The longer a slow system exists the more unlikely you could ever make it faster, much less Googlespeed. A very interesting debate on Premature optimization ver

The longer a slow system exists the more unlikely you could ever make it faster, much less Googlespeed. A very interesting debate on Premature optimization versus Profile-guided Optimization could be had from this comment. Ideally, better algorithms and efficient data structures could be chosen at the design stage to create a more optimized solution. But does this approach always work? In my experience, it often fails in a greenfield implementation. The reason being the lack of understanding of the real performance bottlenecks. This is usually revealed only after several iterations of profiling and optimizing code (often revealing least expected spots). But if it's fairly well known problem say relative ranking of a set of hyperlinked documents, there is a lot of opportunity to study existing literature and come up with a more efficient algorithm. Even then, once the initial implementation is built, the task of making "a slow system faster" needs to be done and often a never ending process. So the conundrum is not often the need for speed being an primary criteria but on lack of familiarity with the performance bottlenecks eliminate them from the design stage itself.