[Orca-users] RE: Workload format in orcallator.se

Cockcroft, Adrian acockcroft at ebay.com
Wed Jun 8 13:31:06 PDT 2005 

I think its best to capture workload oriented CPU usage using total CPU
time rather than dividing by the number of CPUs, this is because it is
common to move workloads between machines (consolidation) and to have
workloads that span dissimilar machines (pooling) and a V880 should have
a CPU capacity of 800% compared to an E2900 having 2400% capacity, with
a workload described as needing 350% showing similar resource usage
metrics on both.

Dividing down for a particular plot is a good option to have for more
static environments, but when system configurations change on a
weekly/daily basis it makes comparisons across time easier if the number
of CPUs is explicit.

It's good to see the workollator code integrated into orca and being
used...

Cheers Adrian
http://perfcap.blogspot.com

-----Original Message-----
From: orca-users-bounces+acockcroft=ebay.com at orcaware.com
[mailto:orca-users-bounces+acockcroft=ebay.com at orcaware.com] On Behalf
Of Dmitry Berezin
Sent: Wednesday, June 08, 2005 12:39 PM
To: 'Buhler Justin'
Cc: orca-users at orcaware.com
Subject: RE: [Orca-users] RE: Workload format in orcallator.se

Justin,

In short yes, this is the reason.

Here is how it works in detail -
The way usr% (and other cpu times) are calculated is by taking total
time
(total user time, total system time, etc.) for the process(es) of
interest
within given interval, dividing it by that interval and multiplying it
by
100 to get a percentage. On a multi-processor system, multiple processes
use
different CPUs at the same time, which is not being accounted for while
percentages are calculated. By dividing that number by the number of
CPUs,
we get an average CPU usage for the whole system, which is what we need
for
performance and trend analysis. I suppose it would be possible to get a
per
CPU usage numbers, but this would require more coding.
I decided to record the actual time (which could obviously be greater
then
100% for a group of processes) in case somebody has a need to plot it.
For
instance, if one has systems with different numbers of CPUs and runs the
same application on them, it might be useful to see how the application
behaves on these systems by comparing the total CPU usage (say 150% on
one
system vs. 125% on the other). For my systems I only need to see the
system
performance, not the application, so I do the division while plotting.

Did this answer your question?

  -Dmitry.


> -----Original Message-----
> From: orca-users-bounces+dberezin=acs.rutgers.edu at orcaware.com
> [mailto:orca-users-bounces+dberezin=acs.rutgers.edu at orcaware.com] On
> Behalf Of Buhler Justin
> Sent: Wednesday, June 08, 2005 11:38 AM
> To: Dmitry Berezin
> Cc: orca-users at orcaware.com
> Subject: RE: [Orca-users] RE: Workload format in orcallator.se
> 
> Thanks for the explanation !!!! Just another clarification about the
cfg
> file. When you list the following:
> 
> data                    (.*)_usr% / ncpus
> data                    $1_sys% / ncpus
> data                    $1_cpwt% / ncpus
> data                    $1_pgwt% / ncpus
> 
> You are dividing the result by the Number of CPU's to get the per CPU
> usage time? Is this because kstat is providing the total % of the sum
of
> the cpus available.
> 
> This is a question/clarification?
> 
> Justin
> 
> -----Original Message-----
> From: Dmitry Berezin [mailto:dberezin at surfside.rutgers.edu]
> Sent: Tuesday, June 07, 2005 7:11 PM
> To: Buhler Justin; orca-users at orcaware.com
> Subject: RE: [Orca-users] RE: Workload format in orcallator.se
> 
> Justin,
> 
> Here is what you need to get workload information with orcallator.se -
> 
> 1. Set and export WORKLOAD_INFO (see examples below)
> 2. Set and export WORKLOAD_FIRSTMATCH=0 if you want to match each
process
> to
> all workloads (optional, default setting is 1)
> 3. Run orcallator.se with -DWATCH_WORKLOAD to enable workload code
> 4. Run orcallator.se with -DWORKINFO_PRINT to have your workload
settings
> printed out for verification (optional)
> 
> There are a couple of examples of setting WORKLOAD_INFO in the
> workinfo_class.se file (take a look at the Usage notes section). Here
are
> a
> few more -
> 
> This defines three workloads:
> WEB   - all httpd processes
> JAVA  - all java processes owned by user javauser
> OTHER - all other processes
> Process names (httpd and java) and user name (javauser) are matched as
> patterns.
> 
> WORKLOAD_INFO="WEB,httpd;JAVA,java, ,javauser;OTHER"
> 
> The same as above, but httpd process is matched exactly.
> 
> WORKLOAD_INFO="WEB,httpd, , ,1;JAVA,java, ,javauser;OTHER"
> 
> The next example collects data for all processes owned by user oracle
> (ORACLE), as well as data for different oracle database instances
> separately
> (ORADB1 and ORADB2). In addition it defines NetBackup workload (NB) -
all
> processes that begin with "bp". Notice that in order to collect data
for
> all
> oracle databases together and separately at the same time, you need to
set
> WORKLOAD_FIRSTMATCH=0 which implies that you can not use OTHER as in
the
> example above, since it will just catch everything.
> 
> WORKLOAD_FIRSTMATCH=0
> WORKLOAD_INFO="ORACLE, , ,oracle,1;ORADB1, ,ORADB1,oracle;ORADB2,
> ,ORADB2,oracle;NB,^bp"
> 
> This will just collect process and thread count data for all httpd
> processes. Replace the first 1 with a space to change to pattern match
if
> desire.
> 
> WORKLOAD_INFO="WEB,httpd, , ,1,1'
> 
> Here is the example of orcallator.cfg file settings -
> 
> http://www.orcaware.com/pipermail/orca-users/2005-May/004325.html
> 
> I hope this is not too wordy and is what you need. If it is not, feel
free
> to ask further questions.
> 
>   -Dmitry.
> 
> 
> > -----Original Message-----
> > From: orca-users-bounces+dberezin=acs.rutgers.edu at orcaware.com
> > [mailto:orca-users-bounces+dberezin=acs.rutgers.edu at orcaware.com] On
> > Behalf Of Buhler Justin
> > Sent: Tuesday, June 07, 2005 10:39 AM
> > To: orca-users at orcaware.com
> > Subject: [Orca-users] RE: Workload format in orcallator.se
> >
> >
> > Dmitry,
> >
> > Can I get an example of the workload format in the orcallator.se
> > configuration file.
> >
> > Cheers.
> >
> > Justin
> >
> >
> > _______________________________________________
> > Orca-users mailing list
> > Orca-users at orcaware.com
> > http://www.orcaware.com/mailman/listinfo/orca-users
> 
> 
> 
> --
> No virus found in this outgoing message.
> Checked by AVG Anti-Virus.
> Version: 7.0.323 / Virus Database: 267.6.5 - Release Date: 6/7/2005
> 
> 
> _______________________________________________
> Orca-users mailing list
> Orca-users at orcaware.com
> http://www.orcaware.com/mailman/listinfo/orca-users


_______________________________________________
Orca-users mailing list
Orca-users at orcaware.com
http://www.orcaware.com/mailman/listinfo/orca-users

More information about the Orca-users mailing list