Sven-Thorsten Dietrich wrote:
I am too looking forward to seeing results against the >= 07.48 RT
kernels incorporating Daniel's recent IRQ disable relief.
Indeed, this is on our list.
I think the comparison should absolutely compare identical community
kernels. The comparison between two different release candidates is
questionable. rc2 to rc4 doesn't seem like much, after all, how much
code could go into a release candidate. (diff | wc -l)
Also, I question testing -rc code in the first place, except for
regression purposes.
On this issue, it has to be said that I don't think any set of test
results will suffice on its own as a definitive benchmark. There will
be a need for continued testing and publication of said results, which
we hope others will take part in when we publish the framework we've put
together.
Finally, there are other big-picture issues. How hard is it to maintain
the code in general? At the risk of ruffling feathers, forward-porting
RT code (or backporting) it a few revisions of rc's isn't too bad.
At Ingo's pace, we have all done some of that.
How does that effort compare for porting ADEOS code? If several weeks of
work are invested in a comparison of rc2 to rc4, how much additional
work is needed to bring Adeos up to the base for the current RT kernel?
Philippe can correct me if I'm wrong, but adeos maintenance is not that
difficult. However, it has to be said that up until now, Philippe has been
the main driving force behind adeos. So while he's been fairly good at
publishing patches for as recent a kernel as possible, the manpower behind
PREEMPT_RT is obviously larger. That, though, only requires interested
parties to participate for it to change. Again, the adeos patch isn't
that big.
In addition, I think the discrepancy between the vanilla kernel and the
RT kernel could be much greater, if the workload specifically, or even
coincidentally exercised bottlenecks.
If you've got any specific test run suggestions, we'll gladly take them.
Karim
