Samba write throughput

[Jeremy Allison]

On Fri, Jan 28, 2011 at 09:09:20AM +0800, Jacky Lam wrote:
> On 1/27/2011 2:33 PM, Volker Lendecke wrote:
> >On Thu, Jan 27, 2011 at 09:24:57AM +0800, Jacky Lam wrote:
> >>On 1/27/2011 12:38 AM, Jeremy Allison wrote:
> >>>On Wed, Jan 26, 2011 at 05:51:13PM +0800, Jacky Lam wrote:
> >>>>Dear all,
> >>>>
> >>>>     I setup my Samba server on my machine and get 300Mbps for
> >>>>reading from server, but only 170Mbps for writing. I found that
> >>>>reading is use a more efficient way to do with sendfile(). But
> >>>>writing just use normal file IO and cause too many memory copy. Is
> >>>>there any way to trigger (from client) or force the server to use
> >>>>recvfile() instead?
> >>>http://wiki.samba.org/index.php/Linux_Performance
> >>>
> >>Thanks. It is new to me, but it doesn't answer why recvfile/splice
> >>implementation has not chance to be used when my XP are writing to
> >>it. Any other idea?
> >Is your system really so CPU-bound? What kind of CPU do you
> >have? Atom? ARM? Small MIPS?
> >
> >Volker
> >
> Yes, my system is seriously CPU-bound. When I copy a file from PC to
> my MIPS's Samba server, I found that almost 30% CPU time is spending
> on copy_to_user and copy_from_user. But I googled that although
> recent Linux does have recvfile(), but Samba has used splice() to
> reduce the memory copy. That's why I curious.

In tests I've done with an OEM, using splice() on write makes
the problem *worse*, unless the kernel has been patched to do
direct socket to buffer-cache zero copy.

Jeremy.