vgrme/avx512-grafana

Demonstrating AVX512 vs AVX2 on Red Hat OpenShift

Authors: Rhys Oxenham

This repository aims to demonstrate AVX512 vs AVX2 performance in the Monte Carlo simulation (which should result in ~50% perf increase), and how to set it up visually in Red Hat OpenShift.

NOTE: You will need to ensure that you have nodes that support AVX512/AVX2 in your OpenShift cluster, or the workloads will fail to schedule!

The deployment script does the following-

• Create a new dedicated OpenShift project "monte-carlo"
• Enable user-workload monitoring in built-in OpenShift monitoring
• Deploy an AVX512 and AVX2 based Monte Carlo simulation pod (with Prometheus Push Gateway)
• Enable ServiceMonitors for the two workload pods to scrape gateways
• Deploy the Grafana Community Operator (v4.7.0) and create an instance
• Create a Grafana service account to have access to Prometheus/Thanos data (built-in)
• Add Grafana Datasource for Prometheus/Thanos data (built-in)
• Deploy a Custom Grafana Dashboard to show "time_elapsed" between AVX2 and AVX512
• Show the Grafana Dashboard Route to the user

NOTE: Make sure that you've deployed the NodeFeatureDiscovery operator, and an equivalent instance, as we tag the workload pods to nodes that support required features:

spec:
  restartPolicy: Never
  nodeSelector:
    feature.node.kubernetes.io/cpu-cpuid.AVX512BW: "true" # requires NFD

Deployment

To deploy the environment, simply run the following, assuming you have already configured your OpenShift client and have system:admin access:

# oc whoami
system:admin

# ./deploy.sh
(...)

Clean Up

To clean up all of the resources, simply run the following-

# ./cleanup.sh
(...)