The CIFSbench package

Luke Kenneth Casson Leighton lkcl at
Wed Mar 10 16:58:45 GMT 1999

On Thu, 11 Mar 1999, Christopher R. Hertel wrote:

> Interesting discussion, re CIFSBench.  I have a question...
> Given the complexities and permutations of SMB, how possible is it (or 
> how easy is it) to tune a client for a particular server?

ms trying to tune their SMB server to 64k block reads, due to stupid app
writers only having 64k buffers:

char buf[65536];
while (!eof(fpos))
	read(fpos, buf, sizeof(buf));

so ms has to optimise to deal with this, _plus_ there is a relatively
expensive overhead per SMB, so they have to bear this in mind too.

andrew tridgell has noticed that if you hammer out several 8k SMBreadX
requests into the samba TCP buffer (see smbclient source) you can force an
SMB server to respond with SMB responses back-to-back on the wire.

with "server-side QOS" limits on a per-client basis you can get max
performance _and_ ensure that one client doesn't do in the network.


More information about the samba-technical mailing list