[email protected] said:
> This could be a contributing factor, and may be worth testing, but I will
> mention that each of these R720's is connected a zpool comprised of a couple
> hundred SAS spindles each, and each R720 has ~192GB of memory.
> 
> So even if we don't have an external log device per se, I feel fairly good
> about the RAM-based ZIL doing a pretty good job of handling synchronous
> writes efficiently.
> 
> In addition -- wasn't sure how effective an external log device is at
> absorbing huge streams of sequential writes?  Seems like it would just get
> filled up and you'd eventually be constrained by your spindles anyway. 

Hi Ray,

When "zfs set sync" is "default" or "always", the cache flushes triggered
by synchronous writes go all the way out to non-volatile media, so server
RAM (ARC) is not very relevant.

For NFS, it's not the bandwidth or IOPS of your ZIL devices that kills
performance.  The latency of your server hardware is small compared to
the latency involved in the network round-trip of the synchronous write
traffic between NFS client and NFS server.

In my experience, even the amount of cache in an H800 (512MB or 1024MB)
will make a huge difference to the "tar x" workload from an NFS client
(unless you're extracting small numbers of big files, in which case there
is little synchronous traffic).  The early generations of NetApp's made due
with only a couple hundred MBytes of NVRAM, so it really doesn't take much.

Granted, my experience has been with gigabit ethernet, so 10GbE latency
could be good enough to alleviate the synchronous-write penalty.  However,
disabling synchronous writes on the NFS server is a quick way to see if this
is what's going on.  And if it is, a utility like "zilstat" will tell you how
much external log you might need.

Regards,

Marion


_______________________________________________
nfs-discuss mailing list
[email protected]

Reply via email to