[THIN] Re: OT : Windows Performance Measure Query
- From: Mark Lee <marklee15@xxxxxxxxx>
- To: thin@xxxxxxxxxxxxx
- Date: Wed, 7 May 2003 10:15:13 +0100 (BST)
Bernd, Tim etc. etc. thanks guy's - I guess we all know how (un)reliable
taskman and perfmon are for exact performance stat's but they are the only real
guide we have with regard to system usage - all other tools (incl. Apptrack I
just posted) relies on the underlying sub-system to get the info out of the box
- there is no other way ! However, IMHO the best way to ID poor performing
servers is the users themselves - they kick and scream big time if you have a
server issue. True we monitor these stat's but for management - they are the
one's wanting the comfort blanket not the admins ! Mark
Bernd Harzog <Bernd.Harzog@xxxxxxxxxxx> wrote:
Tim,
Thanks for the insight. It is also my opinion that the issue lies not
just with the CPU metrics in task manager, but the very same ones in
PerfMon as well. They are both created by the same underlying process,
and are read from the same registry counters. The only difference is
that Task Manager accumulates some of them for you (like cumulative Page
Faults and cumulative CPU time since the process launched).
Bernd Harzog
CEO
RTO Software
678-455-5506 x701
bernd@xxxxxxxxxxx
www.rtosoft.com
-----Original Message-----
From: Tim Mangan [mailto:tmangan@xxxxxxxxxxxx]=20
Sent: Tuesday, May 06, 2003 8:22 PM
To: thin@xxxxxxxxxxxxx
Subject: [THIN] Re: OT : Windows Performance Measure Query
As long as we are bashing the task manager...
Not only does it truncate values, but it misses a bunch. It's not so
much
"did something take less than 10ms", but what happens to be running when
the
check is made. That task is active in the CPU is credited with using
the
entire previous period. In the long run it is an OK approximation, but
it
can be lead astray by conditions. One big one is interrupts. An easy
example. Let's say you have a some really bad software that spends
ungodly
time in interrupts (I never mentioned Lexmark by name, did I?). These
interrupts may account for an overall 5% cpu load at times. But that
processing gets credited to whatever process is active when the OS
checks
who is running. If the system is 50% idle, then 50% of the interrupts
get
credited to the idle task and not counted at all. (In reality more than
50%
of the interrupts end up there due to the nature of the interrupts, but
we're nitpicking). Your servers are busier than the task manager says.
If
you are still reading this and are interested, we have a white paper
about
the task manager under-reporting the white papers section of our web
site
(www.tmurgent.com).
tim
-----Original Message-----
From: thin-bounce@xxxxxxxxxxxxx [mailto:thin-bounce@xxxxxxxxxxxxx]On
Behalf Of Mark Lee
Sent: Tuesday, May 06, 2003 6:25 PM
To: thin@xxxxxxxxxxxxx
Subject: [THIN] Re: OT : Windows Performance Measure Query
Bernd, thanks for that useful snippet of info - must remember this for
future reference. Also, just to add - task manager truncates CPU usage
(does not perform a rounding operation) so 10 processes running at 0.9%
CPU
actually show as 0% CPU in total as opposed to 9% CPU in total !!! Go
figure, I can remember when CALC.EXE in Win3.1 couldn't count correctly
-
thought they would have learnt their lesson by now :-( Cheers Mark ps.
Plug away, TScale does increase server CPU utilisation but you also get
the
performance gain Bernd has mentioned - I speak as a VERY satisfied user.
TScale was installed nearly a year ago on our servers and it works that
well
we've nearly forgetten it's even there !!!!!
Bernd Harzog wrote:
Let me toss in a little kernel of knowledge that I have come across
recently. All of the CPU counters in all of the Windows server operating
systems are only refreshed on the 10 millisecond heartbeat (clock tick)
of the operating system. This means that anything that takes less than
10 milliseconds to occur is not counted at all in the CPU utilization
metric. To prove this to yourself, launch Task Manager on a Windows NT,
Windows 2000, Windows 2003 or Windows XP server or desktop computer.
Then add CPU Time (the cumulative CPU time that a process has used since
it was started), and Page Faults as columns. Now sort on CPU Time.
Notice that there will be several entries for which CPU Time is zero,
but for which a number of page faults have occurred. How can this be?
Answer, the page faults that have occurred took less than 10
milliseconds to occur and were not counted as having used any CPU.
This also explains why you can have a server that is reporting 50% CPU
utilization but which is "busy" or unresponsive to end users. It is 50%
busy doing things that are counted in the CPU utilization metric and 50%
busy doing things that are not. Bottom line is that CPU utilization is a
dangerously useless metric by which to judge how much work is being
done, or how busy the server is. It is also impossible to infer
performance from this metric since a lower CPU utilization metric does
not necessarily mean the server can or is offering better performance to
end users, since some much of what the CPU is doing in not reported in
this metric in the first place.
