Terrible NFS performance under 9.2-RELEASE?

[Rick Macklem]

Garrett Wollman wrote:
> <<On Wed, 29 Jan 2014 18:01:43 -0500 (EST), Rick Macklem
> <rmacklem at uoguelph.ca> said:
> 
> > However, I do suspect we'll be putting a refined version of the
> > patch
> > in head someday (maybe April, sooner would have to be committed by
> > someone else). I suspect that Garrett's code for server read will
> > work
> > well and I'll cobble something to-gether for server readdir and
> > client write.
> 
> Once I can get this mps(4) issue ironed out, I should be in a
> position
> to get some real data on this.
> 
If you can check the network device driver you use and if it looks
like it has a scatter size of less than 36 (often a constant with
"TXSEG" or "TX_SEG" in the name) and calls either m_defrag() or
m_collapse(), adding a counter to see if those functions are being
called, would be nice.

If the m_collapse()/m_defrag() function is being called without
the patch and not with the patch, the performance difference may
be avoiding that call and not a more generic benefit.

I just did a quick find/grep and it looks like a lot of drivers
have *TXSEGS* set to around 32 and then call one of two functions
for more than that. Since without a patch, 64K NFS reads/writes
hand sosend() an mbuf list of 34 entries, it seems like this could
be happening a lot. (I didn't look to see which ones set if_hw_tsomax
to significantly less than 64K.)

Thanks for working on this, rick.
ps: you might want to combine your patch with mine, so readdir and
    client side writes use 4K clusters.

> -GAWollman
> 
>