The way applications access data has never been more important. Bad I/O patterns can limit application performance and can overload shared storage. In this session I will be describing a case study where a live monitoring system was used to locate applications with bad I/O patterns. The customer is a large software vendor with a compute cluster that runs their CI framework. By combining lightweight analytics for the dev ops team with a more rigorous framework that users could opt-in to, we delivered an I/O profiling-as-a-service framework that eliminated storage bottlenecks. We are now working with them to integrate I/O profiling capabilities with their tests to prevent problems returning.
Required audience experience
A high-level understanding of compute clusters or distributed continuous integration frameworks.
Objective of the talk
I/O profiling has previously been a specialist skill performed as part of an optimisation step, but bad I/O is increasingly being a point of failure in distributed systems. In this session we show how big speed ups can be achieved without specific knowledge in the field.