lmbench numbers
lmbench numbers
- Subject: lmbench numbers
- From: email@hidden (Rob Latham)
- Date: Mon, 3 Dec 2001 00:14:01 -0500
nothing like an "objective" os comparison to liven up a relatively young
list :>
lmbench is a suite of microbenchmarks designed to measure various os
functions. it's pretty well-accepted as a good way to profile an
operating system. ( more info:
http://www.bitmover.com/lmbench )
i see from the archives that lmbench numbers were last posted in may.
hope i'm not revisiting the topic too soon.
These numbers came from a g4 400 powerbook, with netbsd, linux, and os x
installed. For kicks, you can see how results vary with several recent
OS X updates.
netbsd-powerpc is basically the GENERIC kernel, recompiled to not have
lots of drivers for hardware not present on the laptop.
linux is from the "benh" tree: it's got better support for the hardware,
but as far as core os features should be similar to a kernel.org kernel.
I've got 2 or three runs for each kernel, just in case there's an
outlying freak result for one.
An empty slot for a test means that for some reason the test was not
run. (in the case of netbsd, i disabled the rpc tests because they just
hung and never completed. the other no-results are mysteries to me)
i would classify the load going on while these tests were running as
"normal idle workstation" ( e.g. mozilla was loaded, but i wasn't
actively browsing the web )
Darwin, or more likely the mach sublayer, has significantly higher
overhead for several classes of operating system functions ( most
notably, context switch, file latency, and VM latency ). would the
darwin kernel hackers mind elaborating on some of the reasons why?
Thanks
==rob
--
Rob Latham Allentown, PA USA
EAE8 DE90 85BB 526F 3181 1FCF 51C4 B6CB 08CC 0897
--
Rob Latham Allentown, PA USA
EAE8 DE90 85BB 526F 3181 1FCF 51C4 B6CB 08CC 0897
Subject: lmbench numbers
lmbench is a suite of microbenchmarks designed to measure various os
functions. it's pretty well-accepted as a good way to profile an
operating system. ( more info:
http://www.bitmover.com/lmbench )
i see from the archives that lmbench numbers were last posted in may.
hope i'm not revisiting the topic too soon.
These numbers came from a g4 400 powerbook, with netbsd, linux, and os x
installed. For kicks, you can see how results vary with several recent
OS X updates.
netbsd-powerpc is basically the GENERIC kernel, recompiled to not have
lots of drivers for hardware not present on the laptop.
linux is from the "benh" tree: it's got better support for the hardware,
but as far as core os features should be similar to a kernel.org kernel.
I've got 2 or three runs for each kernel, just in case there's an
outlying freak result for one.
An empty slot for a test means that for some reason the test was not
run. (in the case of netbsd, i disabled the rpc tests because they just
hung and never completed. the other no-results are mysteries to me)
i would classify the load going on while these tests were running as
"normal idle workstation" ( e.g. mozilla was loaded, but i wasn't
actively browsing the web )
Darwin, or more likely the mach sublayer, has significantly higher
overhead for several classes of operating system functions ( most
notably, context switch, file latency, and VM latency ). would the
darwin kernel hackers mind elaborating on some of the reasons why?
Thanks
==rob
--
Rob Latham Allentown, PA USA
EAE8 DE90 85BB 526F 3181 1FCF 51C4 B6CB 08CC 0897