In the last episode (May 06), Sasha Pachev said:
>
> # Linux 2.0.33 www.sashanet.com i586 : Pentium 166 32 MB EDO RAM
> Static parent/Dynamic child Avg fork() + execl() time = 8.36 ms
> Static parent/Static child Avg fork() + execl() time = 1.95 ms
> Dynamic parent/Dynamic child Avg fork() + execl() time = 8.80 ms
> Dynamic parent/Static child Avg fork() + execl() time = 2.48 ms
>
> # Linux 2.0.36 www.isshost.com i686 : Pentium II 350 128 MB SDRAM
> Static parent/Dynamic child Avg fork() + execl() time = 2.31 ms
> Static parent/Static child Avg fork() + execl() time = 0.43 ms
> Dynamic parent/Dynamic child Avg fork() + execl() time = 2.45 ms
> Dynamic parent/Static child Avg fork() + execl() time = 0.55 ms
# FreeBSD 4.0-CURRENT micron.danhome.emsphone.com i386 : P/166 32MB EDA RAM
Static parent/Dynamic child Avg fork() + execl() time = 6.59 ms v=6.19
Static parent/Static child Avg fork() + execl() time = 1.41 ms v=0.96
Dynamic parent/Dynamic child Avg fork() + execl() time = 7.31 ms v=6.21
Dynamic parent/Static child Avg fork() + execl() time = 2.07 ms v=0.96
# FreeBSD 2.2.8-STABLE emssrv5.emsphone.com i386 : PII/400 128MB SDRAM
Static parent/Dynamic child Avg fork() + execl() time = 2.38 ms
Static parent/Static child Avg fork() + execl() time = 0.63 ms
Dynamic parent/Dynamic child Avg fork() + execl() time = 2.68 ms
Dynamic parent/Static child Avg fork() + execl() time = 0.87 ms
# SunOS 5.6 pd1.emsphone.com sun4u : Sparc/400Mhz x6 CPUs, 6gb RAM
Static parent/Dynamic child Avg fork() + execl() time = 3.14 ms v=2.94
Static parent/Static child Avg fork() + execl() time = 0.88 ms v=0.69
Dynamic parent/Dynamic child Avg fork() + execl() time = 3.68 ms v=2.95
Dynamic parent/Static child Avg fork() + execl() time = 1.41 ms v=0.65
> I'd be interested to see what you get on others. From the results I
> got, it seems that unless your server(Intel hardware/Linux) gets hit
> in the order of magnitude of 1000 times per second, the overhead of
> fork() + execve() is relatively insignificant. Correct me if my
> interpretation of the results is in error.
Of cource, fork() overhead varies on the size of the process, and is
not the best choice of you're going to do an exec() immediately. Why
copy your page tables thwn you're going to throw them away in the next
syscall? The "v" columns are the same program, run with vfork()
instead of fork().
-Dan Nelson
dnelson@stripped
