[gridengine users] CPU = WALLCLOCK time in the share tree, policy ?

Orlando Richards orlando.richards at ed.ac.uk
Wed May 25 08:47:01 UTC 2011

On 24/05/11 18:43, users-request at gridengine.org wrote:
> Message: 6
> Date: Tue, 24 May 2011 17:48:47 +0200
> From: Reuti<reuti at staff.uni-marburg.de>
> To: Dave Love<d.love at liverpool.ac.uk>
> Cc:"users at gridengine.org Group"  <users at gridengine.org>
> Subject: Re: [gridengine users] CPU = WALLCLOCK time in the share tree
> 	policy ?
> Message-ID:
> 	<F4D4A572-EA78-47E7-A2BF-C6AF7234397B at staff.uni-marburg.de>
> Content-Type: text/plain; charset=US-ASCII; format=flowed; delsp=yes
> Am 24.05.2011 um 15:35 schrieb Dave Love:
>> >  ..._RESERVED_USAGE are at least badly documented.  When I came across
>> >  them I interpreted `reserved' to mean resource reservation, i.e. the
>> >  overhead from doing that.  I'll fix the doc; does anyone have a good
>> >  suggestion for a better description?  I'm inclined to say `available
>> >  resources' and explain further.  I haven't got round to checking,
>> >  but I
>> >  assume what's used for cpu is running time multiplied by slot count;
>> >  anyone know what's used for io, which is specifically mentioned?
>> >
>> >  Also, I wonder why SHARETREE_RESERVED_USAGE is deprecated, as it seems
>> >  more useful default.
> Although someone stated on the list to use it in the way:
> (and hence complained about unifiying it), I think most often you will
> set just both to TRUE or FALSE, and so just one setting would be
> enough and imply the second.
> -- Reuti

That was me, but the other way around - SHARETREE_RESERVED_USAGE=TRUE, 

We'd still find it useful to be able to see how much CPU time was used 
on the system, as opposed to wallclock time. However, we use wallclock 
time for accounting purposes, and therefore also in determining share 
tree tickets.

With the above configuration, the system wasn't honouring the 
ACCT_RESERVED_USAGE flag, and just reports wallclock time in the CPUTIME 
field (using 6.2u5 from Sun).

There were two key reasons for wanting to be able to see the CPU time on 
the system:
  - some user groups (notably GridPP who process CERN work) use it as a 
metric for measuring the efficiency of a given system (cpu time / 
wallclock time = efficiency)
  - In general, it's a useful metric to see what how your system is 
being used, and where more can be squeezed out of it.

So I'd still be in favour of keeping both in! Having said that - the 
value of CPU utilisation can be obtained by doing (UTIME + STIME), but 
it still seems a bit odd to have a CPUTIME field which doesn't give you 
CPUTIME. Especially when there's a WALLCLOCK field sitting right there 

If anything - I'd suggest that the ACCT_RESERVED_USAGE field could be 
deprecated with respects to manipulating the data which is put into the 
accounting file. I'd always prefer that the raw data goes into the 
accounting file, which doesn't vary depending on configuration settings.

    Dr Orlando Richards
   Information Services
IT Infrastructure Division
        Unix Section
     Tel: 0131 650 4994

The University of Edinburgh is a charitable body, registered in 
Scotland, with registration number SC005336.

More information about the users mailing list