: O(1) 2.4.17-J7 Tuneable Parameters
Quoted 8 lines
[PATCH]: O(1) 2.4.17-J7 Tuneable Parameters
Hi Ingo,
This patch synchronizes with J7 and I think makes the changes
you wished. A couple of important points:
Quoted 9 lines
- This patch can be applied to EITHER 2.4.17 OR 2.4.18 pre 7 as long as Ingo's J7 patch is applied first.
- While I agree with you on not wanting these in the mainline kernel, I ran Hackbench on one of our new Foster systems with and without the tuneable parameters, and while the numbers do degrade slightly, its rather suprisingly small. So dont be afraid to use this as a system tuning aid.
How big is the actual degradation in your test? IIR, Ingo is afraid
that the tunables could easily screw things up, which of course is true. What about adding a kernel-build option that leaves the sysctl interface read-only by default and enables writing only if it is requested at build time?
That way the external interface stays constant.
Martin
Signature · 5 lines
------------------------------------------------------------------ Martin Knoblauch | email: [email protected] TeraPort GmbH | Phone: +49-89-510857-309 C+ITS | Fax: +49-89-510857-111 http://www.teraport.de | Mobile: +49-170-4904759
Quoted 17 lines · Martin Knoblauch
Quoted 8 lines
[PATCH]: O(1) 2.4.17-J7 Tuneable Parameters
Hi Ingo,
This patch synchronizes with J7 and I think makes the changes
you wished. A couple of important points:Quoted 8 lines
- This patch can be applied to EITHER 2.4.17 OR 2.4.18
pre 7 as long as Ingo's J7 patch is applied first.- While I agree with you on not wanting these in the
mainline kernel, I ran Hackbench on one of our new Foster systems with and without the tuneable parameters, and while the numbers do degrade slightly, its rather suprisingly small. So dont be afraid to use this as a system tuning aid.
Quoted 5 lines
How big is the actual degradation in your test? IIR, Ingo is afraid
that the tunables could easily screw things up, which of course is true. What about adding a kernel-build option that leaves the sysctl interface read-only by default and enables writing only if it is requested at build time?
Nah, as long as the permissions say only root can write to it, you're pretty much covered. (It's certainly no worse than root going cat /dev/zero > /dev/hda...)
Rob
Quoted 8 lines · Martin Knoblauch
Quoted 1 line
[PATCH]: O(1) 2.4.17-J7 Tuneable Parameters
How big is the actual degradation in your test? IIR, Ingo is afraid
that the tunables could easily screw things up, which of course is true. What about adding a kernel-build option that leaves the sysctl interface read-only by default and enables writing only if it is requested at build time?
Running on a machine that I dont think I can really officially give numbers.. However, lets say that without the tuneable code you got a run of hackbench doing 60 groups that took 8.27 secs, when the tuneable code is in it went to a whopping 8.6 secs :)
The results at least on this benchmark were all in that decimal noise.
As for a build option, if the code were integrated I might see that as making sense, but as this is a developmental patch the user is expected to know what they are doing. Only root can write anything to the parameters as well.
Cheers,
Signature · 4 lines
Jack F. Vogel IBM Linux Solutions [email protected] (work) [email protected] (home)