22 Feb
2012
22 Feb
'12
1:16 a.m.
On 21/02/12 2:31 PM, Sean Conner wrote:
That seems more plausible than the caching suggestion.
It was thus said that the Great Toby Thain once
stated:
I'm aware of that. :)
It is a good test of correctness. Every makefile I write is expected to
pass it.
On 21/02/12 1:23 PM, Dave McGuire wrote:
If a "make clean ; make -j" breaks, then the Makefile isn't correct 8-)
On 02/21/2012 08:30 AM, Toby Thain wrote:
Yes, I know how to write correct Makefiles. > Not an embedded system, but at work we have
access to a multicore SPARC
> system (Sun, I don't recall the model since it's actually stashed away
> in a
> data center) with 8 cores. Doing a parallel make (it helps to have a
> properly written makefile; I went to the trouble to do so for the part
> I'm
> responsible for) only takes 1/10th the time of a non-parallel make.
Impressive. I've yet to see a *super*linear speedup for parallel make,
myself.
--Toby (uses 8 core at work)
I've *only* ever seen a superlinear speedup on parallel builds. Are your
Makefiles ok? By superlinear I mean, taking less time than the
serialised process
divided by the number of cores, which is what Sean is seeing. I am not
sure I've ever seen this. Close to 1/N, sure. But 0.7/N?
The directory I posted figures for has 389 files to compile (C++). It
wouldn't surprise me that there are a few files that take a while to
compile, while in the same time, other cores could compile a multiple number
of files. I'm sure there's an answer in queueing theory (it's faster for
one line of customers to N tellers than one line for one teller).
In another section of the codebase, there's a
subsection that takes around
1:30 (one minute, 30 seconds) to compile serially, and 1:15 parallel, but
that's because there's one file that's a few megabytes in size [1].
-spc (I also avoid recursive makes)
Right. They lead to incomplete graphs, and failures. But we all know
that by now...
--Toby
[1] I can't be sure of certain data files being install where the
program runs, so it's easier to just embed said data files into the
executable. This is purely for testing purposes, not release.