Assessing device value & performance as SLOG: Samsung 983 ZET vs Optane 905P.

[TrumanHW]

According to a review on Tom's Hardware, the Samsung 983 ZET is the closest performer to an Optane, (some times faster than an Optane) ... in select categories...well, to ME at least. Maybe though, the exact areas in which an Optane most decisively outperformed the 983 ZET are precisely those indicative & responsible for Optane's high performance as a SLOG.


Are the below assumptions correct..?
Would a SLOG almost always be either: [Writing] ... or ... [Reading + Writing]..?

What metrics best indicate which device (or why a device) will work well as a SLOG ..?
• Mixed R/W performance..?
• I'd assume most R & W will be sequential...as in, no specific reason causing Random IO...
-- even if data can be analyzed in so little time ... it would seem like like it could be read-out sequentially unless the array isn't ready..?
• What causes data to have a low QD ..?
• I'd assume the data-sizes will basically be individual ... except for the extra metadata and perhaps the hash file.?
• One other reason perhaps the SLOG might give a performance advantage (though I haven't seen it mentioned) may be that it allows faster calculation of the hash file..?

Even if a SLOG needs to do 'Random Reads' so long as the SLOG's read-speed exceeded the array's collective write speed (for the given-data) it shouldn't matter, no?
According to calomel.com's benchmarks for different array-types ... even a striped array of 24 7200rpm HDs can't exceed 700MB/s (when exclusively-writing) ..?

ZFS Benchmarks:
Copied from Calomel.com

​

Quantity​	Capacity ​	Config.​	Capacity​	WRITE​	Read + Write​	READ​
24 Drives​	4TB​	12 striped mirrors	45.2 TB​	696 MB/s​	144 MB/s​	898 MB/s​
24 Drives​	4TB​	RAIDz1 (RAID-5)	86.4 TB​	567 MB/s​	198 MB/s​	1304 MB/s​
24 Drives​	4TB​	RAIDz2 (RAID-6)	82.0 TB​	434 MB/s​	189 MB/s​	1063 MB/s​
24 Drives​	4TB​	RAIDz3 (RAID-7)	78.1 TB​	405 MB/s​	180 MB/s​	1117 MB/s**​
24 Drives​	4TB​	STRIPED RAID-0	90.4 TB​	692 MB/s​	260 MB/s​	1377 MB/s​

THANKS!







After thoughts:
It'd be a nice option for some people at least ... for mirrored array's could be configured to separate their Read Write tasks ... and upon completing the Read ... synchronize the Write data across the other mirror ... in that way their Read + Write performance would be the aggregate of their peaks...at a small increase of risk...which the user could define: (maximum time before mirroring data: 5 min, an hour, whatever they're comfortable with)...

Allowing mirrored array's to offer a R+W (mixed) performance option to sync written data after the read-task completes..?
This'd allow each mirror to provide the READ-ONLY performance from one array & the WRITE-ONLY from the other..?

 

[Constantin]

Have you considered a refurbished RMS-200? Much faster than either option and about the same price.

The SLOG is only written to unless the system has a power failure or some other catastrophic failure in which case the data is read back from the SLOG and then written to disk. No idea about all the other factors you mention. I'd focus on

• getting something that is power-loss-protected (which the SAMSUNG 983 ZET and most Optane are).
• planning how to use the thing - partition the ZET into multiple smaller pieces to set aside a SLOG, Metadata sVDEV, etc?

The SLOG usually needs very little room - 10's of GB and that 983 comes in much larger capacities - so what to do with all that extra room? Partitioning for metadata and small files might make sense. See here for a great primer re: SLOGS from @Stilez .

 

[TrumanHW]

Have you considered a refurbished RMS-200? Much faster than either option and about the same price.

Awesome. I was only able to find the 8GB version ... is that adequate..?
How big of a SLOG can be beneficial..? as in, what's the largest SLOG that is likely to provide much benefit..?
I'm assuming a SLOG isn't just a fast temporary layer to write all incoming data but rather, specific kinds of data..?
(I'm assuming a SLOG does a specific task that more RAM can't mimick)

An L2arc directly handles Reads for (only) frequently used data, yes..?
An L2arc will 'improve' reads & writes ... but only by preserving the resources to access performance by offloading commonly read tasks to the L2..?

Thanks!

 

[TrumanHW]

Also it looks like NVDIMMs are crazy fast ... and 16GB isn't that expensive. (though I'm wondering if a SLOG is even relevant to the kind of use I have for this FreeNAS machine... as it's just for my personal use).