I'm reading about similar issues with zfs + nfs + vmware esxi; so this is definitely NOT a SmartOS problem. It's an unfortunate problem of mixing technologies that do things differently.
My view is that regardless of the upper level applications (NFS or KVM or whatever), if your zpool is going to be doing lots of sync writes, you design your resources to accommodate. That means either stripped SSDs or ZeusRAMs as an SLOG. I'm curious what kind of risk one is taking for disabling sync on a zvol for an nfs mount of typical MS Office type files, no database type transactions? Others will probably have more precise answers, but the primary risk is the loss of writes. ZFS by default buffers the writes into RAM, but sync writes will only complete once the data is on persistent storage. ZFS remains compliant to this requirement by using the ZIL (ZFS Intent Log) which by default lives in the normal zpool disks or on the SLOG if you have one. By setting sync=disabled your effectively disabling the ZIL and the only copy that exists of your writes before they are committed to the spinning disks is in RAM, which will be lost in the event of a power failure. It isn’t particularly about what the application is (Office/Database etc) but more what kind of I/O semantics the higher level applications are expecting the underlying storage to adhere to. If application X gets the okay from the other end to say ‘your write is done and safe’ and suddenly the other end loses power before committing the write to disk, how will application X recover/behave? Thats a gross simplification of the scenario, but the basic question is how much risk are you happy to run with in your environment when it comes to data integrity? I think the only recommended scenario where sync=disabled is usable is when your virtual machine is entirely stateless and just needs really high performance I/O. On 19 Jan 2015, at 9:38 am, Greg Zartman <[email protected]<mailto:[email protected]>> wrote: On Sun, Jan 18, 2015 at 3:14 PM, David Finster <[email protected]<mailto:[email protected]>> wrote: Having the problem ‘solved’ is probably subjective based on what your I/O expectations/requirements are. I'm looking for something that will at least give me a workable disk speed. Right now, the disk speed is so slow that a web management GUI that's been used widely for 20 years times out and doesn't work. I'm reading about similar issues with zfs + nfs + vmware esxi; so this is definitely NOT a SmartOS problem. It's an unfortunate problem of mixing technologies that do things differently. I'm curious what kind of risk one is taking for disabling sync on a zvol for an nfs mount of typical MS Office type files, no database type transactions? Adding at least one write-optimized SLOG should improve the situation significantly. Adding another SLOG should improve things further (more fast I/O capacity available for sync writes) but perhaps not to the same degree. Here we use Intel 3700 SSDs (the 200G version), which also have power protection, and they seem to work rather well. I've only used Samsung type SSDs, but I'll get one of these and give it a shot and see how it helps performance. Greg ------------------------------------------- smartos-discuss Archives: https://www.listbox.com/member/archive/184463/=now RSS Feed: https://www.listbox.com/member/archive/rss/184463/25769125-55cfbc00 Modify Your Subscription: https://www.listbox.com/member/?member_id=25769125&id_secret=25769125-7688e9fb Powered by Listbox: http://www.listbox.com
