Re: more perf results

Larry McVoy (
Sat, 25 Jul 1998 10:49:48 -0600

Oops, I was running an SMP kernel when I reported the last set of results.
This time I include 2.034/2.1.110 SMP/2.1.110 UP all on the same hardware.
Things look quite a bit nicer in UP land. And SMP is at least no worse
than 2.0.34 UP.

: I mentioned a while back that I had a little bit of Linux vs Irix
: performance testing. I have expanded numbers which include Sun, Linux
: 2.0 and Linux 2.1.110, as well as Irix. I'm happy to say that 2.1 isn't
: worse, but it isn't a lot better either, maybe a few percent but that
: is probably not statistically significant.
: The test is a 1700 line shell script which is a regression test for my
: source mgmt system. Think of it as the checker-inner, checker-outer
: script and you're close.
: The Sun is a 167Mhz Ultra1, pretty old machine I think. The regression
: stats are:
: Sun Ultra1@167:
: tmpfs 9.22secs user, 8.80secs sys, and 23secs elapsed
: ufs 8.70secs user, 9.68secs sys, and 61secs elapsed (!!)
: SGI R10K@175: 1.8u 8.6s 0:12
: AMD K6@300:
: 2.0.34 4.25user 3.59system 0:07.98elapsed
: 2.1.110 4.19user 3.47system 0:07.70elapsed SMP(!)
2.1.110 4.25user 2.87system 0:07.13elapsed Uniprocessor

: The interesting thing is the ratio of user to system time. We would
: expect that the user times are all pretty much similar, when normalized
: for processor performance. So this gives us a pretty good look at how
: the system is eating up CPU cycles. In the table below, the numbers are
: user/system and bigger is better - it implies that the system cost less,
: relatively speaking.
: Sun tmpfs 1.04
: ufs 0.89
: AMD ext2fs 1.21 2.0.34
ext2fs 1.21 2.1.110 SMP
ext2fs 1.48 2.1.110 UP
: SGI xfs 0.21
: Sun's doing pretty well in tmpfs but that is cheating. They are still
: doing more than 4x better than Irix in UFS, but remember that they take
: 5x as long wall clock wise. Linux, as usual, rocks :-) I can't wait
: to see Stephen's journaling changes, I think at that point, Linux will
: be the hands down winner in all dimensions, both performance and safety.

To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to
Please read the FAQ at