Ugly slowdown in Windows file sharing after SP2 - definitely an SP2 issue

[Cactus]

Hey.

Just for the record, so we can all share out info. I have two
state-of-the-art Intel/ XP Pro desktops, connected thru a GigE switch.

I frequently copy large numbers of 100kB - 500kB sized files between
the two machines.

With XP SP1, performance was about 3.5 MB/sec sustained (about 10% of
the speed attainable with larger files, because of Windows file
sharing overhead). This number is pretty normal.

After installing XP SP2, the same operations rarely exceed 500 kB/sec,
or about 7X slower.

This is completely reproducible. I checked all the obvious things,
such as NIC driver setttings. All looks fine.

When I removed SP2, the problem went away completely. So this is 100%
certainly an SP2 bug.

Have fun out there.

Cactus.

 

[Hans-Georg Michna]

With XP SP1, performance was about 3.5 MB/sec sustained (about 10% of
the speed attainable with larger files, because of Windows file
sharing overhead). This number is pretty normal.

Cactus,

unless your hard disks are rather slow, this is not normal at
all. Even older hard disks should achieve 6 MB/s. The network
itself should reach a persistent 11 MB/s, at the very least 10
MB/s. Anything below that is definitely a sign of a problem.

So one suspicion is that one of the computers had a problem all
along, and Service Pack 2 only made it more obvious.

Could you check the kernel processor times while copying large
files? One suspicion is that one of the network adapters
overuses the processor. You can check this in Task Manager after
enabling the kernel time display.

Hans-Georg

 

[Cactus]

Cactus,

unless your hard disks are rather slow, this is not normal at
all. Even older hard disks should achieve 6 MB/s. The network
itself should reach a persistent 11 MB/s, at the very least 10
MB/s. Anything below that is definitely a sign of a problem.

So one suspicion is that one of the computers had a problem all
along, and Service Pack 2 only made it more obvious.

Could you check the kernel processor times while copying large
files? One suspicion is that one of the network adapters
overuses the processor. You can check this in Task Manager after
enabling the kernel time display.

Hans-Georg

I completely disagree with your performance estimates. My previous
speeds are normal (for many smallish files) based on my experience at
work across many XP systems. One can certainly argue that it
represents crappy performance, but hey that's WIndows. The further
slowdown is quite definitely a bug in SP2.

 

[Cactus]

unless your hard disks are rather slow, this is not normal at
all. Even older hard disks should achieve 6 MB/s. The network
itself should reach a persistent 11 MB/s, at the very least 10
MB/s. Anything below that is definitely a sign of a problem.

Oh and by the way, the application I am using accesses the network
drive files in parallel, not sequentially. So that would easily
contribute a factor of 2 slowdown. Say from 7 MB/s to 3.5 MB/s.

The problem I saw is DEFINITELY, 100% a bug in WinXP SP2. I realize
you are a Microsoft enthusiast, but please try to be open-minded.

 

[Hans-Georg Michna]

On Sun, 29 Aug 2004 07:39:28 +0200, Hans-Georg Michna

I completely disagree with your performance estimates. My previous
speeds are normal (for many smallish files) based on my experience at
work across many XP systems. One can certainly argue that it
represents crappy performance, but hey that's WIndows. The further
slowdown is quite definitely a bug in SP2.

Oh and by the way, the application I am using accesses the network
drive files in parallel, not sequentially. So that would easily
contribute a factor of 2 slowdown. Say from 7 MB/s to 3.5 MB/s.

The problem I saw is DEFINITELY, 100% a bug in WinXP SP2. I realize
you are a Microsoft enthusiast, but please try to be open-minded.

Cactus,

when you copy smaller files, that's something different. To test
network performance, it is better to copy very big files. Even
better is a test program that generates data on the fly, so the
disk performance does not come into play.

I cannot rule out a bug in Service Pack 2, I only wanted to
mention that you should get 11 MB/s when transferring data over
a 100 Mbit/s network, and I have actually tested that. You get
it when you use newer, fast hard disks or RAID arrays. My first,
now no longer valid, thought was, if you got only 3.5 MB/s,
there may have been another problem already before the SP2
installation.

Regardless whether or not Service Pack 2 is the cause or only
the trigger, we should try to pinpoint the cause and repair it.
I have currently not many ideas and hope that others will chime
in and give us some good proposals on how to proceed.

Meanwhile could you simply copy a large file across the network
and measure the time? Ideal would be a file of 100 MB, so we see
at least some measurable 10 seconds copy time. I would like to
take disk performance out of the consideration and check whether
the problem is really the network. I have seen too many cases in
which the problem was something else, fragmented files, many
small files, disk controller problems, things like that.

Hans-Georg

 

[Cactus]

I have seen too many cases in
which the problem was something else, fragmented files, many
small files, disk controller problems, things like that.

Yes I understand. By the way, I remember you from the old Canopus
forum on Compuserve, maybe 12 years ago.

As I stated in my original mesage, I ran my application with SP2
installed, got ~500 kB/sec very consistently. Then I uninstalled SP2.
Then I re-tried the exact same things and got ~3.5 MB/sec very
consistently.

SP2 is off my system permanently now, so I'm sorry I can't try the
large-file thing you suggested. Actually I am guessing that this is
specifically a small-file issue, i.e. it's related to the NetBios
subsystem processing of remote file open operations.

On the bright side, SP2 did fix a very annoying problem of cumulative
slowdowns in the explorer shell. For example, with SP1, after several
days without a reboot, the context menu in windows explorer would take
many seconds to come up after clicking.

I read through the list of fixes in SP2, and found 4 which referred to
fixes for resource leaks (handles). I suspect that what I was seeing
was the XP kernel scanning through a linked list of millions of leaked
handles. So I will get the hot-fixes for those specific problems. If
they fix my problem, IMO I have no need for SP2 at all.

Cactus

 

[Cactus]

And by the way, my slowdown was NOT explicitly related to windows
firewall (as has been suggested elsewhere for LAN slowdowns with SP2).

Before I nuked SP2, I disabled the firewall and the firewall service,
and rebooted. No improvement in my problem.

However, I do suspect this problem will turn out to be a side-effect
of the firewall changes. Something got screwed up in NetBios.

 

[Hans-Georg Michna]

By the way, I remember you from the old Canopus
forum on Compuserve, maybe 12 years ago.

Ha, incredible! Greetings!

As to Service Pack 2, if you delay its installation, I believe
you will slowly, but surely miss future patches, so you probably
can't hold out for very long. So the hope is that the problem
will be fixed in a patch soon.

It would be interesting to find out whether anybody else has the
same problem. If nobody has it or very few have it, it won't get
fixed, but perhaps then there is a way to fix it individually,
if the cause can be found.

The long file test would still be interesting too, just to find
out whether it is indeed a problem of many short files.

Hans-Georg