From: torvalds@klaava.Helsinki.FI (Linus Benedict Torvalds) Subject: Re: Benchmarking under Linux (was Re: New 486 Suggestions? Date: 4 Sep 1992 10:07:06 GMT
In article <1992Aug31.210041.21832@novell.com> bboerner@novell.com (Brendan B. Boerner) writes:
>
>This has me wondering: does anyone have any benchmarking or performance
>metric programs for Linux? I've got a 386/16Mhz w/a tad under 5MB of
>RAM and it takes me somewhere between 3-4hours to build the kernel. On
>my girlfriend's machine, a 486/33 w/8MB RAM, it takes under 10
>minutes. I'd like to get an idea of where I could improve
>performance.
RAM space is the first thing to look for: gcc eats memory like mad, and
swapping is slow especially with the older kernels. I hope my changes
to the swap-algorithms in 0.97.pl3 will help people with slow machines:
it does seem to help even on my 8MB machine (kernel compilation time
when running X+some xterms fell from 18 minutes to 15). The main change
was to use a page-used counter instead of just a swp bitmap, which
allows fork() to be speeded up /a lot/ when swapping.
But processor speed can be very important under linux: not just for the
obvious user-level speedup, but due to better response to disk-drive
interrupts and the like. Faster machines may simply read the disk at
the full 1:1 interleave - with slower systems, it's possible that the HD
driver doesn't keep up, and you get the dreaded 1-block/rotation
syndrome, which really hurts when swapping. This problem is probably
especially notable on 386SX machines: the 386 interrupt handling is
inherently slow, but if the memory badwidth is further reduced by the
16-bit bus, interrupt response is probably ever worse.
Linus