Software developers tend to only focus on application performance when they are near code completion and the business users have complained that the application is too slow. Even if the application meets all of the requirements, it will receive a large amount of criticism and may potentially be scrapped. Sometimes the proposed solution is to add more hardware, but a system that can't handle 3 concurrent users is not going to scale effectively to 3000 users by multiplying the hardware by 1000 times.

Unfortunately, when the team finally starts to look at performance problems, the project delivery deadline has usually passed, so there is enormous pressure to look for the silver bullet fix. This hunt may go on for months with no results.

Only by carefully working through a performance improvement methodology will the project gain any real results. Briefly, these are the steps that we follow:

1. Determine the metrics to be used, such as pages/sec, transactions/sec, messages/sec, etc.
2. Agree on the desired targets for the metrics.
3. Estimate the theoretical performance the current infrastructure should be able to produce.
4. Identify a representative, measurable and reproducible workload to be used during testing.
5. Baseline the current system.
6. Measure disk i/o, processor load, memory usage, network usage, etc.
7. Look for bottlenecks and track them back to the code.
8. Re-engineer the code to remove the bottlenecks.
9. Re-baseline to measure improvements. Repeat the cycle until the targets are reached.

We've found that many small improvements will eventually get the desired results.

We do our best to leverage the existing work as much as possible instead of the common practice of throwing everything away and starting over from scratch.