Shared ZIL between pools?

[hailthemelody]

Hi,

I'm looking for share a single (or mirrored) ZIL device between multiple pools. Is this possible? I read on a thread on another site from back in 2013 (https://forums.freebsd.org/threads/42691/) that it's possible via different partitions, but thought there may be a more up-to-date answer/approch now.

Any thoughts? Would it be limited to up to 4 pools since drives can only have up to 4 partitions?

[raulfg3]

Please explain why you need to share ZIL, and if you do test of your actual performance.

[hailthemelody]

Sure. I'd like to (not need to) share a ZIL so that I can maximize the amount of disks I can use in my environment. In the environment in question, I currently have 5 pools of 5 disks each (not including ZIL or L2ARC.) Currently, each pool has it's own dedicated ZIL and L2ARC SSDs. I'd keep an L2ARC SSD per pool, but if I could use a single (or mirrored) ZIL between multiple pools I would have enough free bays to add a 6th pool of 5 disks.

Edit: By single/mirrored ZILs I mean a single (or two, if mirrored) disks, separated by partitions.

[Onichan]

Partitioning the ZIL is still the only way. What are you using these pools for, are you sure a ZIL is even helpful? Also you should be using GPT which has a 128 partition limit, not 4.

Also the device you choose for your ZIL makes a big difference https://b3n.org/ssd-zfs-zil-slog-benchm ... omparison/

[hailthemelody]

Wow - thank you very much for your post and link Onichan.

The pools are used for VMs with various purposes such as voip, web apps, web hosting, light database etc. The pools are 5x 1TB HGST 7200RPM SATA drives running RAID 10 with one hot spare, meaning 2TB usable each. Obviously SAS would be better, but I'm trying to keep it economical. In front of each pool is a 240GB Toshiba OCZ Trion 150 SSD used for L2ARC, and the SAN server has 96GB ram (going up to 144gb or 192gb in a year or so.) Currently, I have an additional SSD (a dedicated 240GB Toshiba OCZ Trion 150 SSD) in front of each pool for ZIL, which is what I'm rethinking in this thread.

The more I've been furthering my knowledge of ZIL the more I think I've been misusing it. My main intent of using ZIL was to lower wear-and-tear of my spinning disks (in my pools) by having the writes first cache to the ZIL, and also to protect from data loss in the event of an unexpected power outage. At present, during moderate read/write usage my spinning disks only "light up" every 4-5 seconds, so I think I've been doing it correctly (unless this has been the RAM's doing). However, I've since been learning more about SLOG and maybe this is more what I'm going for - would I need both ZIL and SLOG?

From my present understanding of your advice, I'm considering buying two Intel DC S3700's to work as a mirrored ZIL between all 6 pools, using GPT to make 15GB partitions for each pool (100GB / 6 pools, minus a few gb to spare). Is this a fair understanding of your advice, or perhaps should I get four Intel DC S3700 SSDs and use 2 mirrored pairs for 3 pools each to get better performance? I have room for that, if it will make a appreciable difference. Or another direction altogether to utilize SLOG?

[Onichan]

ZIL vs SLOG: the ZIL stands for ZFS Intent Log and it always exists for every pool, just by default it resides on the same disks as the pool itself. A SLOG is a basically just a separate/dedicated ZIL. This page explains it a bit more and talks about the purpose of the ZIL and synchronous vs asynchronous https://pthree.org/2013/04/19/zfs-admin ... ntent-log/

As far as using a SLOG for a VM pool, that is actually a reasonable idea for performance reasons. They would normally be synchronous writes so whatever is requesting the data to be written will generally wait for the confirmation it was successfully written before proceeding with whatever. SSD's fast IO and low latency can make the writes faster as it doesn't have to wait around as long for each write request, but like the previous article mentions it depends on the SSD as well. Also I think ZFS flushes writes every 5 seconds, which is why you are seeing the disk activity at those intervals. Though I think if you have enough incoming writes then it would be active more often, not too familiar with the 5 second flush thing.

Now I don't really see an SLOG helping with wear and tear, it should help make write requests be more sequential and less random, but that's unlikely to help much if at all as far as wear on the disks go.

Also as I mentioned the ZIL always exists so SLOG wouldn't necessarily be better to protect against power outage. That's the whole point of the ZIL.

Also just FYI not sure if you know, but a L2ARC requires RAM itself. Never looked into it much, but basically the system has to keep a table of what's currently in the L2ARC and that table is stored in RAM.

I'm not saying you should go buy some S3700 SSDs, just mentioning that the SSD does matter. Is this for business use? For personal or even a small business the S3500 should be more than enough. Also that was just the one site I know of that has those tests, it would be good to see if other people have done some SLOG tests with more SSDs. Also I personally haven't tested a 6 partition SLOG so before you buy anything I would test that ZFS is even happy with that. As far as if how many pools you should put on one set of mirrored SSDs, that I have no idea. Those S series are the datacenter ones designed for heavy use so they should be fine with sustained long term use, but it really depends on how write busy your pools are.