[Samba] Re: Marginal write performance & pauses in outgoing transfers

[Nicolas Gieczewski]

> If necessary, use ethereal next time and store the 
> fully detailed display in a print file. After  compression it
> is much smaller. Your observation is correct, there are
> long pauses between payloads, but I don't see what it's
> waiting for.

I'll try that when I get some more time and post the results here,
but I doubt we'll find anything fixable (yep, I'm the pessimistic kind).
Actually, I'm starting to doubt Win2k SP4 regarding this issue. I'm not
certain it's what caused the problem, but I installed it like last week
and I've read many posts on different places about all the problems it's
caused to thousands of people (much more serious than mine in many
cases). It's a shame I hadn't used Samba to transfer large files for
weeks, so now I'm not sure if things broke before or after I installed
SP4, but there's a possibility it has something to do with this. The
worst thing I could have done is delete the SP3 backup files one day
before I found out about this problem. *sniffles*

> It's best to suppress the pop-up info in the file browser (clear the box
> at Options->Folder options->View->Popup-info) even when you don't
> have problems.

Pop-up info is alright, and I use it often actually. It's the status bar info
what causes the slowdown while I have this problem and scroll through the
files with the arrow keys (I use my keyboard extensively anywhere I can
avoid using the mouse :)  When I didn't have this problem, though, that
wasn't an issue. It was nearly as fast as it is for local files. It doesn't
really bother me that much anyway.

> It's a clean little smb.conf. I don't believe those
> 8k buffer specifications in socket options do much 
> good, but they're probably irrelevant. Therefore throw
> them out. On the other hand I wouldn't exclude 
> "SO_KEEPALIVE" from socket options, so add it. I don't 
> know what "dead time" does, but if you're sure you do, 
> it's another toggle you can use in tests.

Been there, done that. I've fiddled with exactly all the settings
you mentioned and tried all possible combinations (I even made up
a table to make sure I didn't skip any). Absolutely none of them
had an effect on the traffic pauses, it was just the buffer settings
which caused an impact on the size of the data bursts in between
the pauses but, as I had previously said, the pauses were never
affected. *bangs head against keyboard*

> I don't see much use for "interfaces" and "bind interfaces only".
> If you only have one NIC, you almost certainly don't  need them.

I have two NICs, and as paranoid as I am, I always restrict all
my daemons only to that/those interface(s) where I need them
to be listening. Even syslogd is restricted to listening only on the
loopback interface and on one of my NICs where my Cisco DSL
modem is connected, which logs to syslog. However, I think I
even commented out that stuff while testing as a last resort.
I've tried literally everything.

> But who's the browser master in your workgroup? If you
> leave it undefined, then any of your clients will claim it and
> there'll be some broadcast storms from  time to time when
> they're settling the issue who shall win the election. You
> should add the following options in [global]:
> 
>    wins support = Yes
>    preferred master = Yes
>    domain master = True
> 
> and tell your clients to never start elections. That's
> "MaintainServerList"="No" under HKLM key
> SYSTEM\CurrentControlSet\Services\Browser\Parameters
> in the Win2K registry.

Still no improvement. I wasn't even doing name-based browsing,
by the way, as I always access my machines by IP address. I know
they're harder to type than names, but hey, I once had a dream
where I got a new hosting service and I remembered the IP address
they had assigned to me even after waking up. I think I need a
break :)  Anyway, I even made Samba listen on port 445 during my
tests so as to get rid of any NetBIOS-specific overhead, with no
success.

> I wish you good luck.

Thanks a bunch, I need it ;)

Cheers,

Nicolas Gieczewski
Nix Software Solutions
http://www.nixsoftware.com/