From a fundamental perspective, performance and scalability are orthogonal. Scalability in the realm of technical infrastructure simply means that it can grow and shrink without fundamental change. Growing, of course, means serving more traffic and/or more users, and that directly relates to the performance of the system as a whole.
If you have two load-balanced web servers that serve static HTML pages only, this is an inherently scalable architecture. More web servers can be added, and capacity can be increased without redesigning the system or changing dependent components (because no dependent components are in the system).
Even if your web servers can serve only one hit per second, the system scales. If you need to serve 10,000 hits per second, you can simply deploy 10,000 servers to solve your performance needs. This scales, but by no means scales well.
It may be obvious that one hit per second is terribly low. If you had used 500 hits per second as the performance of a single web server, you would only have needed 20 machines. Herein lies the painful relativity of the term scales wellserving 500 hits per second of static content is still an underachievement, and thus 20 machines is an unacceptably large number of machines for this task.
It should be clear from the preceding example that the performance of an individual component in the architecture can drastically affect how efficiently a system can scale. It is imperative that the performance of every introduced architectural component is scrutinized and judged. If a component performs the needed task but does not scale and scale well, using it will damage the scalability of the entire architecture.
Why be concerned with the performance of individual components? The only way to increase the performance of a complex system is to reduce the resource consumption of one or more of its individual components. Contrapositively, if an individual component of a complex system performs slowly, it is likely to capsize the entire architecture. It is fundamental that solid performance-tuning strategies be employed through the entire architecture.
Every architecture has components; every component runs software of some type. If your architecture has performance problems, it is usually obvious in which component the problems are manifesting themselves. From that point, you look at the code running on that component to find the problem. A few common scenarios contribute to the production of slow code:
Given that there is no magical solution, how does one go about writing high-performance code? This is an important question, and there is a tremendous amount of literature on the market about how to optimize code under just about every circumstance imaginable. Because this book doesn't focus on how to write high-performance code, we will jump to how to diagnose poorly performing code.
Gene Ahdmal stated that speeding up code inside a bottleneck has a larger impact on software performance than does speeding up code outside a bottleneck. This combined with classic 90/10 principle of code (90% of execution time is spent in 10% of the code) results in a good target.
Do not choose the slowest architectural component or piece of code to focus on. Start with the most common execution path and evaluate its impact on the system. The thing to keep in mind is that a 50% speedup of code that executes 0.1% of the time results in an overall speedup of 0.05%, which is small. On the other hand, a 50% speedup of code that executes 5% of the time results in an overall speedup of 2.5%, which is significant.
At the end of the day, week, month, or year, there will be code that is bad. Even the best developers write bad code at times. It is important that all infrastructure code and application code be open to review and revision and that performance review and tuning is a perpetual cycle.
I honestly believe the most valuable lessons in performance tuning, whether it be on the systems level or in application development, come from building things wrong. Reading about or being shown by example how to do a task "correctly" lacks the problem-solving skills that lead to its "correctness." It also does not present in its full form the contrast between the original and the new.
By working directly on an application that had performance issues, you can work toward improvement. Realizing performance gains due to minor modifications or large refactoring has tremendous personal gratification, but there is more to it than that. The process teaches the analytical thought processes required to anticipate future problems before they manifest themselves.
Sometimes performance tuning must be "out of the box." Analysis on the microscopic level should regularly be retired to more macroscopic views. This multire-solutioned problem analysis can turn a question such as "How can I merge all these log files faster?" into "Why do I have all these log files to merge and is there a better way?" Or a question such as "How can I make this set of problematic database queries faster?" becomes "Why am I putting this information in a database?"
Changing the scope of the question allows problems to be tackled from different angles, and regular reassessment provides an opportunity to use the right tool for the job.