Questions regarding Log drive

Ralms

Dabbler
Joined
Jan 28, 2019
Messages
29
Hi everyone,

So, I'm using some NFS mounts synced with FreeNAS, which currently are slow to a crawl, as the raid is simply 5 x 3TB HDDs

Due to that, I've seen around that adding a Log drive to the pool, which should dramatically help with this.
I've decided to go with an Nvme drive on my HP DL380p server and to achieve that I will be trying using an EZDIY-FAB NVME PCIe Adapter (the simple one, not the RGB versions lol).

My budget for the Log drive is up to 150€ max.
Right now the one I've been leaning to is an Intel Optane 800P 60G which I can get for around 140€ from Amazon Germany.

The questions I have are:
  1. Will the Pool sequencial read and write speeds be limited by the Log drive?
  2. What is more important, read or write latency? or both?
  3. In an 15TB pool in Raid Z1, how much capacity is really needed on the Log drive? I heard is really low in the single digit GB.

Also, If you have any other suggestion for other drives would be welcome, I was thinking on using my 960 Evo 250GB, but seems to have fairly high latencies.
And I don't need an enterprise battery powered Log drive, as the server is UPS powered, most of those drives even used in the market don't have great price to performance value.

Thank you,
Ralms.
 

HoneyBadger

actually does care
Administrator
Moderator
iXsystems
Joined
Feb 6, 2014
Messages
5,110
  1. Will the Pool sequencial read and write speeds be limited by the Log drive?
  2. What is more important, read or write latency? or both?
  3. In an 15TB pool in Raid Z1, how much capacity is really needed on the Log drive? I heard is really low in the single digit GB.

0. RAIDZ1 is not the optimal vdev if you are after performance; mirrors are.
1. Adding SLOG will not impact reads directly, but will definitely help writes.
2. Depends on your workload, but most of them bottleneck on writes.
3. No more than 4G with default tunables.

And I don't need an enterprise battery powered Log drive, as the server is UPS powered, most of those drives even used in the market don't have great price to performance value.

Yes you do; a UPS will not protect you from your PSU failing. The battery-backed or power-loss-protected status of your SLOG device is primarily how it will achieve good results, as the limit to how fast an SLOG device is in practice is limited by "how fast can it respond to a request to flush cache to storage."

The Optane 800p is fine for this with the caveat that its write endurance is only 365TBW over 5 years - if you need more than this in M.2 form factor prepare to spend significantly more to get an Optane P4801X (110mm)
 

Ralms

Dabbler
Joined
Jan 28, 2019
Messages
29
0. RAIDZ1 is not the optimal vdev if you are after performance; mirrors are.
1. Adding SLOG will not impact reads directly, but will definitely help writes.
2. Depends on your workload, but most of them bottleneck on writes.
3. No more than 4G with default tunables.

Yes you do; a UPS will not protect you from your PSU failing. The battery-backed or power-loss-protected status of your SLOG device is primarily how it will achieve good results, as the limit to how fast an SLOG device is in practice is limited by "how fast can it respond to a request to flush cache to storage."

The Optane 800p is fine for this with the caveat that its write endurance is only 365TBW over 5 years - if you need more than this in M.2 form factor prepare to spend significantly more to get an Optane P4801X (110mm)

Thank you for the information.

RaidZ1 for me has enough performance because is for home use only and also ends up having better then mirror because of very low CPU loads due to the low traffic. Also I'm usually limited by networking anyway, from Proxmox to FreeNAS is a 10Gbit link, but the rest is 1Gbit, so even if I only have around 500MB/s sequencial, that is plenty for what I need.

The workload is VMs disks and Docker containers, I have a couple NFS shares mapped on Proxmox.
Yeah, the major performance bottleneck seems to be with writes, because setting Proxmox VMs to have the Disk with Cache Write Back (I know is not the safest option) did help slightly, I enabled that just temporarily while I sort the LOG drive.

Regarding the PSU failure and data endurance, this server is being used as only a home nas, so it has fairly low amount of data written and I'm aware of the point of failure with the PSU, the server itself has redundant hot swap PSUs, so they would need to both fail at the same time which is very unlikely.
 

HoneyBadger

actually does care
Administrator
Moderator
iXsystems
Joined
Feb 6, 2014
Messages
5,110
Regarding the PSU failure and data endurance, this server is being used as only a home nas, so it has fairly low amount of data written and I'm aware of the point of failure with the PSU, the server itself has redundant hot swap PSUs, so they would need to both fail at the same time which is very unlikely.
The desire for your SLOG device to have PLP is from a performance need, not just safety. An SLOG device only "needs" PLP for data at rest. That ensures that data written there is safe. But a device with PLP-at-rest only will actually have to commit the data to NAND immediately when it receives a "flush data" which a sync write will generate.

A device with PLP-in-flight will be able to immediately respond "complete" to that request, because its RAM is protected by capacitors or other power hold-up, and then lazily flush the cache in the background. That's where the speed comes in.

So you're free to use a non-PLP SSD for SLOG; but you will get what you pay for. Have a look at the link in my signature for SLOG devices to get an idea of the difference between a "good" and "poor" SLOG.

https://www.ixsystems.com/community/threads/slog-benchmarking-and-finding-the-best-slog.63521/
 
Top