Special VDEV

[sonoracomm]

Hello and thank you in advance for all replies.

Regarding Napp-IT and Optane Special VDEV:

http://napp-it.org/doc/downloads/special-vdev.pdf

(I read this several times.)

Background
==========

I recently purchased a pair of 480GB Optane 900p for a primary storage server with Napp-IT Pro.

I did a fair amount of benchmark testing with the drives partitioned with 3x 20GB and 1x 400GB partitions intending to use the 20GB partitions as SLOGs and the 400GB partitions as a mirrored Special VDEV for metadata and small files on a large(ish) 7200rpm disk pool.

I added 2x 20GB SLOG partitions (one from each SSD) to three different pools with great success...dramatic performance improvement over the old SSD SLOG drives.

Questions
========

However, when I added the mirrored Special VDEV (simple default settings), I did not note a benchmark performance improvement. In fact, the benchmark performance decreased, so I removed the Special VDEV and I'm trying to understand why the performance appears to have decreased?

Is there something wrong with my plan?

Is there something wrong with my testing plan?

Is the noted benchmark performance decrease the whole story?

Is the noted benchmark performance decrease due to the fact that the mirrored Special VDEV is based on the same pair of SSDs as the SLOGs?

Any insights you might have?

Thanks in advance,

G

 

[gea]

I would expect

- there is a performance degration if you use a disk for Slog and other io.
With Optane this should be minimal

- a special vdev for metadata
All metadata entries for "last accessed data" is in Arc what is RAM and much faster than a special vdev for metadata. I am not sure if metadata in Arc is preferred over special vdev but in general a special vdev for metadata helps mainly in situations where metada is not in Arc (RAM). This is the case with high volatile data for many users ex a server with many users. If you repeatedly access same data, a special vdev will not help unless you force a filesystem to use a special vdev for all data (recsize is the key)

A special vdev for metadata (without forcing a whole filesystem to the special vdev) was developped by Intel for "big data" not for soho/lab use as an alternative to data tiering (move date between a fast and a slow array area)