Jim, I'd say your append can introduce confusion.

The vcpus still count for absolute shares.  But, indeed, the number of real
CPUs come into account: the % written on the SET SHAER command is expressed
as a % of the whole system.
If you have 10 CPU's, a SET SHARE ABS 5% LIMITHARD means that a uniprocessor
guest can consume 50% of one CPU.  If the guest has two virtual processors,
each one can consume 25% at most.  This is very important if this guest
would run one heavy uniprocessing task, this task will have a hard time to
get more than 25% of a CPU.  We learned this the hard way when installing a
6 engine 9672 ages ago (our heavy process was a VM DB2 server)

2007/8/3, James Vincent <[EMAIL PROTECTED]>:
>
> There is actually more science to this than one would imagine.  We just
> went through this and chatted with Bill Bitner on it to try and understand
> it all.  Here are the things to keep in mind for setting SHARE (besides what
> is already written in the HELP):
>
> For RELATIVE share, the amount should be X times the number of virtual
> processors the guest has.  So if you want all servers to be "equal" at REL
> 100 and you have a server with two virtual CPUs, if you set it to REL 100 -
> it is really setting to REL 50 for EACH vcpu.   It would need REL 200 to be
> equal to all others at REL 100.
>
> For ABSOLUTE the vcpus don't matter, but the REAL cpus do.  If you want to
> cap the web server to not use more than 50% of a real processor, then it is
> X% / #_real_cpus    That means if you have a server you want to cap at 50%
> and you have 11 IFLs, you would use SET SHARE REL nnn ABSOLUTE 4.5%LIMITHARD  
>  (that is 50 / 11 )  The LIMITHARD will prevent it from going
> over it's limit, otherwise if the cpu is available the server will get it.
>
> It takes a little to get used to and I just covered some basics.
> Hopefully with everyone elses update and this it gets you set for what you
> need.
>
> Jim Vincent
>
> On 8/2/07, Rob van der Heij <[EMAIL PROTECTED]> wrote:
> >
> > On 8/2/07, Anne Crabtree <[EMAIL PROTECTED]> wrote:
> > > We have five linux instances running, two of which aren't really being
> > > used, two are production and one is test.  In the test linux instance,
> >
> > > they are running a web app (which I know nothing about).  At certain
> > > times during the day, this test instance will take close to 100% of
> > the
> > > IFL.  We want to back it off so that it can take no more than 50% to
> > see
> > > how that affects them.  Would it be logical to issue SET SHARE linux
> > > RELATIVE 50 since it is defaulting to REL 100?  Will that do what I
> > > think and limit it to half of what the other virtual machines "get"?
> >
> > Really depends on what you want to achieve.
> > * When you set REL 50 and others have REL 100, the one with smaller
> > share will get only half of what another would get *if there is CPU
> > shortage*  It's something you do for test versus production servers
> > * If you want to "kneecap" the virtual machine to see what the
> > application behavior would have been when you did not have all that
> > spare capacity, then a share with LIMITHARD is useful (and you want
> > that absolute% because otherwise you have a moving target).
> >
> > And for virtual machines with multiple virtual CPU's, you need to
> > realize share is divided by the number of virtual CPU's
> >
> > When a virtual machine is looping or doing useless work that you want
> > to get rid of, then FORCE is more appropriate ;-)   But yes, I have
> > sometimes set virtual machines to REL 1 LIMITHARD to let them run
> > until I have time to start diagnose the problem.
> >
> > Rob
> > --
> > Rob van der Heij
> > Velocity Software, Inc
> > http://velocitysoftware.com/
> >
>
>

-- 
Kris Buelens,
IBM Belgium, VM customer support

Reply via email to