Aims of the analysis

31
Comparison between Gridice and Boss data related to LCG0 production G. Maggi, M. Maggi, A.Pierro, N. De Filippis, G. Donvito, T. Coviello

description

Comparison between Gridice and Boss data related to LCG0 production G. Maggi, M. Maggi, A.Pierro, N. De Filippis, G. Donvito, T. Coviello. Aims of the analysis. To extract some parameters from LCG0 production data such as farm performance job lifetime CPU load and used memory - PowerPoint PPT Presentation

Transcript of Aims of the analysis

Page 1: Aims of the analysis

Comparison between Gridice and Boss

data related to LCG0 production

G. Maggi, M. Maggi, A.Pierro, N. De Filippis, G. Donvito, T. Coviello

Page 2: Aims of the analysis

Aims of the analysis• To extract some parameters from LCG0

production data such as– farm performance

– job lifetime

– CPU load and used memory

– LAN and WAN occupancy

for the simulation of the CMS analysis environment

• To select some reference distributions to be reproduced by the simulation

Page 3: Aims of the analysis

User Interfaces & Boss databases • Bari• Padova• Bologna• Ecole Polytecnique (France)

Job type dataset Number of Jobs

(250 events/job)

period

Cmsim133 mu03_bb2mu 400x10=4000 October 7-2003

Cmsim133 mu03_MB 400x5=2000 September 11-2003

genPYT110 mu03_MB 400x5=2000 July 11-2003

Total 8000 2000000 events

Page 4: Aims of the analysis
Page 5: Aims of the analysis

CMS-LCG0 testbed

   Farm   N

 PC Type

 HT

 Clock (MHz)

Memory(Mbytes)

             1 Bari 1 dual P III no 1000 1000

2 dual P III no 1133 1000

6 dual P III no 1266 1000

2 Bologna 10 dual P III no 1000 500

1 dual P IV(Xeon)

no 2400 1000

3 Padova 9 dual P III no 1000 500

8 dual P III no 1266 1000

9 dual P IV (Xeon) no 2400 1000

4 LNL 13 dual P III no 1000 500

14 dual P IV(Xeon)

no 2400 1000

5 CERN 10 dual P IV (Xeon) no 2800 1000

6 CNAF 2 dual P III no 1000  

7 E. Polytecnique 2.5 dual P III no 800 500

Page 6: Aims of the analysis

Jobs per farm and computer type

Page 7: Aims of the analysis

Performances

Page 8: Aims of the analysis

Job lifetime (dataset CMSim mu03 bb2mu)

Page 9: Aims of the analysis

Table of jobs lifetime (dataset CMSim mu03 bb2mu )

(*) Long tail (however the peak is at 40000 s)

8338440,0841905226102400

4084760,0642571401101000Lnl

9849840,049944193902800Cern

8338370,0591355227902400

6236820,0802245279801266

4084660,1184850409401000Padova

8338560,0601337222802400

(*)4084540,1656944420401000Bologna

6236930,1123072275401266

5626160,0541686309701133

4084780,0622490399201000Bari

 Specint 2000Measured Specint

RMS/averageRMSAverage timePC clockFarm

Page 10: Aims of the analysis

Specint

• The “Measured Specint” is obtained by this formula:

XtypePCtimeaverage

MHzPIVtimeaverageMHzPIVSpec

XtypePCSpec

2800)2800int(

)int(

• “Specint 2000” is obtained from the web: http://www.specbench.org/osg/cpu2000/results/cint2000.html

Page 11: Aims of the analysis

Measured Specint

Page 12: Aims of the analysis

Notes about measured Specint

• The Padova Farm, in dataset sim_mu03_MB and Kin_mu03_MB, did not behave as the others. Why?

• The “measured Specint” does not depend on the Farm– ie: does not depend on element’s number in

the Farm.

• The “measured Specint” depends on assignment (kin, sim, etc)

Page 13: Aims of the analysis

Measured Specint used for jobs lifetime normalization

CPU SPEC CPU2000obtain from

www.specbench.org

Measured specint from cmsim mu03_bb2mu

1000MHZ 408 469

1133MHZ 562 616

1266MHZ 623 688

2400MHZ 833 846

2800MHZ 984 984

Page 14: Aims of the analysis

Measured and theoretical Specint

Page 15: Aims of the analysis

dataset sim_mu03_bb2mu

Page 16: Aims of the analysis

Job lifetime (dataset CMSim mu03 MB)

Page 17: Aims of the analysis

dataset sim_mu03_MB using “measured

specint” of sim_mu03_bb2mu

Page 18: Aims of the analysis

Job lifetime (dataset CMKin mu03 MB)

Page 19: Aims of the analysis

dataset kin using “measured specint” of sim_mu03_bb2mu

Page 20: Aims of the analysis

Comparison between gridice and boss data

• Number of jobs running and waiting according to gridice and boss

• The agreement is reproduced except for a shift in time due to a not perfect synchronization between gridice and UI time (about 7 minutes shift for the Bari farm)

Page 21: Aims of the analysis

Running jobs on Padova farmdate: 01 September – 08 September

Page 22: Aims of the analysis

Running jobs on Bari farmdate: 08 September – 15 September

Page 23: Aims of the analysis

Running jobs on single pc of Bologna farm, load cpu and load ram vs time (s)

Page 24: Aims of the analysis

Used Memory

 cmsim cmkin

PC clock Average Ram Used

Average Ram Used

1000 75MB 10MB

1266 80MB  

2400 90MB 

Page 25: Aims of the analysis

Running job vs Transfer rate (Byte)

Page 26: Aims of the analysis

Summary

• With this analysis we have– Measured the execution time of jobs on different CPU

– Modelled the CPU performance

– Measured the farm performance

– Measured the resource load (RAM, CPU, transfer rate)

– Validated the Boss data with Gridice ones

• Now we are starting with the simulation of the computing model (Ptolemy, Monarc, etc.)

Page 27: Aims of the analysis

job lifetime – cpu time = I/O Time dataset CMSIM

PDAss3195

BAAss3195

BOAss3195

Page 28: Aims of the analysis

job lifetime – cpu time = I/O Time dataset CMSIM

LNLAss3340

CERNAss3340

Page 29: Aims of the analysis

Lifetime of job –execution time cpu = I/O Time dataset CMKIM

BAAss2653 BO

Ass2653

CERNAss2653

Page 30: Aims of the analysis

Lifetime of job –execution time cpu = I/O Time dataset CMKIM

LNLAss2653

PDAss2653

Page 31: Aims of the analysis

Notes about I/O Time

• The I/O time of all Farm for CMSIM has a peak about to 60-100 seconds except CERN farm where the peak is to 35 seconds

• The I/O time of CMKIN is about ten seconds according to kind of Assignment