Einstein on self compiled Linux
7 Jul 2006 10:51:17 UTC
Topic 191528
(moderation:
Hello fellow-Crunchers!
I am thinking of compiling my own kernel. I wondered if this would have impact on the crunching performance of my laptop.
I can imagine that a self-compiled kernel works more efficient, so more CPU could be available for crunching (isn't it?). But would it also have a direct effect on Einstein performance? And what about additional heat generation due to more efficient CPU usage( as p.e. seen by optimizations on the app for S4)?
Many thanks,
Bert
Somnio ergo sum
Einstein on self compiled Linux
)
It won't any notable effect, neither on heat or crunching speed.
There might a slight
)
There might a slight difference if you compile boinc - I'm just in the process of trying that out now, having spent the last 24 hours installing Gentoo. What I am aiming to do is to also have as few system calls as possible - basically allow einstein to run with as few interruptions as possible :)
Gray
Thanks for the
)
Thanks for the answers!
But if I would compile my OS, then it should be optimised with respect to my hardware, and making less unneccesary calls. IMHO, not only BOINC would require less CPU cycles, but also other apps. Then, there are more CPU cycles available for the actual computation of the WU. Right? I would also conclude that the more efficient my OS would be, the more heat the CPU would generate. Possibly not a problem, but I've gotten the idea that my fan is weak..
I've compiled my own BOINC, but that was only useful for better benchmarking results. It did not actually improve the calculations as these are done by the Einstein-app.
Regards,
Bert
Somnio ergo sum
For sure compiling custom OS
)
For sure compiling custom OS kernel helps. But the gain is not that great. On a typical modern computer, OS kernel uses way less than 1% of total CPU time. I estimate that optimized OS kernel would save you 10% of it's CPU consumption yielding less than 0.1% of CPU time more available to applications. That gives you at most 1.5 minutes of CPU time per day.
It's up to the user to decide wether it's worth it. I might add that compiling a custom linux kernel on typical modern machine takes about 15 minutes of CPU time (more than 10 days of savings) depending on amount of drivers compiled. And quite some expertise on what to omit from kernel as well.
Metod ...
Just a once-off comparison
)
Just a once-off comparison between a Linux installed straight off the CD and using Kde (SUSE 10.1), and an optimised Linux (Gentoo) running Fluxbox.
I admit that Kde is not exactly ideal for a crunching box, but either way there does seem to be a speedup between the 2 thus far.
Time--Credit requested--OS
34,169.41----176.27----SUSE 10.1
33,892.35----179.19----Gentoo
Gray
PS: I mentioned the credit requested here as an indication of the WU size
RE: Just a once-off
)
There are a couple of problems with the comparision you've made:
* on gentoo you also have custom compiled system libraries such as math library - libm.so. This should contribute to speed up considerably more than custom compiled kernel as application (especially official one) spends quite some time using library functions. Time spent using kernel functions is far far shorter (basically only IO operations).
The only fair comparision would be to run application in off-line mode on a couple of WUs (different ARs) a couple of times on each of them to get significant statistical numbers (average, median, etc). This should be done both running distribution-supplied kernel as well as the custom-built one. Only then the times should be compared to see if there's any difference.
Metod ...
Ok, so compiling the kernel
)
Ok, so compiling the kernel for the improvement of crunching power does not make much sense as there are just a few IO operations...
But does it make sense to compile (part of) my distro? I'm running Debian Sarge at the moment. I haven't read info on it yet, at the moment debian.org seems to be down...
Thanks for the input!
Somnio ergo sum
RE: RE: on gentoo you
)
Even though I dismissed the comparison made by Gray (about 2.5%) as (almost) irelevant it's the best one you can get just like that. If you want to get better comparison, you'll have to follow the procedure I outlined in my previous post.
If you decide to follow that path, by all means post your findings here.
Metod ...
What I meant, was: If
)
What I meant, was:
If compilation of the kernel alone does not signifficantly improve crunching performance, would compilation of the entire (or part of) Debian Sarge generally make an improvement?
BTW, hardly five minutes gain is indeed insignifficant. Moreover, I've noticed that a WU requiring more time does not neccesarily mean more credit (I checked for results in a weekend, when the computer was not used). So your suggestion for comparision would indeed be more suited. But that can only be achieved AFTER I have recompiled Sarge. I was just curious if crunching power would improve, as an additional justification to recompile Linux.
Somnio ergo sum
Einstein comes as a static
)
Einstein comes as a static app, no? So any changes to the userland are irrelevant...