[OpenAFS-devel] Andrew Deason's OpenAFS RX performance patches

[David Howells]

sweetpotatopie2021 <sweetpotatopie2021@protonmail.com> wrote:

> However I have a plea: AFS has never noted as a real speed demon for data
> transfer. But its lack of performance is cited as a contributing factor
> leading some who might otherwise use AFS to consider alternatives (smb, nfs,
> cloud).

One problem that I've been seeing with kAFS on Linux is that UDP packets
appear to get batched by the NIC.  The NIC may wait for, say, 32 packets to be
buffered before starting transmission.  This means that ACKs get batched.

I have patches to add tracepoints into a number of Linux NIC drivers if anyone
is interested in having a look for themselves.

I have a pair of Intel 82599ES 10G NICs, one in the server and one in my test
client.  It seems like they cheat for TCP in a couple of ways: firstly, ACKs
seem to happen faster and secondly, the TCP window size seems to get larger
than the hard limit - I see packets up to ~64K.  Possibly the NIC is stitching
them back together.

David