[Beowulf] Joe Blaylock's notes on running a MacOS cluster, Nov. 2007
bill at cse.ucdavis.edu
Tue Nov 20 23:44:26 PST 2007
Tim Cutts wrote:
> I have to say I have some sympathy with that view. Much though I love
> OS X as a desktop OS, I have to wonder why anyone would jump through
> hoops building a cluster running just Darwin, when they could probably
> have done it much more easily, and with much more community support, if
> they used Linux... It just seems like a lot of extra pain and
> difficulty for no tangible benefit.
Yeah, a bit of a mystery to me, I was comparing G5 + OSX vs opteron + linux
for a cluster. We had the source code and got some fairly horrible numbers
for the mac. Several of our codes required source code changes for OSX,
mostly semaphores, pthreads, and shared memory related. Performance was
sometimes ok (in user space), but packets/sec (with a cpu job running),
context switching, and performance with shared memory was often off by a
factor of 10. I wasn't sure if I should blame the hardware, the scheduler,
or the OS (besides the scheduler).
One of our more important codes did run, and was faster per CPU on the
G5 (at least until I picked a better compiler on opteron), but when we used
MPI (I think it was lam at that point) it performed rather poorly.
So as a learning experience, sure. Practical advantages, I can't see any.
The 2.0 GHz dual core xservs for $3k with 1GB ram aren't exactly burning up
the price/performance curve, at least you can get ECC these days.
Seems like if linux is going to see a competitor for HPC type usage soon it's
going to be opensolaris.
BTW, if you want to price out an xserv I couldn't find it at the apple store
unless I used the search widget.
More information about the Beowulf