Improving Performance in <128kB File Workloads

We all know that small files can be a bit of an Achilles' heel on the Isilon platform for both capacity overhead and performance reasons. I had a customer in the healthcare space using Isilon to support their imaging platform complaining about performance for Cache 1 and Cache 2 on McKesson PACs. It was learned that for every image taken there were several thousand small metadata files created to support it. This meant that the bigger image would not suffer from the capacity overhead but obviously the thousands of smaller files would recognize the overhead.

They were well aware of the capacity overhead but were having performance issues. We decided to try some options and thinking that the mixed workload of file sizes was a potential impact we decided to try setting the protection policy to 3x thinking that OneFS trying to calculate parity and getting a mixed messages was a potential cause.

In this our assumptions were correct. Mirroring rather than leveraging 2:1 netted us over 30% performance increase on the workload (measured in terms of time to write a case to the cluster)

However, they didn't want to suffer the capacity penalty on their larger files. What we did was use SmartPools to change the directory's protection policy to 2:1 on a schedule. This meant data coming in wasn't impacted by the parity algorithm and we reclaim the space in the off hours when performance isn't a priority.

I've recommended this to several customers with good success to date and Jason Sturgeon was kind enough to write up a paper explaining the process.