Timings for optimised poll(2)

Richard Gooch (rgooch@atnf.CSIRO.AU)
Sat, 23 Aug 1997 22:13:05 +1000

Hi, all. To follow on from my previous message where I suggested
that the poll() syscall checks a "poll_events" field in the "struct
file" structure rather than calling the indirect poll function, I
hacked up a fast_poll() syscall, rebooted and did some timings. For a
zero timeout poll I get the following results:

old_poll: 4 900 microseconds (virgin 2.1.51)
new_poll: 2 900 microseconds (2.1.51 + no wait queue manipulation)
fast_poll: 850 microseconds (poll_events field processing)

Timings for 1021 file descriptors on a Pentium 100. These numbers look
quite promising!

Regards,

Richard....