Yes, to avoid exactly the type of question resultant from using a crummy SLOG. The results do indeed seem to indicate the SLOG isn’t your problem.
It’s pretty clear that the pool you have is very much write limited. I suspect you have a problem stemming from the fact that these consumer drives are limiting your write performance for one reason or another. If any are QLC drives, your write performance is expected to fall off a cliff once the internal SLC cache is exhausted. The same is true for even TLC drives which are more common, but the drop off isn’t usually this severe.
Did you allow the individual disk benchmarks to run? Since you have a mix of random different drives, it may be helpful to identify if one of the models slowing down the pool.
Thanks for this, I will fix in the next version.
He used the script that I wrote which is well documented in the link I posted. I’m fairly confident in his results.
He ran the test with my script which creates a dataset with 1M record size, no compression and sync disabled. While his SLOG may not be appropriate for use as a SLOG, thats definitely not the core of the problem. His writes are far lower than you’d expect for sequential 1M writes.
I suspect at least some of the drives he has are QLC sata drives. I haven’t looked, but the results certainly point in that direction.
yes I’m aware of the crummy slog drive choice as for the drives they are all SLC drive at least from the listings i bought them from. I’ll be nuking the pool today and individually testing each and every drive. As for for slog. nvme choices for slog drive recomendations are welcome.
Maxing your RAM is 10000000x faster than slog (okay, not really, but it is WAY better than using SLOG unless you have a very specific use case for SLOG.
In this situation, his SLOG device is probably more than capable of serving that role, given the maximum write performance of his pool is between 200-400 MB/s depending on thread count.
NVME latency is far better than SATA in general, and the amount of data the pool is actually capable of writing is small enough that I wouldn’t expect the SLOG to introduce much additional bottleneck if the same test was ran with sync=always.
Of course, the write performance of this pool is lackluster for what the @brandon_evans was expecting, so its probably a moot point. But I say all of this to say, if those were HDDs I’d have no qualms with the same drive being used for a SLOG in a desktop’s Steam drive LUN like this appears to be.
im going to need more pci lanes for that to work sad are the any good options for m.2 drives i know just from looking that slog is not going to work not without some serious rebuilding of the entire server really kicking my self in the face for not going AMD epic on my server…
SUPERMICRO MBD-X13SAE-F-B Motherboard anyone got any good or bad news about this motherboard its got the lanes…
ok after nuking the pool and testing each brand/model of SSD I’ve discovered that the PNY drives are definitely NOT SLC. Ran Crystal Disk Mark on them and to say there write speed is slow is an understatement they burst upto 350 MB/s and then drop to a sad 90 MB/s average in less than 5 seconds ouch.
drive benchmarks by brand
Profile: Default
Test: 64 GiB (x1) [E: 0% (0/932GiB)]
Mode: [Admin]
Time: Measure 5 sec / Interval 5 sec
Date: 2025/01/13 18:08:38
OS: Windows 11 [10.0 Build 22631] (x64)
WD Red SA 500 best performance
Continuous writes at 370MB/s without slowing down
[Read]
SEQ 1MiB (Q= 8, T= 1): 439.37 MB/s
SEQ 1MiB (Q= 1, T= 1): 417.8 MB/s
RND 4KiB (Q= 32, T= 1): 240.27 MB/s
RND 4KiB (Q= 1, T= 1): 27.22 MB/s
[Write]
SEQ 1MiB (Q= 8, T= 1): 390.99 MB/s
SEQ 1MiB (Q= 1, T= 1): 370.85 MB/s
RND 4KiB (Q= 32, T= 1): 141.2 MB/s
RND 4KiB (Q= 1, T= 1): 40.34 MB/s
=============================================
T-force Vulcan Z NOT there QLC model I looked
not as good but doesn't drop off a cliff after
being loaded for 5 seconds
douse have a cash for writes.
Have to write at max speed for over a minute
for it to slow down
[Read]
SEQ 1MiB (Q= 8, T= 1): 439.54 MB/s
SEQ 1MiB (Q= 1, T= 1): 520.24 MB/s
RND 4KiB (Q= 32, T= 1): 159.24 MB/s
RND 4KiB (Q= 1, T= 1): 21.88 MB/s
[Write]
SEQ 1MiB (Q= 8, T= 1): 219.96 MB/s
SEQ 1MiB (Q= 1, T= 1): 370.85 MB/s
RND 4KiB (Q= 32, T= 1): 141.2 MB/s
RND 4KiB (Q= 1, T= 1): 26.69 MB/s
============================================
PNY CS900 SSD (avoid like the plague)
NO Cash CHIP just 2 Nand chips TA8HG63AWN
R2241 1004070 these are 500 GB Nand chips
write performance. what write performance
strait to SDD speeds with this drive
[Read]
SEQ 1MiB (Q= 8, T= 1): 423.01 MB/s
SEQ 1MiB (Q= 1, T= 1): 400.65 MB/s
RND 4KiB (Q= 32, T= 1): 236.86 MB/s
RND 4KiB (Q= 1, T= 1): 26.33 MB/s
[Write]
SEQ 1MiB (Q= 8, T= 1): 94.35 MB/s
SEQ 1MiB (Q= 1, T= 1): 92.69 MB/s
RND 4KiB (Q= 32, T= 1): 55.98 MB/s
RND 4KiB (Q= 1, T= 1): 21.74 MB/s
============================================
Great work digging. The best path forward would be to standardize on the WD Red SA 500 drives if you’re planning on spending some money. While not quite to the level of the Intel drives @etorix mentioned, they are marketed and designed for this use case where the other two drive models are clearly not.
yea im going for the The WD Red SA500 umm not spending an arm and a lag for the intel ones. As for the PNY Crapy drives. well lets just say i found a new door stop. now to rebuild the SSD pool
I have accumulated a number of second-hand 3.84 TB Intel/Micron/Samsung DC SSD drives for my own flash pool. They do not cost and arm and a leg, and have a lot of life remaining even when no longer at 100% health. As you’re building with 1 TB drives, 960 GB and 1.92 TB sizes look like a very realistic option.
So i’ve got the pool rebuilt and it works wonderfully now and is working to my standards.
I ran a benchmark on it using crystal benchmark over SMB and this is the results is it faster yes. Could it go faster yes if i went intel SSDs. Im not going intel SSD for a long time. expense is to high. even used. this NAS is used for home use. Not some business prepose. so unless they fall out of the sky for free it is’nt happening.
