Latency and Rsync Transfers

[Neal B]

Hi Everyone,

Thanks for your replies.  The autotuning in the 2.6 kernel seems to be much
better than manually setting the buffer sizes for the rsync client and
server.

Tuning the kernel allowed us to triple the speed.
http://fasterdata.es.net/TCP-tuning/linux.html

Neal



On Mon, Feb 1, 2010 at 6:59 PM, Jamie Lokier <jamie at shareable.org> wrote:

> Neal B wrote:
> >    Thanks for your reply.  I have been experimenting with the buffer
> >    settings and when specifying it actually causes the transfers to go
> >    slower.
> >    I am running an rsync server using xinet.d and an rsync client.  I
> >    have tried specifying the sockopts on just the client, server, and
> >    both.
>
> Did you set the buffers large enough?  For GigE, 75ms, if the link is
> all yours and that's really the bandwidth, you will need at least
> about 1.5 megabytes of buffer to keep the link busy.  Double it to 3
> megabytes for good measure (latency variance).
>
> On Linux, you have to write larger values to
> /proc/sys/net/core/wmem_max and rmem_max before the SNDBUF and RCVBUF
> options are effective.  Some other OSes have similar limits that need
> to be raised.
>
> In addition to SNDBUF and RCVBUF, some OSes dynamically increase the
> buffer size, and using SNDBUF and RCVBUF disables that, so they are
> not always beneficial.
>
> -- Jamie
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.samba.org/pipermail/rsync/attachments/20100202/64cf4374/attachment.html>