Less slow versus faster

I describe my work as making things less slow rather than making things faster. While making something less slow tends to make it faster I think there are two different tasks here and both are important:

• By making things faster I mean a faster response time and more throughput for a single-threaded workload. For example, cache-friendly memory allocation might increase throughput from 100 QPS to 125 QPS and decrease response time from 10ms to 8ms.
• By making things less slow I mean that the throughput gap - the difference between linear scaling and actual throughput - has been reduced. Assume the server can do 100 QPS at one thread and with linear scaling it can do 1000 QPS at 10 threads. But linear scaling is hard to achieve and the server might be limited to 400 QPS on a real workload. Eventually someone has time for performance debugging and finds things to make better with PMP and we get 500 QPS at 10 threads. This is making something less slow.

The USL provides a formal model to reason about making software less slow for concurrent workloads. I learned about Dr. Gunther's work on the USL thanks to Baron.