Hi Min,
Unfortunately, this CPU count strategy does not work with CCIN2P3-LCG2
site.
Indeed, the real number of physical CPUs supplied to the grid at CC is
100 CPUs at this moment, and more than 700 CPUs very soon (hopefully
next week).
With our batch system, we define in general 2 queues by physical CPU,
but each "queue" (called a WorkPoint) can be configured to accept
several classes of jobs (something like short (A), medium (G) and long
(T)).
At this moment, the Glue schema does not allow us to express the subtely
of our batch system queue mechanism. As a consequence, it is impossible
for you to infer the total of real CPUs from the published data. What we
currently publishes are virtual CPUs instead of real CPUs.
The solution consisting in adding CPU counts at Subcluster level,
proposed by stephen, should be the solution to our problem.
Anyway, in our specific and current case, the solution is to take the
max of the queues.
But a solution could be to sum systematically the CPUs of each queue by
site. Indeed, this value has the same meaning for all the sites.
According to me, it reflects the number of jobs a site pretends to be
able to perform simultaneously, what I call the virtual CPUs. It is the
role of the site administrator to set correctly this value in
concordance with the real capacities of his/her site.
When we will have the possibility to get the real CPU count, it will be
very interesting to compute the ratio between virtual CPUs and real ones.
Hope this helps ;).
Pierre
Min Tsai a écrit :
>Hi All,
>
>The fix is in for the CPU count. Three other sites had their CPU stats
>change: CCIN2P3-LCG2, INFN-LNL-LCG, INFN-PADOVA. Let me know if these
>number are inaccurate for some reason.
>
>Best Regards,
>Min
>
>-----Original Message-----
>From: LHC Computer Grid - Rollout [mailto:[log in to unmask]] On
>Behalf Of Min Tsai
>Sent: Wednesday, January 26, 2005 12:41 PM
>To: [log in to unmask]
>Subject: Re: [LCG-ROLLOUT] TotalCPU count on the GOC Mon
>
>Dear Anar,
>
>Typically CPU stats for queues on a single CE repeat all refer to the same
>set of CPUs. So to prevent recount of CPU Gstat adds up the CPU stats for
>the first queue it encounters for each unique CE. So in your case:
>
>It adds up:
>GlueCEUniqueID=lcg03.gsi.de:2119\/jobmanager-torque-alice 2 CPU
>GlueCEUniqueID=lcg06.gsi.de:2119\/jobmanager-lcglsf-alice 16 CPU
>
>I have not noticed a configuration like yours before, so I will make a
>modification by adding CPU from queues that have different total CPU
>statistics even though they reside on the same CE. The only problem we will
>have if when 2 queues on a single CE has the same total CPU count even
>though they are referring to 2 completely different clusters.
>
>I hope this will correct the CPU problem for you site. Thank you for
>providing this feedback! I will let you know once this I have tested and
>complete this change.
>
>Cheers,
>Min
>
>
>
>
>
>-----Original Message-----
>From: LHC Computer Grid - Rollout [mailto:[log in to unmask]] On
>Behalf Of Anar Manafov
>Sent: Wednesday, January 26, 2005 11:59 AM
>To: [log in to unmask]
>Subject: [LCG-ROLLOUT] TotalCPU count on the GOC Mon
>
>Good day to ALL!
>
>I have mentioned that on the monitoring (http://goc.grid.sinica.edu.tw/
>gstat/lcg03.gsi.de/) we (GSI) publishing only 18 CPU (Total CPU). So, I
>wonder how this number is calculated and why not all of the queues are
>affected.
>We have 2 different CE:
>Torque CE (with 2 CPU).
>LSF CE (more than 300 CPU),
>in LSF we have “dteam” and “alice” queues.
>For “alice” ~ 16 PCU
>For “dteam” ~ 344 CPU or something (Later on, when we finish
>the test of
>our new pool-accounts algorithm we will publish more CPU on the
>“alice”).
>
>So, my question would be which algorithm monitoring uses to calculate Total
>CPU amount?
>
>I would appreciate any comment on this.
>
>Thank you very much in advance.
>
>Best of luck,
>
>Anar
>
>
>
--
______________________
Pierre GIRARD
Grid Computing Team Member
IN2P3/CNRS Computing Centre - Lyon (FRANCE)
http://cc.in2p3.fr
Tel. +33 4.78.93.08.80 | Fax. +33 4.72.69.41.70 | e-mail: [log in to unmask]
|