Benchmarks

Hardware

* Performance per node.

The green values are measured according to CERN procedures.

Virtual Machines

Available Resources

This are the CPU recourses available on Phoenix. HEP-SPEC06 is measured according to the CERN procedures.

* 4 HEP-SPEC06 = 1 kSI2k

How to Update Published Values

• On ARC CEs it is enough to update the corresponding lines in arc.conf

[root@arc01 ~]# vim /etc/arc.conf
[...]                                                                                                     
jobreport_options="archiving:<%= scope.lookupvar('arc_ce::archiving_dir') %>,topic:/queue/global.accounting.cpu.central,gocdb_name:CSCS-LCG2,benchmark_type:HEPSPEC,benchmark_value:11.46,use_ssl:true"             
[...]                                                                                                     
cpu_scaling_reference_si00="2600"                                                                         
processor_other_description="Cores=35.84, Benchmark=11.46-HEP-SPEC06"                                     
[...]                                                                                                     
opsys="6.7"                                                                                               
[...]                                                                                                     
benchmark="SPECINT2000 2600"                                                                              
[...]                                                                                                     

[root@arc01 ~]# /etc/init.d/nordugrid-arc-inforeg stop      
[root@arc01 ~]# /etc/init.d/nordugrid-arc-ldap-infosys stop 
[root@arc01 ~]# /etc/init.d/nordugrid-arc-ldap-infosys start
[root@arc01 ~]# /etc/init.d/nordugrid-arc-inforeg start     

• On CREAM CEs the correct values should be set in site-info.def and then Yaim run to apply the new values into GIP configuration

[root@cfengine ~]# vim -R /srv/cfengine/files/ANY/opt/cscs/siteinfo/site-info.def
[...]
CE_CPU_SPEED=2600                                                                           
CE_OTHERDESCR="Cores=35.84,Benchmark=11.46-HEP-SPEC06"                                      
CE_CAPABILITY="CPUScalingReferenceSI00=2600 Share=atlas:40 Share=cms:40 Share=lhcb:20 glexec
CE_SI00=2600                                                                                
CE_MINPHYSMEM=32768                                                                         
CE_MINVIRTMEM=96256                                                                         
CE_OS_="6.7"                                                                         
[...]

[root@cream02:~]# /opt/glite/yaim/bin/yaim -r -s /opt/cscs/siteinfo/site-info.def -n creame
 -n SLURM_utils -f config_bdii_5.2 -f config_cream_gip_info_dynamic_setenv -f config_cream_gip_info_dyna
c -f config_cream_gip_setenv -f config_cream_gip -f config_cream_gip_glue2_setenv -f config_cream_gip_gl
2 -f config_cream_gip_scheduler_plugin_setenv -f config_cream_gip_scheduler_plugin -f config_cream_gip_s
tware_plugin_setenv -f config_cream_gip_software_plugin -f config_cream_vo_tag_dir -f config_info_servic
cream_ce_setenv -f config_info_service_cream_ce -f config_info_service_cemon_setenv -f config_info_servi
_cemon -f config_cream_info_service_rtepublish_setenv -f config_cream_info_service_rtepublish        

 # if this does not work, just modify by hand

[root@cream01:~]# vim /var/lib/bdii/gip/ldif/static-file-CE.ldif     
[root@cream01:~]# vim /var/lib/bdii/gip/ldif/static-file-Cluster.ldif
[root@cream01:~]# vim /var/lib/bdii/gip/ldif/ComputingManager.ldif   
 
 # and restart bdii service

After a while the new values should be published by the site BDII:

Feb 11 15:08 [rgianni@ui:~]$ ldapsearch -x -LLL -h bdii.lcg.cscs.ch -p 2170 -b 'o=grid'

Possible Improvements

Not all of the jobs are CPU bound. Many of them are waiting for data. Therefore we could enable Hyper Threading and allow 12 jobs per WN. The other 4 cores will be reserved for the OPS/SGM and CSCS queue only.

• Advantages
  • More CPU slots 1152 vs. 768
  • Better utilization of CPU
  • Better overall performance
  • Better match of CPU pledges for 2010 (about 13% more performance, see below)
  • Cores of CSCS queue available for testing; faster detection/debugging of problems
• Disadvantages
  • Possibly less performance per single job slot
  • Accounting less accurate
    • Performance of a single job depends a lot on the number of running jobs / WN

* 4 HEP-SPEC06 = 1 kSI2k

Resources

• HEPSPEC06 at hepix
• HEPSPEC06 at CERN
• Spec2k at hepix
• Spec2k at sinica.edu.tw