samba - Jul 1998 - Slowdown when copying large files (PR#8624)

tl@fairplay.no wrote:
> Hmm, I missed the original post, but this sounds like it could be the bug
> I reported 15 days ago to samba-bugs (PR#8316) and still haven't got
any
> responses to.
> 
Sorry, I'm still integrating incoming patches.
> If it's the same bug it's mainly the write performance that's
bad, and in
> reality the performance is worse than Win95/WinNT for all files, it's
just
> get worse for large files (exactly how much depends on the OS).
> 
> To check if it's the same problem turn up the debug-level to 3, and
copy a
> moderately large file to the Samba server from the local disk of a Win98
> workstation. If you get *lots* of SMBwrite and SMBflush in the log-file
> you have the same problem.
> 
> A "good" log created from WinNT features one (or sometimes two)
SMBwrite
> per ~63k and zero SMBflush. A log from a Win98 machine features one or two
> SMBwrite per 1.5k (!) and one SMBflush per 1.5k!!!
> 
> Anything above ONE SMBflush per file (at the end) or one SMBwrite per 10k
> or so indicates a serious problem.
> 
> A stop-gap measurement to get *usable* performance is recompile with
> 'NO_FSYNC' defined (or the fsync() call in server.c:sync_file()
> commented).
> 
This is what I'm going to do for 1.9.18p9, but I will
parameterize it.
> This removes the silly 'wait until everything is on disk before
continuing
> at every 1.5k', but doesn't fix the small block-size used. The
small
> block-size used still limits the performance, but it's actually usable
for
> larges files.
> 
> If you can confirm that this is what happens to you too it might be time
> to send it to 'samba-urgent', as it's likely to affect LOTS of
people as
> more and more people start using Win98.
Well the problem is the small block size used is *client*
defined - not server. My guess is that the Win98 explorer
is optimised for Web performance, (ie. doing MSS size 
reads/writes) and as the same binary is used for both SMB
file browsing and Web browsing the silly block size has been
used throughout.

Does anyone know if there is a Win98 registry setting to
set the SMB read/write size ?

Jeremy Allison,
Samba Team.
-- 
--------------------------------------------------------
Buying an operating system without source is like buying
a self-assembly Space Shuttle with no instructions.
--------------------------------------------------------

On Mon, 27 Jul 1998, Jeremy Allison wrote:
> tl@fairplay.no wrote:
> > Hmm, I missed the original post, but this sounds like it could be the
bug
> > I reported 15 days ago to samba-bugs (PR#8316) and still haven't
got any
> > responses to.
> 
> Sorry, I'm still integrating incoming patches.
No harm done, it's just that after 15 days without a reply for what I
suspect may be a rather serious and growing problem(growing Win98
userbase) I was beginning to think that perhaps it should be forwarded to
samba-urgen just in case...

> Well the problem is the small block size used is *client*
> defined - not server. My guess is that the Win98 explorer
> is optimised for Web performance, (ie. doing MSS size 
> reads/writes) and as the same binary is used for both SMB
> file browsing and Web browsing the silly block size has been
> used throughout.
The interresting thing is that it doesn't appear to use small block
against WinNT4 servers...

Since I can't get SAMBA level 3 logs from the WinNT server I have to rely
on tcpdump's, and they are a whole lot harder to understand, but it does
look like Win98 is using much larger blocks against WinNT than against
Samba servers!

As an example, a copy of a 259261 bytes file generates:
WinNT4SP3: 1 SMBwrite(REQUEST), 1 SMBwrite(REPLY), 
	   10 SMBwritebraw(REQUEST), 10 SMBwritebraw(REPLY)

Win98:    119 SMBwrite(REQUEST), 28 SMBwrite(REPLY)

Due to other reasons these copies was done by different Win98 boxes, but
that shouldn't do any difference (if necessary I could do it with the same
machine, but not just now).

The Samba server was a custom-compiled 1.9.18p8 on RedHat 5.1, and 'write
raw' / 'read raw' is both in their default setting (enabled
according to
documentation and source).


This isn't actually especially surprising, I would have been a lot more
surprised if Win98 *didn't* use large blocks against WinNT, inferior 
performance against WinNT servers compared to Win95 would have been a
major show-stopper... Even Microsoft should have noticed THAT! And if they
had a lot of people would have yelled at them by now...

I didn't actually check it then, it was simply to unlikely in my opinion.


This was why I assumed that it should be possible to get Win98 to use
decent sized blocks against SAMBA, and that it was likely that it was
something in the handshakes that differed.

> Does anyone know if there is a Win98 registry setting to
> set the SMB read/write size ?
How does WinNT get it to use large block sizes??


The tcpdump logfiles are rather large, so they are available on request.

I also made an tcpdump RPM for RH5.1-intel, with the latest
tcpdump/libpcap (3.4 and 0.4, both final!), SMB support and various other
patches. I'm trying to upload it and the SRPM to RedHat's contrib area,
but ftp.redhat.com is a bit busy at the moment...
(Available by email if someone wants it before it's in the regular RH
contrib/hurricane directory).