all set to raidz2? >>> >>> Two important things will suffer: >>> >>> Performance: will suck compared to a stripe of smaller vdevs. > >> What is the optimal number of disks in a vdev? > > It depends what you want to do with the pool. I never go beyond 8 in a > raidz2 configuration. > >>> Reliability: you expose your self to a much greater risk or multiple >>> drive failures. Resilver times will also be slower, compounding the >>> risk. >>> >> Then the HBAs must also be concidered I guess. > > Not really, they fail way less often than drives. It would still be good to balance the load to use both HBAs as equally as possible. System can write to 2 HBAs faster than one. -- -------------- Politicians don't provide anything, the tax payers do.
From: Ian Collins on 28 Mar 2010 14:30 On 03/29/10 06:01 AM, Canuck57 wrote: > On 26/03/2010 3:33 PM, Ian Collins wrote: >> On 03/27/10 02:55 AM, Michael Laajanen wrote: >>> Hi, >>> >>> A ZFS vdev should be up to 9 disk I have read, what is the disadvantage >>> with having 16 disk in one vdev, compared to two vdevs with 8 disk each >>> all set to raidz2? >> >> Two important things will suffer: >> >> Performance: will suck compared to a stripe of smaller vdevs. >> >> Reliability: you expose your self to a much greater risk or multiple >> drive failures. Resilver times will also be slower, compounding the risk. > > How would risk of drives be an issue? Say one raidz2 of 16 versus 2 of 8 > each? All your eggs are in on basket. If you have 2 vdevs, you stand a reasonable chance of surviving 3 failures and a smaller, but real chance of surviving 4. -- Ian Collins
From: Michael Laajanen on 29 Mar 2010 11:15 Hi, Canuck57 wrote: > On 26/03/2010 3:33 PM, Ian Collins wrote: >> On 03/27/10 02:55 AM, Michael Laajanen wrote: >>> Hi, >>> >>> A ZFS vdev should be up to 9 disk I have read, what is the disadvantage >>> with having 16 disk in one vdev, compared to two vdevs with 8 disk each >>> all set to raidz2? >> >> Two important things will suffer: >> >> Performance: will suck compared to a stripe of smaller vdevs. >> >> Reliability: you expose your self to a much greater risk or multiple >> drive failures. Resilver times will also be slower, compounding the risk. > > How would risk of drives be an issue? Say one raidz2 of 16 versus 2 of > 8 each? > > I would really hate to think that if one drive failed, it's mirror is > spread across all others as if so it is inferior to traditional RAID > 0+1. For any mirror to totally fail, it's opposing mirror must also > fail. So spreading the risk across 15 other disks...stupid as any 1 in > 15 can fail and it is dead. Where as the probablitities of the exact > opposite is 1/15. > > I dont see the risk. And larger stripes should help also. > > But admit, I am no ZFS expert. But if doing this with SVM I would raid > 0+1 the 16 disks as one on opposing disks assuming I needed all the > space as one. > Below is my write tests using two BHA each having two channels with 4 drives, all standard E450. 4 400MHz and 4GB RAM. This is only write ofcourse. bash-3.00# time mkfile 1000G /pool00/foo non raidz real 368m55.000s user 6m1.477s sys 329m37.046s raidz2 1 vdev real 475m48.784s user 6m6.856s sys 338m45.906s raidz2 2 vdevs time mkfile 1000G /pool00/foo real 485m18.255s user 5m57.202s sys 346m38.191s /michael
From: Ian Collins on 29 Mar 2010 15:15
On 03/30/10 04:15 AM, Michael Laajanen wrote: >> > Below is my write tests using two BHA each having two channels with 4 > drives, all standard E450. 4 400MHz and 4GB RAM. > > This is only write ofcourse. Try a tool like bonnie++ to get a better idea of the all round performance. -- Ian Collins |