Scalable Network Programming
The fork-and-do-something latency on my notebook on Linux 2.6 is 200
microseconds. That means my notebook can create 5.000 processes per second. Thus my notebook can handle about 13 billion forks per month. My Athlon XP 2000+ desktop can do 10.000 processes per second, or 26 billion per month. Heise Online, the biggest German site, had 118 million page impressions in September.
Scheduling
Why does the fork benchmark include writing to the pipe? Because having many processes not only makes creating more processes more difficult, it also makes choosing which process to run more difficult. The part of the operating system that chooses which process to run next is called the scheduler. A typical workload is having two dozen processes, with one or two of them actually having something to do (they are runnable).