On Tue, Jun 21, 2005 at 02:17:59PM -0400, Malita, Florin wrote:
On Tue, 2005-06-21 at 11:47 -0400, Dave Jones wrote:
> Which is an apples to oranges comparison. 2.6.11-1.1369_FC4 is
> actually based on 2.6.12rc6. I'll be interested in seeing results
> rerun against this kernel.
Back with some apples:
http://lufs.sourceforge.net/unixbench.html
Now I have:
1. Linux 2.6.12-rc6 (nodebug+p4+nose+nohm+lean): 355.7
2. Linux 2.6.11.12 (nodebug+p4+nose+nohm+lean): 345.8
3. Linux 2.6.11-1.1369_FC4 (nodebug+p4+nose+nohm+lean): 269.3
4. Linux 2.6.11-1.1369_FC4 (nodebug+p4+nose+ nohm): 253.1
5. Linux 2.6.11-1.1369_FC4 (nodebug+p4): 239.4
6. Linux 2.6.11-1.1369_FC4 (nodebug): 236.7
7: Linux 2.6.11-1.1369_FC4 (orig): 213.2
8: SunOS 5.11 (orig): 122.3
(1, 2 & 3 here share the same configuration)
So 2.6.12-rc6 is slightly better overall than 2.6.11.12 and still a lot
faster than FC4 (especially in the syscall overhead & pipe throughput
area).
Thanks, I'll take a look at this later.
Am I correct in assuming the FC kernel doesn't use the
vsyscall/sysenter
mechanism thus taking a serious performance hit on P4s?
Correct. (Unless you have a CPU with NX).
I think we'd also be able to reenable sysenter if we booted with exec_sheild=0,
but currently we don't handle that case (we just always disable if no NX present)
Dave