[OmniOS-discuss] re-tune round-robin reading from a mirror
Hello list; is it possible with OmniOS to have a multi-way mirror with one disk being an SSD and the rest magnetic; then to tune ZFS to perform all reads from the SSD? for the sake of performance. The default case is round-robin reading, which is the best if all disks are of approximately equal performance, especially if they're on separate controllers. But SSD changes that. __ Michael Mounteney ___ OmniOS-discuss mailing list OmniOS-discuss@lists.omniti.com http://lists.omniti.com/mailman/listinfo/omnios-discuss
Re: [OmniOS-discuss] re-tune round-robin reading from a mirror
This is a very interesting idea. It could allow for the creation of a scratch pool with great ROI. I have a particular need for an extremely high read rate pool for data analysis that will leverage the fattest read optimized SSDs I can get for the dollar. I was considering this to be raidz1, but if mirroring with disks would work this way, it could be even better bang for the buck. Did you ever do any actual testing with this type of setup? I'd love to see some real world performance data. -Chip On Wed, Jul 15, 2015 at 10:09 AM, Jim Klimov jimkli...@cos.ru wrote: 15 июля 2015 г. 14:10:15 CEST, Michael Mounteney gat...@landcroft.co.uk пишет: Hello list; is it possible with OmniOS to have a multi-way mirror with one disk being an SSD and the rest magnetic; then to tune ZFS to perform all reads from the SSD? for the sake of performance. The default case is round-robin reading, which is the best if all disks are of approximately equal performance, especially if they're on separate controllers. But SSD changes that. __ Michael Mounteney ___ OmniOS-discuss mailing list OmniOS-discuss@lists.omniti.com http://lists.omniti.com/mailman/listinfo/omnios-discuss When I last asked a few years ago (but IIRC forma mirror of local+iSCSI vdevs), the answer was along the lines that round-robin first considers the available devices. If the faster (ssd, local) device has no queue, it gets the load while the slower device still struggles with the task it has, so on average the faster device serves more io's - but not 100%. Queue depth tuning can also help here. Jim -- Typos courtesy of K-9 Mail on my Samsung Android ___ OmniOS-discuss mailing list OmniOS-discuss@lists.omniti.com http://lists.omniti.com/mailman/listinfo/omnios-discuss ___ OmniOS-discuss mailing list OmniOS-discuss@lists.omniti.com http://lists.omniti.com/mailman/listinfo/omnios-discuss
Re: [OmniOS-discuss] re-tune round-robin reading from a mirror
On Jul 15, 2015, at 11:45 AM, Schweiss, Chip c...@innovates.com wrote: This is a very interesting idea. It could allow for the creation of a scratch pool with great ROI. I have a particular need for an extremely high read rate pool for data analysis that will leverage the fattest read optimized SSDs I can get for the dollar. I was considering this to be raidz1, but if mirroring with disks would work this way, it could be even better bang for the buck. Did you ever do any actual testing with this type of setup? I'd love to see some real world performance data. There's a great (computer) science experiment in here waiting to be run. Setting up a 1-SSD + N-spinning-rust mirror (given the capacity mismatch, one could theoretically use N-slices of N-spinning-rusts) and then measuring read vs. write, which disks used, etc. would be spectacularly useful. It would also probably make a nice presentation at the next OpenZFS conference. Speaking of that, this question should be posed on the OpenZFS mailing list. It's possible someone in the wider ZFS community has considered this or even run tests already. Dan ___ OmniOS-discuss mailing list OmniOS-discuss@lists.omniti.com http://lists.omniti.com/mailman/listinfo/omnios-discuss
Re: [OmniOS-discuss] re-tune round-robin reading from a mirror
15 июля 2015 г. 14:10:15 CEST, Michael Mounteney gat...@landcroft.co.uk пишет: Hello list; is it possible with OmniOS to have a multi-way mirror with one disk being an SSD and the rest magnetic; then to tune ZFS to perform all reads from the SSD? for the sake of performance. The default case is round-robin reading, which is the best if all disks are of approximately equal performance, especially if they're on separate controllers. But SSD changes that. __ Michael Mounteney ___ OmniOS-discuss mailing list OmniOS-discuss@lists.omniti.com http://lists.omniti.com/mailman/listinfo/omnios-discuss When I last asked a few years ago (but IIRC forma mirror of local+iSCSI vdevs), the answer was along the lines that round-robin first considers the available devices. If the faster (ssd, local) device has no queue, it gets the load while the slower device still struggles with the task it has, so on average the faster device serves more io's - but not 100%. Queue depth tuning can also help here. Jim -- Typos courtesy of K-9 Mail on my Samsung Android ___ OmniOS-discuss mailing list OmniOS-discuss@lists.omniti.com http://lists.omniti.com/mailman/listinfo/omnios-discuss