How Samba let us down
Mathew McKernan
mathewmckernan at optushome.com.au
Wed Oct 23 09:58:01 GMT 2002
Hi Chris,
I can only comment on personal experience here, what I have commented on may
be wrong in some cases. The Samba developers may be able to comment on my
ideas.
By the look of it, the reason why it is so slow is the fact that you may not
be running a WINS Server. We had this problem with NT boxes, yes Windows
Servers. We installed a Windows NT Server to be our WINS server, it
increased the speed of the LAN dramatically. We now run the WINS Server on a
Linux box running Samba.
We have a home drive server which serves about 1800 users with 400 logged on
at one time drawing about 30MBps out of it server. This box is a Pentium 4,
512MB RAM. 400GB RAID server running Linux and Samba. We originally ran a
similar setup on Windows NT, but found it was terrible slow once connections
got over 200 users. This box is providing access using ACLs, like in your
setup.
My suggestion:
Install a WINS Server (simple 400MHz box even) running Linux, and if you
like run an internal DNS too which is syncronised to the WINS database using
the "wins hook" option in smb.conf. Point all your devices' WINS addresses
to this new WINS server. You will notice a dramatic improvement in
performance.
Thanks
Mathew
----- Original Message -----
From: "Chris de Vidal" <cdevidal at yahoo.com>
To: <samba at lists.samba.org>
Cc: <samba-technical at lists.samba.org>
Sent: Wednesday, October 23, 2002 4:13 PM
Subject: How Samba let us down
> Before you read this, I want to state (for reasons
> listed below) that I don't expect an answer (advice is
> welcomed, but please read this email carefully before
> answering). I'm sharing this with the community with
> the hope that better software results from our sad
> experience...
>
> BACKGROUND
>
> I've been using NT for 4 years, Netware and Linux for
> 3 years, and Samba for almost 2. I work in the IT
> department of a medium-sized unit of a global
> advertising company. We have a Netware and NT
> environment with a bit of Linux.
>
> We installed a 280GB IDE Samba archive server (rare
> usage) and a 15GB SCSI Mac/Samba file server (medium
> usage). We also use Samba for more menial tasks like
> smbmounts and file transfers. We thought we were
> comfortable with Samba. We knew we were comfortable
> with other types of file servers.
>
> OUR SETUP
>
> Going from my tired memory:
> Athlon MP 1.8GHz (mem=nopentium)
> 2GB ECC SDRAM
> Tyan S2460(I think?)
> Antec 450W PS
> Lots of cooling
> 5 IBM DeskStar 120GB drives with 8MB caches in RAID 5
> 3ware 7580(I think?) 8-port hardware RAID
> 3ware hot-swappable drive cages
> Intel e1000 Gigabit NIC, full duplex, 1000MBit,
> autonegotiation off
> 3com Gigabit switch, autonegotiation off
> RedHat 7.3
> Kernel 2.4.19 with ACL support
> ext3 with ACL support
> Samba 2.2.5 with ACL support installed from a
> recompiled SRPM from the samba.org FTP site.
> Winbind
> NO nfs daemon (I hear it's buggy w/ ACLs)
>
> We have a variety of clients, from DOS and OS/2 to
> Windows (9x-2000) and Linux. The server acts as a
> print spooling area (the actual queues are on an NT
> server) and scratch area for database programmers to
> manipulate their flat database files. As far as I
> know, these files are not commonly accessed by more
> than one user at a time.
>
> THE PROBLEM
>
> For the past year, our heaviest-used Netware server
> has been under more and more stress.. filling up,
> running out of licenses, slowing down, etc.
> Preliminary tests using Samba on a fast Linux box
> showed anywhere from 70% to 1000% speed improvements,
> depending on the task. The decision was made to
> switch it to Linux; the whole company is migrating
> away from Netware and we (as a unit, not speaking for
> the company) don't want to be completely trapped into
> Windows if we can help it.
>
> The new hardware arrived and more preliminary tests
> indicated all looked good. We were set to switch last
> Saturday night. We turned off logins to the Netware
> box, backed it up, restored it to the new Linux box,
> set permissions, then made sure the various computers
> in the building could log in.
>
> Yesterday, our first day, was rough. For most of the
> day we fought random slow browsing with no
> explanation. Clients would appear to lock up for
> several seconds. We found some misconfigurations in
> smb.conf but the problems reappeared. No errors were
> seen in any machines' logs on debug level 2. I
> trimmed the smb.conf to a minimal number of options
> and that seemed to help with the slowness. Today,
> however, the problem reappeared a few times with no
> errors in the logs that we could see.
>
> The printers were missing some of the records sent to
> them to print, something that had never happened with
> Netware. Every time the missing records were
> different. Occasionally, it would work right.
> Oplocks (kernel, level I and II) were left to defaults
> (turned on).
>
> THE OUTCOME
>
> Sadly, tonight we are installing a Windows NT server.
> Installing a brand new server is actually cheaper for
> us than the 8 or so hours of downtime to back up the
> server, install NT on it, and restore the data to it.
> We don't want to revert to Netware because so many
> clients have been reconfigured to log on only to the
> domain (DOS, OS/2, etc.) and that would require many
> more hours reversing those changes. Also, some files
> have been added since leaving Netware. We also
> decided to proceed to use NT because is more proven in
> this capacity.
>
> CONCLUSION
>
> To be fair, the problems could be related to some
> misconfiguration. I have pasted the smb.conf below.
>
> I fear it might just be an oplock problem, but it is
> not clear what would result if more than one user
> happened to try to write to a file with them disabled.
> Every advice we found said to leave them on to
> prevent corruption and to improve performance. We ran
> out of time to test it, and feared what failure would
> bring. Running this:
> grep -r -B5 -A5 oplock /var/log/samba/ | grep -B5 -A5
> error
> produced only 5 of these errors
> oplock_break: receive_smb error (Connection reset by
> peer)
> from the same DOS machine from 2 days worth of all
> machines' logs running at debuglevel 1 (some at level
> 2). I don't know if that is a good indicator of an
> oplock problem. I can do other greps on request.
>
> Unfortunately, we can't test out your suggestions in
> production, and our off-production testing apparently
> can't stress it well enough. So please just take this
> email as input - I'm not looking for answers here,
> though advice is appreciated.
>
> The problem could also have been environment or
> hardware. We should know soon, as we are going to
> reinstall the original Samba server with NT, and the
> problems should reappear if hardware or environment.
> If we do find that to be true, I will certainly reveal
> our findings to this mailing list.
>
> And perhaps the problem was with ACLs. We couldn't
> turn them off in production to test that theory.
>
> It is likely that we will try Samba in this capacity
> again in the future with a more mature version.
>
> Thanks for listening,
> /dev/idal
>
>
> [global]
> server string =
> workgroup = <our domain>
> password server = <our PDC>
> security = domain
> encrypt passwords = yes
> smb passwd file =
> /etc/samba/smbpasswd
> veto files = /lost+found/
> winbind uid = 10000-20000
> winbind gid = 10000-20000
> winbind separator = +
> create mask = 660
> force create mode = 660
> directory mask = 0770
> force directory mode = 0770
> log file =
> /var/log/samba/%m.log
> debuglevel = 2
>
> [print]
> path = /share/print
> writeable = yes
>
> __________________________________________________
> Do you Yahoo!?
> Y! Web Hosting - Let the expert host your web site
> http://webhosting.yahoo.com/
More information about the samba-technical
mailing list