How to debug 100x variance in writing the same file?
rsharpe at richardsharpe.com
Fri Jan 21 19:27:53 GMT 2005
On Thu, 20 Jan 2005, John Gerth wrote:
> I'm writing out of some desperation and not so much looking for a direct
> answer as some clues about what directions to look in.
> The problem is occuring on a 1.3 GHz Apple G4 Xserve 10.3.6 which is Samba 3.0.5
> I've been working to reduce the number of variables and so all clients
> and the servers are now on the same gigabit VLAN (Cisco switches). I have
> mapped drives from the Xserve and from a 2.8GHz Win2k3 server
> The share on the Xserve is in /tmp on a local disk. At Apple's advice:
> socket options = TCP_NODELAY SO_RCVBUF=64240
> The client is 3.2GHz WinXPpro also on gig-e and my test cases are of the form:
> timethis xcopy c:\temp\nnnM f:nnnM
> where nnn is the size of the file in megabytes, 10,20,40,100, etc.
> and chose at time
> Just to get a baseline I also did "scp" of the files from a Linux gig-e
> box to the Xserve and it runs at a very nice 50MB/s as do the xcopys
> of 100M to the Win2K3 box. This leads me to believe that I don't have
> any network problems and that the Xserve is fully capable of absorbing
> the data at network speeds.
> However, writing to the Xserve's Samba is just all over the map.
> xcopys of even 10M files, vary from .2 to 55 seconds (files of
> other sizes up to 100M show similar ratios). I even went
> so far as to symlink a target filename to /dev/null to try and
> remove any vagaries with respect to actually writing to disk,
> but the wild variations continue (usually very fast or very slow).
OK, as someone has already pointed out, the request-response nature of the
SMB protocol can lead to some problems.
I would be interested in seeing some traces of the good times and the bad
If you have fast-retransmit and fast-recovery switched off, for example,
things can go bad if the network loses packets, and the number of lost
packets will determine the amount that things suck by.
However, stream protocols like FTP and SCP would tend to recover from
events like that.
Richard Sharpe, rsharpe[at]richardsharpe.com, rsharpe[at]samba.org,
More information about the samba-technical