Re: Is tcp autotuning really what NFS wants?

[NeilBrown <neilb@suse.de>]

[-- Attachment #1: Type: text/plain, Size: 1202 bytes --]

On Mon, 15 Jul 2013 09:42:11 -0400 Jim Rees <rees@umich.edu> wrote:

> Here's the thread on netdev where we discussed this issue:
> http://marc.info/?l=linux-netdev&m=121545498313619&w=2
> 
> Here is where I asked about an api to set a socket buf minimum:
> http://marc.info/?l=linux-netdev&m=121554794432038&w=2
> 
> There are many subtleties, and I suggest anyone who wants to try to fix this
> code should read the email thread. The netdev people were pretty insistent
> that we turn on autotuning; David Miller said the old behavior was
> equivalent to "turn[ing] off half of the TCP stack."

Thanks for those pointers.
Certainly autotuning is what we want.
The issues with the receive buffer getting tuned too low appear to have been
addressed some time ago.
The issues with the send buffer getting tuned too low are not what I at first
thought they were, but still are not completely resolved.
My previous patch which allowed N requests beyond the available space is, I
now think, more than necessary.  You just need to make sure one can get
though.  That should be enough to push the sndbuf size up quite quickly.

I'll revise it, test, and repost.

Thanks,
NeilBrown

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 828 bytes --]