One of our team’s field engineers recently sent a link to a Channel 9 video: Steve Michelotti of e.magination on High Performance Web Solutions. This company built a 64-bit web server that handles over 3 billion transactions a day and guarantees a 250 ms response time. And it goes without saying that they built it on top of .NET.
The team optimized their code to avoid implicit allocations, pool and reuse objects and allocate large collections directly on the Large Object Heap. But they still saw some latencies that were greater than their contract allowed. So they worked with their field engineer and the GC team to get a new feature implemented in .NET 3.5 SP1 that let them work around the problem.
The problem comes from the size of their managed heaps. When you try to manage memory on a 16 GB server it can take a few seconds to run through the heaps. There’s not much you can do on a single machine besides keeping your heaps reasonably sized. Having millions of objects in memory is expensive. But their architecture is such that they can direct requests from the web server to any of a set of identical application servers. This means that they can take any individual server offline for the few seconds it takes to do a full collection. All they needed is to know when a GC is going to happen and when it’s completed. Full GC notifications is a new feature in 3.5 SP1 that gave e.magination necessary insight into the GC’s behavior that lets them redirect requests away from servers that are doing full GCs.
This video is only about 20 minutes long but it’s packed with great insights into how to make the most of your .NET server code. It shows the team at e.magination doing a fantastic job of analyzing and tuning their system for optimum performance. I found it interesting and inspiring. I hope you do too.