Event query discovery from Google cluster traces · run 2026-10-03 13:46–18:25 CEST Workflow Run uri icon

start time

  • October 3, 2026 1:46 PM

end time

  • October 3, 2026 6:25 PM

duration (s)

  • 16701

duration time calculation method

  • Slurm ElapsedRaw of the job (start to end). Display timestamps use Europe/Berlin.

run status

  • Succeeded

CPU efficiency (%)

  • 0.5

data read (GB)

  • 142.9910061056

data written (GB)

  • 0.9391990374400001

CPU time (s)

  • 22260.0

CPU time calculation method

  • Slurm TotalCPU of the job (user + system CPU time of all steps); 256 CPUs allocated of 256 node CPUs.

task count

  • 3

peak memory (GB)

  • 0.549265408

GPU usage status

  • GPU requested in Slurm allocation

memory calculation method

  • Peak memory: the largest resident memory (Slurm MaxRSS) of any job step, converted from bytes to gigabytes. No average memory: Slurm does not record memory use over time.

energy consumed (kWh)

  • 1.9421783179398344

energy measurement coverage

  • 100% of the node (whole node, measured)

node power average (W)

  • 418.6

node power maximum (W)

  • 545.0

node sleeping idle power (W)

  • 283.5

node awake idle power (W)

  • 372.6

carbon emission (kg CO2e)

  • 0.6924147429053855

carbon intensity (kg CO2e/kWh)

  • 0.35651450565047216

energy metric source

  • IPMI power of the compute node read through its Prometheus ipmi_exporter; Slurm acct_gather_energy/ipmi for comparison

energy calculation method

  • Node energy: IPMI power of the node (ipmi_exporter) read about every 10 s while the job ran (1671 readings, 384 to 545 W, covering 100% of the job), integrated over time and scaled to the job duration (16701 s): 6991842 J. The whole node was allocated to this job, so all of it is the job's energy, converted to kilowatt-hours (1 kWh = 3,600,000 J): 1.942178 kWh. Slurm accounting (acct_gather_energy/ipmi) reports 6119223 J; it is not used, because it would mean an average power of 366 W, outside the node's IPMI readings during the job (384 to 545 W). Single node hpc-cms01-007 (HPC@HU), allocated entirely to this job (all 256 CPUs). Energy is the measured energy of the whole node, including its idle power; cooling and network are excluded.

carbon calculation method

  • Time-matched: the run's energy 1.942178 kWh x the 15-minute carbon intensity of the grid (Electricity Maps, zone DE, lifecycle, flow-traced) in the quarter-hours the run used, weighted by the run's energy in each quarter-hour (split by power readings; 19 quarter-hours from 2026-10-03T11:45 to 2026-10-03T16:30 UTC, 220 to 525 g/kWh) = 356.5 g/kWh, so 692.41 g CO2e. Window start and end are Europe/Berlin times. Recalculated on 2026-11-07: replaces the earlier value 690.74 g CO2e (355.7 g/kWh, Electricity Maps).

idle power method

  • Node hpc-cms01-007 saves energy when nothing runs: its idle processor cores sleep (low-power states), while the node itself stays on and takes new jobs at once. Sleeping idle power: 283.5 W, the median IPMI power of the node in the cluster monitoring (Prometheus, a reading every 5 s) over the 30 days up to 2026-10-30, counting only times when nothing had run on the node for at least 30 minutes (load below 0.2, both GPUs below 60 W). Awake idle power: 372.6 W, the average IPMI power with the whole node allocated to an empty job while its power was read once a second on the node (Slurm job 1598720, 2026-10-03, 302 readings). The readings keep the cores from sleeping, so the node draws 89 W more. No energy above idle is given for this run: the node's cores sleep when idle, and its idle power without sleep is not known. The node is assumed to have had its current power profile also on the day of the run.

git commit

  • c898ae4fda50a3463e992344c946f9b983fae7e7

job name

  • event-query-discovery (Slurm job 1598721, partition standard, exit code 0:0)

code version

  • sha256:89d9538633ba919f433b818e2dc32398966c4b5e8f71cef7553dde493c1a6625

trace types

  • Slurm accounting (sacct); node CPU and IPMI power samples; node hardware inventory; node IPMI temperature readings; stage records (times, exit status, CPU time, peak memory)

trace data format

  • TSV, JSON and plain text

CPUs used

  • All 256 CPUs (hardware threads) of the node: 128 cores on 2 processors.

fan speed average (rpm)

  • 7899.0

node CPU utilisation average (%)

  • 0.6

node conditions method

  • Node CPU utilisation: CPU time of all processes on the node (node_exporter, all modes except idle and iowait) / (256 CPUs x sampled time) = 0.6% (1.5 of 256 CPUs busy on average). Node power: IPMI power of the node (ipmi_exporter), 1671 readings about every 10 s: average 419 W (over time), lowest 384 W, highest 545 W. CPU efficiency: job CPU time 22260 s / (256 allocated CPUs x 16701 s) = 0.5%. CPU frequency: current frequency of the 256 CPUs the job could use (/proc/cpuinfo), 1671 readings. At each reading the fastest of these CPUs is taken, because the CPUs that are working run fastest: average 3641 MHz, highest 3704 MHz. Average over all of these CPUs, including idle ones: 1575 MHz. Fan speed: IPMI fan sensors of the node (ipmi_exporter), 6 fans, 1671 readings: average 7899 rpm (mean of the fans, over time), highest 8400 rpm. Data read and written: Slurm accounting MaxDiskRead and MaxDiskWrite of the job's 1 step(s): 142.991 GB read, 0.939 GB written. This is all data the job's processes read and wrote, on any file system.

workflow engine

execution host

  • hpc-cms01-007

node allocatable CPU

  • 256

node allocatable memory (GB)

  • 1081.494478848

node architecture

  • x86_64; 2 x AMD EPYC 7713 64-Core Processor; 256 hardware threads

os image

  • openSUSE Leap 16.0

kernel version

  • 6.12.0-160000.36-default

backend