Databricks

jamesw · ‎01-10-2023

Setup:

custom docker container starting from the "databricksruntime/gpu-conda:cuda11" base image layer
10.4 LTS (includes Apache Spark 3.2.1, Scala 2.12)
multi-node, p3.8xlarge GPU compute

When I try to view Ganglia metrics I am met with "502 Bad Gateway":

Even after ~1hr of my compute cluster running there are no logs at all:

As a sanity check I booted another compute without a custom docker container (using 11.3 LTS ML (includes Apache Spark 3.3.0, GPU, Scala 2.12)) and the Ganglia metrics work fine.

Are there any limitations with Ganglia metrics and custom docker containers?

Also when I am using the custom docker container, I am forced to use the standard runtime (10.4 LTS) as the Machine Learning runtimes do not support custom containers (see https://docs.databricks.com/clusters/custom-containers.html#requirements).

I am thinking this could be a source of the issue too. Does the ML runtime provide any needed libraries for Ganglia to work on GPU compute?

Vivian_Wilfred · ‎01-11-2023

Hi @James W , Ganglia is not available for custom docker containers by default. This is a known limitation.

However, you can try this experimental support for ganglia in custom DCS:

https://github.com/databricks/containers/tree/master/experimental/ubuntu/ganglia

Kaniz · ‎01-16-2023

Hi @James W (Customer) , We haven’t heard from you since the last response from @Vivian Wilfred, and I was checking back to see if his suggestions helped you.

Or else, If you have any solution, please do share that with the community as it can be helpful to others.

Also, Please don't forget to click on the "Select As Best" button whenever the information provided helps resolve your question.