1 |
Justin posted on Wed, 14 Apr 2010 08:46:15 +0200 as excerpted: |
2 |
|
3 |
> On 14/04/10 04:12, Zac Medico wrote: |
4 |
>> Hi everyone, |
5 |
>> |
6 |
>> Should we add a RESTRICT=parallel value for ebuilds that can't be built |
7 |
>> at the same time as other ebuilds? Brian says we need it for things |
8 |
>> like xorg-server which calls eselect opengl. |
9 |
> |
10 |
> There is at least one other example which benefits from singlular build, |
11 |
> atlas libs. They run a benchmark suite to create platform specific |
12 |
> headers, which is heavily influenced by the system load. So having |
13 |
> RESTRICT=parallel would make the emerge more reliable. |
14 |
|
15 |
sci-libs/blas-atlas (and perhaps lapack-atlas, etc, too), right? |
16 |
"Automatically tuned..." |
17 |
|
18 |
Wow! Yeah, that sounds like a reasonable example. Sort of like the |
19 |
kernel does for md/RAID-5 and 6 at boot, I'd guess, choosing the fastest |
20 |
algorithm on the platform, only they're doing it during system runtime |
21 |
when who /knows/ what else is running! Having a second highly |
22 |
parallelizable (MAKEOPTS version) build go from config to build and its |
23 |
load go from say .8 to 8. in the middle of those benchmarks /could/ screw |
24 |
things up "just a little!" |
25 |
|
26 |
Thanks. That's just the sort of additional practical example I was asking |
27 |
for to try to get my mind around this. Excellent example as, unlike the |
28 |
various xorg/mesa/drivers thing, it's pretty hard to argue "just code |
29 |
around it", for this one. The only technical way out of it here would |
30 |
seem to be to change the build-and-benchmark strategy itself, which would |
31 |
rather defeat the "automatically tuned" bit entirely. |
32 |
|
33 |
BTW, gcc seems to do some stage output comparing in its bootstrap |
34 |
process. Is that all absolute code correctness, or is there some |
35 |
performance benchmarking there that could benefit from this as well? |
36 |
|
37 |
-- |
38 |
Duncan - List replies preferred. No HTML msgs. |
39 |
"Every nonfree program has a lord, a master -- |
40 |
and if you use the program, he is your master." Richard Stallman |