As the vendor of TScale, we have first-hand experience with this. Our
customers often find that after TScale has optimized their applications,
the server reports higher CPU utilization, but runs better, gives better
response time to users, and can support more concurrent users. The
explanation for this is that TScale is eliminating many of those
operations that take less than 10 milliseconds to occur and which were
never counted in the first place. This frees up the CPU to spend more
time doing work on behalf of users which gets counted, resulting in both
a higher reported CPU utilization number, and a server that provides
much better applications response time to end users.
Bernd Harzog
CEO
RTO Software
678-455-5506 x701
bernd@xxxxxxxxxxx
www.rtosoft.com
-----Original Message-----
From: Mark Lee [mailto:marklee15@xxxxxxxxx]=3D20
Sent: Tuesday, May 06, 2003 9:32 AM
To: thin@xxxxxxxxxxxxx
Subject: [THIN] Re: OT : Windows Performance Measure Query
Ruud is correct, it's showing Kernel times (ie. when the CPU is
performing a kernel mode operation as opposed to a user one!).
Basically CPU usage can be spilt into several broad area's - on Win2K
there are five - % DPC Time- % Interrupt Time- % Pivileged Time- %
Processor Time- % User Time
and if you check out Perfmon under the Processor object and click on
Explain it will give you a breif text about each option ! Suffice to
say, unless you are particularly bothered about each of these leave the
'Show Kernel Mode' box unticked - this will then show you the raw CPU
usage data as 'Total CPU Usage' for your system rather than splitting it
into Kernel mode etc. !!! Hope that explains things a little !!! Mark
eddie.vermeer@xxxxxxxxxx wrote:I know this is OT, but I'm kind of
stumped at the moment. I have a server
that when I look at the CPU usage through task manager on the small
graph,
I'm also getting a red line showing at different percentages as it moves
along in time with the green CPU Usage line. This is a single CPU server
and the CPU is running at a 'normal' level for the server. Everything
else
seems to be fine.
Has anyone see anything like this before or at least know what this red
line
means ?
TIA
Eddie
Web Site : www.tel-europe.com=3D20
eBusiness : www.telcustomer.com=3D20
-Delivering the technology connection-=3D20
Important Notice: This message is intended for the attention of the
person
or organisation to which it has been addressed. It may contain
information
that is confidential, proprietary and/or exempt from disclosure by law.
If
you have received this message in error please be aware that the sender
does
not waive any confidentiality, ownership rights or privilege applicable
to
it. You are therefore strictly prohibited to disseminate, distribute or
copy
this message in any way.
********************************************************
This Week's Sponsor - Emergent Online
EOL's Universal Printer new Features include:
Network Printing, Pagestreaming, 2400 DPI.
No Client Software Required!
http://www.go-eol.com/
**********************************************************
For Archives, to Unsubscribe, Subscribe or=3D20
set Digest or Vacation mode use the below link:
http://thethin.net/citrixlist.cfm
---------------------------------
Yahoo! Plus - For a better Internet experience
********************************************************
This Week's Sponsor - Emergent Online
EOL's Universal Printer new Features include:
Network Printing, Pagestreaming, 2400 DPI.
No Client Software Required!
http://www.go-eol.com/
**********************************************************
For Archives, to Unsubscribe, Subscribe or=3D20
set Digest or Vacation mode use the below link:
http://thethin.net/citrixlist.cfm
********************************************************
This Week's Sponsor - Emergent Online
EOL's Universal Printer new Features include:
Network Printing, Pagestreaming, 2400 DPI.
No Client Software Required!
http://www.go-eol.com/
**********************************************************
For Archives, to Unsubscribe, Subscribe or
set Digest or Vacation mode use the below link:
http://thethin.net/citrixlist.cfm
---------------------------------
Yahoo! Plus - For a better Internet experience
********************************************************
This Week's Sponsor - Emergent Online
EOL's Universal Printer new Features include:
Network Printing, Pagestreaming, 2400 DPI.
No Client Software Required!
http://www.go-eol.com/
**********************************************************
For Archives, to Unsubscribe, Subscribe or
set Digest or Vacation mode use the below link:
http://thethin.net/citrixlist.cfm
********************************************************
This Week's Sponsor - Emergent Online
EOL's Universal Printer new Features include:
Network Printing, Pagestreaming, 2400 DPI.
No Client Software Required!
http://www.go-eol.com/
**********************************************************
For Archives, to Unsubscribe, Subscribe or=20
set Digest or Vacation mode use the below link:
http://thethin.net/citrixlist.cfm
********************************************************
This Week's Sponsor - Emergent Online
EOL's Universal Printer new Features include:
Network Printing, Pagestreaming, 2400 DPI.
No Client Software Required!
http://www.go-eol.com/
**********************************************************
For Archives, to Unsubscribe, Subscribe or
set Digest or Vacation mode use the below link:
http://thethin.net/citrixlist.cfm
---------------------------------
Yahoo! Plus - For a better Internet experience
********************************************************
This Week's Sponsor - Emergent Online
EOL's Universal Printer new Features include:
Network Printing, Pagestreaming, 2400 DPI.
No Client Software Required!
http://www.go-eol.com/
**********************************************************
For Archives, to Unsubscribe, Subscribe or
set Digest or Vacation mode use the below link:
http://thethin.net/citrixlist.cfm
Other related posts:
