WLCG Sustainability activities

Sustainable HEP

General

• e.g. Data Center best practice example, W/HS06

2022-23 Energy Crisis

GDB

Power reduction strategies

• DESY HH cgroup frequency modulation experience
  • for reducing power consumption while working on power supply lines, added a path service unit to reduce HTCondor parent cgroup CPU quotas https://sas.desy.de/index.php/s/fPKkyjLm3SQ4p7F
  • setting LRMS cgroup CPU time quota to a fraction of the CPU quota period, e.g., /sys/fs/cgroup/cpu,cpuacct/system.slice/condor.service/cpu.cfs_quota_us (= 500000) /sys/fs/cgroup/cpu,cpuacct/system.slice/condor.service/cpu.cfs_period_us (= 1000000) to allow all batch jobs on that node to run only for 33% of the total CPUs/cores walltime (could also be applied on the jobs' sub-groups)
  • transparent for the jobs
  • example: power consumption of node batch0502.desy.de
    • 1hour of 99% CPU period quota assigned to the htcondor parent group vs max 33% period quota allowed https://sas.desy.de/index.php/s/LCmktyAtcjBP6Tt
      • 99% @ 1h = ~0.2 kWh vs 50% @ @ 1h = ~0.1 kWh -- within the limits of the power reading accuracy... (#.# kWh) https://sas.desy.de/index.php/s/wZf5iKTLdRrD5yy
    • power consumption estimated from the (Dell) chassis power readings (`/opt/dell/srvadmin/bin/omreport chassis pwrmonitoring`)
    • would have to weight in total wall time with the hyperthreaded cores vs. non-HT set up

Tools

• Pull forecast, schedule reduction and execute various strategies: a first implementation

-- RodneyWalker - 2022-09-15