RE: [Xen-users] Release 0.9.10 of GPL PV Drivers for Windows

["James Harper" <james.harper@xxxxxxxxxxxxxxxx>]

> for 0.9.10:
> For a tcp test, 'iperf-1.7.0 -c dom0-name -t 60 -r':
> 
> domu->dom0: 37.5 Mb/s (better)
> dom0->domu: 101 Mb/s (wow!)
> 
> While I would love it if the two directions were more equal, it's
> definitely
> better for the download speeds to be better than the upload if I can
only
> pick one ;-)
> 
> (Playing with all combinations of csum and lso off and on resulted in
> minor differences.)

Off the top of my head, here are a list of reasons why performance might
not be all that it could be:

1. Linux supports Large Send and Large Receive (for when a Large Send
packet doesn't get routed out of physical hardware). Windows can do
Large Send but cannot handle Large Receive, so I have to fake it by
splitting up the large packet into smaller ones and re-calculating
checksums.

2. Windows seems to only ever give me one packet in its 'SendPackets'
routine, which will obviously result in higher per-packet overheads than
if it could batch them (eg we potentially have to 'notify' dom0 of a new
packet each time instead of having to batch them). Large Send may lessen
that problem a bit though.

3. We don't try and do any interrupt moderation or anything.

4. We try and be a bit conservative with resources (eg packet buffers
etc).

So there is a chance of getting a bit more performance out of it yet...

James



_______________________________________________
Xen-users mailing list
Xen-users@xxxxxxxxxxxxxxxxxxx
http://lists.xensource.com/xen-users