mx-exporter - MetaX Prometheus Data Exporter

The mx-exporter is a standalone app that can be run as a daemon, written in Python Language, that exports MetaX GPU metrics to the Prometheus server. The mx-exporter uses MXSML C-Library for its data acquisition. The exporter and the MXSML library have a [PYTHON binding] that provides an interface between the MXSML C,C++ library and the PYTHON exporter code.

Build docker image

$ sudo make mxc MACAVER=xxxx ARCH_L=amd64

Run as Docker

$ sudo docker run -d --name=mx-exporter --device=/dev/dri -p 0.0.0.0:<host port>:8000 <image name>

Add "-v /var/lib/kubelet/pod-resources:/var/lib/kubelet/pod-resources" to support export pod resources.

if you want to change port and process interval settings, use this:

$ sudo docker run -d --name=mx-exporter --device=/dev/dri -p 0.0.0.0:<host port>:<http port> <image name> -p <http port> -i <interval>

if you want to use self defined counter configuration file on host, here's an example:

$ sudo docker run -d --name=mx-exporter --device=/dev/dri -p 0.0.0.0:<host port>:<http port> -v <new config file>:/work/counters.csv <image name> -c /work/counters.csv

Run as python application

$ python3 -m mx_exporter -p <port> -i <interval> -c <config_file>

• port: http listen port, default:8000
• interval: Metrics gathering interval, default:10000ms
• config_file: Metrics configuration file in CSV format, default: ./default-counters.csv

Deployment

Deploy with kubernetes

All the deployment configuration files stored in the deployment folder

1. Create Namespace

   kubectl apply -f namespace.yaml
   

2. Startup Data Exporter as Daemonset

   cd mx-data-exporter
   kubectl apply -f .
   

3. Startup Prometheus

   cd prometheus
   kubectl apply -f .
   

4. Startup Grafana

   cd grafana
   kubectl apply -f .
   

Deploy with Helm

Helm install using package cd mx-exporter/deployment/mx-exporter/helm; helm install metax-mx-exporter mx-exporter \ --create-namespace -n metax-monitor \ --set image.repository=xxxx \ --set image.tag=x.x.x \ Options:

• service.port - set container port, default is 8000
• gatherInterval - unit: ms, default is 10000