[cvsnt] Poor performance on Windows 2003 server

[Arthur Barrett]

Mark,

OK you've ruled out the most obvious things including anti-virus.  

There was an exhaustive thread here about a year ago with an almost
identical problem - it ended up being the SCSI RAID drivers.  A search
of the newsgroup for "RAID" should find it including my extensive
comments about how to isolate the problem.

There are so many variables in a CVSNT installation what you need to do
is start isolating things.

Firstly isolate the network.

Do the same checkout on the server using :local: protocol then also on
the server using :pserver: protocol.  If this is slow then the problem
is on the server someplace.  Try moving the repository to a USB stick
(5GB should fit) or a RAM disk and re-test to isolate the disks.

If the speed is fine on the server then you need to start looking at
different clients.  Again there was a thread some time ago where one
switch or network card (can't remember which) was causing the problem.

Finally if you trace (cvs -ttt co module) then each action has a
timestamp - it may be possible to "see" some specific points in the
trace which are taking too long to execute.

Regards,



Arthur Barrett
 

-----Original Message-----
From: cvsnt-bounces at cvsnt.org [mailto:cvsnt-bounces at cvsnt.org] On Behalf
Of Mark Donnelly
Sent: Thursday, 19 April 2007 12:59 AM
To: cvsnt at cvsnt.org
Subject: [cvsnt] Poor performance on Windows 2003 server


Hello all:

I've been tasked by my company to convert an existing linux-based CVS
1.11.1p1 repository to a windows 2003-based CVSNT repository (CVSNT
2.5.03 Build 2382).  However, I've run into a problem where a checkout
is taking two and a half times as long with CVSNT as it is with CVS.

Our repository size is a bit over 5GB, with about 175 modules in it -
though probably only about half of the modules are active.  The old
server is a RedHat 7.1 (kernel 2.4.2) system on a first-generation
Compaq ML370 with dual P-III 933MHz processors, 1.1GB of RAM, and CVS on
a dedicated mirrored pair of 10k-RPM SCSI drives.  The new server is a
Windows 2003 SP1 system on an IBM x335 with dual Xeon 3.06GHz
processors, 4GB of RAM, and the entire system on a pair of mirrored
15k-RPM Ultra-320 SCSI drives.

If I run the command "tar -cf - [large_module] > /dev/null" and "tar -cf
- [large_module] > nul" on the Linux and Windows system, respectively, I
find that the command runs more quickly on the Windows machine, as I
would expect.

The CVS conversion that I've done is based off the mail list archive
message at http://www.cvsnt.org/pipermail/cvsnt/2003-June/007066.html -
I TARred up the CVS repository, copied that TAR over to the Windows
system, and extracted everything but the CVSROOT into the CVSNT
repository.  I made sure to remove the LF -> CR/LF translation within
WinZip.

The system is using pserver authentication, at least until I get this
performance problem settled.  There are a couple of perl scripts that
run, but they're called via the CVSROOT\verifymsg and CVSROOT\commitinfo
files, which should both be triggered only upon a commit.

The active modules also have a large number of tags associated with
them, as our company runs nightly builds.  Also, the checkout and active
development happens on a branch, as we've abandoned HEAD/trunk for
historical reasons.

I appreciate any help that people can offer.  If you need any more
information in order to make a suggestion, then I'll be happy to supply
it.

Thanks,
--Mark Donnelly
_______________________________________________
cvsnt mailing list
cvsnt at cvsnt.org http://www.cvsnt.org/cgi-bin/mailman/listinfo/cvsnt