[linux-cifs-client] Re: cifs_readpages improvements

[Steve French]

What about over mounts over localhost interface?

On 6/23/07, Kirill Kuvaldin <kirill.kuvaldin at gmail.com> wrote:
> Hello Steve,
>
> I've started working on the summer project a while ago. I'm currently
> investigating CIFS read performance. Recently I've done a patch for
> adding asynchronous read support for the cifs_readpages operation.
> I'll post the patch for review to the mailing list soon.
>
> I'm currently testing changes in my setup, where the samba server
> (3.0.24) runs on a linux box and the cifs client runs in a qemu
> virtual machine running on the same box, the networking is emulated by
> the tun/tap driver (that gives about 70-80Mbit/sec throughput
> according to the iperf benchmark).
> I don't have another linux machine yet to conduct a decent performance
> analysis (e.g. over GigE).
>
> With these asynchronous changes, CIFS is generally doing something
> better (even in my setup), see the graphs below. I've been running
> iozone using the following command:
> iozone -azc -U /cifs -f /cifs/iozone_data -s 32m -i 0 -i 1
>
> rsize=16k:
> http://epsmu.com/~kuvkir/cifs/benchmarks/cifs_read_rsize_16k.png
>
> rsize=32k:
> http://epsmu.com/~kuvkir/cifs/benchmarks/cifs_read_rsize_32k.png
>
> rsize=64k:
> http://epsmu.com/~kuvkir/cifs/benchmarks/cifs_read_rsize_64k.png
>
> rsize=127k:
> http://epsmu.com/~kuvkir/cifs/benchmarks/cifs_read_rsize_127k.png
>
> The performance difference  between the sync and async version is not
> significant enough though. I anticipate to see more performance gains
> in the gigabit ethernet setup.
>
>
> Kirill
>


-- 
Thanks,

Steve