Case study:

Solving build configuration mistakes

The problem:

When a software build is run in parallel on multiple machines the file system gets overloaded and the scaling performance drops.

Breeze *Trace - ARM

The solution:

By profiling the system with Breeze, it is possible to see how and where the build is storing tmp data.

In this case, Breeze showed that ARM’s build was storing tmp data in /scratch space on a shared disk, instead of in the local /tmp dir. The solution was to identify the scripts that are setting up the environment and reconfigure them to use local disk.

The result:

The scaling performance of the build improved significantly. By putting less load on the file system all the jobs on the cluster ran faster.