On 07/17/10 12:32 PM, William Stein wrote:
On Sat, Jul 17, 2010 at 1:16 PM, Robert Miller<rlmills...@gmail.com> wrote:
With the rest of the packages, we just do
export MAKE="make -j6"
to enable parallel building. Why don't we just use this in atlas as
well? Why do we need to make a special case?
I think the OP was suggesting building ATLAS so that when it is later
run it can use multiple cores.
The above MAKE option merely impacts build time, which is an orthogonal problem.
I'm -1 to adding too many environment variables that impact the built
Sage, because it makes testing sage much, much more difficult, since
it leads to a combinatorial explosion in the number of different sage
versions we should test.
-- William
So what do you suggest with ATLAS?
* Always build with no threading, as is done now?
* Always build so that ATLAS choses the number of threads automatically? That
might be an issue on systems like t2, where that would be 128, or even on some
of your machines with 24 cores.
My concerns are
1) Building ATLAS with no support for threads will be slow.
2) Building ATLAS with multiple threads is likely to be more buggy, as
experience tells me that its very hard to create code that works multi-threaded
on lots of platforms.
I think the present situation (building without threads) is a pretty damm poor
choice, but I'm not convinced the other choices of always building with threads
would be as reliable. In general, and this is no reflection on ATLAS in
particular, writing parallel code is a lot harder than serial code, and the bugs
are hard to trace.
Dave
--
To post to this group, send an email to sage-devel@googlegroups.com
To unsubscribe from this group, send an email to
sage-devel+unsubscr...@googlegroups.com
For more options, visit this group at http://groups.google.com/group/sage-devel
URL: http://www.sagemath.org
