svn commit: r253002 - head

[Peter Wemm]

On Mon, Jul 8, 2013 at 5:12 PM, Teske, Devin <Devin.Teske at fisglobal.com> wrote:
> On Jul 7, 2013, at 2:15 PM, Alfred Perlstein wrote:
>
>> On 7/7/13 2:01 PM, Garrett Cooper wrote:
>>> Why the magic number 12?
>>
>> Numbers higher seem to result in worse performance as reported by some members of my team.
>>
>
> We've run as high as -j48 ... number of logical cpu's "times-two" (it was on a box with 24 logical CPUs).

There's an additional constraint.  -j48 and 48 running instances of
clang++ and its memory demands when compiling C++ code each starts to
get non-trivial.

> We did buildworld in under 9 minutes. Beyond the "times-two" rule we saw a slow-down.
>
> NOTE: I might also mention this was on RAID-1 SSD.

On the cluster we get about 9 minutes on RELENG_9 with WITHOUT_CLANG.
About 11 minutes including the kernel.  Boring old disks with -j24.

I'd be pleasantly surprised if you were doing HEAD in < 9 minutes
given that it frequently fails to build with more than -j8 or -j10.
(I've heard this might be fixed.. but I'm skeptical as I had a -j2
race last week)

But we are talking about a commit giving advice for building kernels
here, right?

>>>> Log:
>>>>  Document tip on how to build all kernels quickly.

I used to use -j1000 to stress test context switching with a LINT
build - that was when we ran into FD_SETSIZE vs make.  Kernel builds
have always been highly tolerant of  very large -j values because
idiots like me used to use it to try and break things.

-- 
Peter Wemm - peter at wemm.org; peter at FreeBSD.org; peter at yahoo-inc.com; KI6FJV
UTF-8: So you can \342\200\231 .. for when a ' just won't do
<brueffer> ZFS must be the bacon of file systems. "everything's better with ZFS"