Hi all,
Does anyone have any recent SPEC 2000 benchmarks comparing Linux to L4Linux running on Fiasco? I'm getting surprisingly high results running on a Pentium IV 2.53 GHz forcing memory to be 256M in both systems. I haven't even tweaked all the performance critical options in L4 to the "off" state yet.
Thanks!
Julian
On Sat, 28 Jan 2006 19:02:19 -0500, Julian Grizzard grizzard@ece.gatech.edu said:
JG> Hi all, JG> Does anyone have any recent SPEC 2000 benchmarks comparing Linux to JG> L4Linux running on Fiasco? I'm getting surprisingly high results JG> running on a Pentium IV 2.53 GHz forcing memory to be 256M in both JG> systems. I haven't even tweaked all the performance critical options in JG> L4 to the "off" state yet.
What exactly do you expect to see when running SPEC CPU benchmarks? They do almost no syscalls, and would hardly swap pages, so on a processor with a hardware-walked page table, the underlying OS is pretty much irrelevant.
Gernot
Hi Gernot,
Right, so it makes sense that the benchmarks should be close. The surprising results to me are that in some cases, the SPEC programs even performed better under L4Linux than Linux.
I'm just looking for some comparison to other people to make sure that my experiments are close to others. Also, there is some mention of testing operating system performance in the SPEC documents, although the benchmark does not seem to be focused on that. Does anyone have any suggestions for the best way to compare Linux to L4Linux then?
Thanks!
Julian
Gernot Heiser wrote:
On Sat, 28 Jan 2006 19:02:19 -0500, Julian Grizzard grizzard@ece.gatech.edu said:
JG> Hi all, JG> Does anyone have any recent SPEC 2000 benchmarks comparing Linux to JG> L4Linux running on Fiasco? I'm getting surprisingly high results JG> running on a Pentium IV 2.53 GHz forcing memory to be 256M in both JG> systems. I haven't even tweaked all the performance critical options in JG> L4 to the "off" state yet.
What exactly do you expect to see when running SPEC CPU benchmarks? They do almost no syscalls, and would hardly swap pages, so on a processor with a hardware-walked page table, the underlying OS is pretty much irrelevant.
Gernot
On Sun Jan 29, 2006 at 09:40:05 -0500, Julian Grizzard wrote:
Hi Gernot,
Right, so it makes sense that the benchmarks should be close. The surprising results to me are that in some cases, the SPEC programs even performed better under L4Linux than Linux.
I'm just looking for some comparison to other people to make sure that my experiments are close to others. Also, there is some mention of testing operating system performance in the SPEC documents, although the benchmark does not seem to be focused on that. Does anyone have any suggestions for the best way to compare Linux to L4Linux then?
Well, it really depends on what workload you want to test. lmbench is often used for measuring system call overheads.
SpecWEB, or the traditional "kernel compile" benchmark, are also often used as application benchmarks.
HTH,
Benno
On Mon, 30 Jan 2006 07:41:09 +1100, Benno benjl@cse.unsw.edu.au said:
B> On Sun Jan 29, 2006 at 09:40:05 -0500, Julian Grizzard wrote:
Hi Gernot,
Right, so it makes sense that the benchmarks should be close. The surprising results to me are that in some cases, the SPEC programs even performed better under L4Linux than Linux.
I'm just looking for some comparison to other people to make sure that my experiments are close to others. Also, there is some mention of testing operating system performance in the SPEC documents, although the benchmark does not seem to be focused on that. Does anyone have any suggestions for the best way to compare Linux to L4Linux then?
B> Well, it really depends on what workload you want to test. B> lmbench is often used for measuring system call overheads.
B> SpecWEB, or the traditional "kernel compile" benchmark, are also B> often used as application benchmarks.
AIM (or reaim).
Gernot
Benno, Gernot, and All,
Here's the benchmarks I will try to use then:
lmbench kernel compile time AIM SPEC CPU 2000 SpecWEB*
*Won't be able to use this one in the near future as I don't have immediate access to it.
Thanks,
Julian
Gernot Heiser wrote:
On Mon, 30 Jan 2006 07:41:09 +1100, Benno benjl@cse.unsw.edu.au said:
B> On Sun Jan 29, 2006 at 09:40:05 -0500, Julian Grizzard wrote:
Hi Gernot,
Right, so it makes sense that the benchmarks should be close. The surprising results to me are that in some cases, the SPEC programs even performed better under L4Linux than Linux.
I'm just looking for some comparison to other people to make sure that my experiments are close to others. Also, there is some mention of testing operating system performance in the SPEC documents, although the benchmark does not seem to be focused on that. Does anyone have any suggestions for the best way to compare Linux to L4Linux then?
B> Well, it really depends on what workload you want to test. B> lmbench is often used for measuring system call overheads.
B> SpecWEB, or the traditional "kernel compile" benchmark, are also B> often used as application benchmarks.
AIM (or reaim).
Gernot
l4-hackers@os.inf.tu-dresden.de