Understanding Kepler Model Server Benchmark Tests

[husky-parul]

Hi

I am trying to use KMS benchmarks to collect metrics for PEAKS following this tutorial

I am running this tests on single node kind cluster which I set up using ./script.sh prepare_cluster

kubectl get nodes -o wide
NAME                              STATUS   ROLES           AGE   VERSION   INTERNAL-IP   EXTERNAL-IP   OS-IMAGE             KERNEL-VERSION           CONTAINER-RUNTIME
kind-for-training-control-plane   Ready    control-plane   19h   v1.25.3   172.18.0.2    <none>        Ubuntu 22.04.1 LTS   6.4.12-200.fc38.x86_64   containerd://1.6.9

After that I ran the test benchmark ./script.sh quick_collect. I have the output in model_training/data/sample_kepler_query.json. In the json file I see the following metrics

I am interested in capturing the total power consumption of each of the nodes in the cluster. For this which metrics should I use from the following. Also where can I find more information on the relationship/difference between these metrics.

• kepler_node_core_joules_total
• kepler_node_platform_joules_total
• kepler_node_uncore_joules_total
• kepler_node_other_joules_total
• kepler_node_package_joules_total

Thanks.

cc @knarayan @sunya-ch @marceloamaral

[sunya-ch]

I would suggest to start from kepler_node_package_joules_total. In the most system, core and uncore are not available.

Simple formula is
platform = package + dram + gpu + other (https://github.com/sustainable-computing-io/kepler/blob/63fb1c4434cb693642fb7b024900dcdfda101a6a/pkg/collector/metric/node_metric.go#L422)

rapl reports package, core, uncore, and dram
nvml reports gpu
acpi reports platform
other is computed from the above formula.

You may check metric description here: https://sustainable-computing.io/design/metrics/#kepler-metrics-for-node-information.
We may need to enrich the explanation if it is unclear.