How much does the computer language influence the speed of software?
A great rebuke to those who would suggest that software will always run faster if it is written in C. A good reminder that there are many things that influence the speed of software, and efficient use of CPU cycles is only one.
I was recently drawn into another discussion about a claim that project Foo was faster than project Bar because Foo is written in C (or maybe C++) and Bar is written in Java. In my experience, as a long-time kernel programmer and as someone who often codes in C even when there are almost certainly better choices, such claims are practically always false. The speed at which a particular piece of code executes only has a significant effect if your program can find something else to do after that piece is done – in other words, if your program is CPU-bound and/or well parallelized. Most programs are neither. The great majority of programs fit into one or more of the following categories.
I/O-bound. Completing a unit of work earlier just means waiting longer for the next block/message.
Memory-bound. Completing a unit of work earlier just means more time spent thrashing the virtual-memory system.
Synchronization-bound (i.e. non-parallel). Completing a unit of work earlier just means waiting longer for another thread to release a lock or signal an event – and for the subsequent context switch.
Algorithm-bound. There’s plenty of other work to do, and the program can get to it immediately, but it’s wasted work because a better algorithm would have avoided it altogether. We did all learn in school why better algorithms matter more than micro-optimization, didn’t we?
If you look at this excellent list of performance problems based on real-world observation, you’ll see that most of the problems mentioned (except #5) fit this characterization and wouldn’t be solved by using a different language. It’s possible to run many synchronization-bound programs on one piece of hardware, with or without virtualization, but the fewer resources these programs share the more likely it becomes that you’ll just become memory-bound instead. On the flip side, if a program is purely disk-bound or memory-bound then you can obtain more of those resources by distributing work across many machines, but if you don’t know how to implement distributed systems well you’ll probably just become network-bound or synchronization-bound. In fact, the class of programs that exhibit high sensitivity to network latency – a combination of I/O-boundedness and synchronization-boundedness – is large and growing.