[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: tcp write performance testing

On 11 Jul 2012, at 21:36, Balraj Singh wrote:

> A quick update on tcp performance after the recent code refactoring:
> TCP read speeds have gone down from about 1 Gb/s between VMs on the
> same host to about 450 Mb/s.  It looks like this is because of a flow
> control / pkt dispatch problem.  TCP write seems to have a more
> serious problem that it is sending pkts outside the available window -
> these pkts are dropped as expected and then the stack spends most of
> the time re-xmiting and recovering from the losses.  It looks like the
> write calls succeed instead of failing or blocking when there is no
> window available.
> So if you are using it, expect TCP to be very fragile for now.  Also
> it looks like DHCP is not working and there is a small race condition
> in ARP.

Thanks for running those tests Balraj! I did the port from Bitstring to
Cstruct as a huge patch, and so I'm not surprised if I introduced small
changes in the way that writes are acked from the OS to Net interfaces.

I'll see if I can reproduce some of the problems on UNIX in Cosener's




Lists.xenproject.org is hosted with RackSpace, monitoring our
servers 24x7x365 and backed by RackSpace's Fanatical Support®.