All SSD - any benefit of L2ARC

[memhog]

Every so often I see a terrific deal on SSD's and wonder about an all SSD head unit that could also offer expansion. My concept is to have a 1U rack with 8 2.5" hot-swap drives.

32 GB system - dual 1G Enet
Raid-1 2 drives for OS/Swap/VM's
Raid-Z1 5 drives (480GB) for storage - with compression

I have one slot available - Is there an application architecture that would benefit from that L2ARC / SLOG? (iScsi, VMware, database, etc.)

I can't think of an advantage of having an L2ARC (or separate SLOG) since I believe it would be at least as fast to get/put whatever pages needed reloading/ from the SSD array.

The other benefit of not having a L2ARC is all the RAM being available for the ARC itself.

---------
I'm not sure how I would utilize the last slot. My thoughts would be that if I wanted an external HDD JBOD for large storage I would then populate it with a SSD for SLOG for that JBOD.

Yes, I know that this would only offer a relatively small amount of storage for now. The price of $270 for 480GB of Sandisk II extreme SSD's just got my attention again (1.75TB @ $1350). I just offered this as an example and did not intend to get into the merits of a discussion of different SSD's.

 

[jgreco]

The write path to a SLOG device is simpler than the write path to the in-pool ZIL. The write performance of an all-SSD pool is likely to be disappointing.

 

[memhog]

The write path to a SLOG device is simpler than the write path to the in-pool ZIL. The write performance of an all-SSD pool is likely to be disappointing.

I appreciate your insights and I may be misinterpreting what you mean by "write path"

I personally have not had the chance to benchmark.
Using this as a reference (only of throughput from Mem to SSD Zpool) it quotes a write speed of 721MB write to a 6 disk Z2 pool.
1/3 of the way down the page is "Solid State (Pure SSDs) RAIDS"
https://calomel.org/zfs_raid_speed_capacity.html
If that number is achievable would there be another benefit to using / needing a separate SLOG?

Is it possible that the throughput is higher but also the latency is much higher?

 

[jgreco]

To be totally fair, you'd want to see how fast the writes are with sync=always set on the target. I have no doubt that ZFS could easily blast 1GB/sec or more to an SSD pool, but some of us have HDD arrays capable of such speeds - but sync is still an issue. The problem is that sync writes require that a given write be committed to stable storage, in the SLOG-less case, that would be the in-pool ZIL, and for a RAIDZ2 pool I expect that means a minimum of three SSD's (block plus two parity blocks) being involved, and if the SSD's are already busy with other traffic, then... suck?

Just thinking about the latency gets you your answers.

 

[memhog]

To be totally fair, you'd want to see how fast the writes are with sync=always set on the target. I have no doubt that ZFS could easily blast 1GB/sec or more to an SSD pool, but some of us have HDD arrays capable of such speeds - but sync is still an issue. The problem is that sync writes require that a given write be committed to stable storage, in the SLOG-less case, that would be the in-pool ZIL, and for a RAIDZ2 pool I expect that means a minimum of three SSD's (block plus two parity blocks) being involved, and if the SSD's are already busy with other traffic, then... suck?

Just thinking about the latency gets you your answers.

Thank you, those are aspects I hadn't considered.
I'm thinking now what I was looking for was turning the SLOG off but keeping the sync so that the applications can decide if it is async or sync writes.
You are right, latency will definitely be the key as to if this arch will perform well.

Would you help me understand where the "other traffic" might come from?
If the system normally writes to the SLOG (and later flushes the equivalent of the slog from RAM to disk) isn't the total traffic to the disk going to be the same using my example All SSD?

Grateful for your insights.

 

[jgreco]

Other traffic like pool reads and pool writes. The ZIL writes are in ADDITION to those.