win98 vs win95 with TCP/IP

Marcus Graf mg at graf.weinheim.de
Mon Oct 12 20:00:10 GMT 1998


> We crave stability.  But Linux+Samba is just much slower as a file 
> server.
> 
> 	It's normally quite fast: are you running over a
> 	dial-up network, requiring you to adust MTU's for
> 	speed???
> 
> 	If not, try ``socket options = TCP_NODELAY'', which
> 	adapts server TCPs to a MS feature (:-)) and produces considerable
> 	speed increases...

There is another point I've reported to samba-bugs a few days ago: 

We've experienced a great *loss* in performance when migrating an 
Access application (MDB file > 16 MB) from a slow NT-Server 
(486DX-66, NT 3.51 server) to a muxh faster Samba server (P166, 64 
MB, Linux 2.0.35, Samba 1.9.18p10.) The workstations are connected 
via a 100 MBit fast ethernet.

With one workstation using the database it ran normal. With two 
warkstations it slowed down. With four workstations the application 
was unusable (> 2 minutes for one query)

Playing with smb.conf we've tracked ist down to be a locking problem. 
Fith 'fake oplocks = yes' it ran at normal speed. (Of cause we've got 
database consistency problems.)

A friend of me looked over the Samba code and told me to reduce the 
LOCK_RETRY_TIMEOUT in local.h from 100 ms to 5 or 10 ms. After that 
the application ran at full speed again. The server load is still 
normal. Maybe 100 ms for the retry timeout are to much wasted time 
for a fast network...

I'm not sure if this was the right way to solve the problem. But it 
works :-)

[We've also set -O6 -m486 -fomit-frame-pointer in the Makefile. Maybe 
it makes Samba a little bis faster :-)]

Ciao
  Marcus

* We build our computers the way we build our cities -- over 
* time, without a plan, on top of ruins. 
* (Ellen Ullman, "The dumbing-down of programming")


More information about the samba mailing list