[OpenAFS] Re: jumbo

Simon Wilkinson sxw@inf.ed.ac.uk
Mon, 28 Feb 2011 18:48:39 +0000

On 28 Feb 2011, at 18:32, Harald Barth <haba@kth.se> wrote:

>> (fragmented packets are actually more efficient in our RX
>> implementation than having to process more single packets)
> Are there actual numbers from this century to validate this
> theory? If yes, I would be interrested to see those.

I don't have any numbers in front of me right now, but I did verify this at t=
he end of last year. On a network that passes them, raw rx connections using=
 jumbograms are almost 4 times faster than those without. You can verify thi=
s yourself using the rxperf tool.

When you are using RX to carry AFS RPCs, some of this performance difference=
 is hidden, but jumbograms remain substantially faster. There are a few reas=
ons for this - producing packets for transmission is expensive, as is handli=
ng ACKs for those packets. Jumbograms reduces both of these costs.

I've been doing work for YFS on improving RX performance which is starting t=
o narrow this gap.