Abstract

As commodity computers become more powerful with ever increasing amounts of RAM and now 4 CPU cores as standard (with further increases expected in the near future), is it the I/O subsystem that is the bottleneck? And if it is, then given the abundance of RAM and CPU power available could we use them to help alleviate the problem?

When writing Linux programs today programmers are faced with a plethora of choices of how to do I/O, and the “wrong” choice could mean their program takes twice as long to execute. In this talk I present results from experimenting with the Linux I/O APIs and show that performance can differ substantially depending on which and how the API is used. Clearly educating the programmers and re-writing all the existing code would be a solution, but with the abundance of RAM and CPU power available could other techniques allow the programmer to always use the same API , but with the kernel giving it the performance equal to if it had used the “fastest” API .