[G4] OS X SMP

Alex alist at sprint.ca
Mon Feb 9 09:48:46 PST 2004


On Sunday, Feb 8, 2004, at 12:04 Canada/Eastern, sr ferenczy wrote:

>> Does the VT cluster run Mac OS X -- and is it really relevant to our 
>> issue?

> yes it does, and yes, i believe it is highly relevant. it is the 
> largest working example of SMP we have on the macintosh, and i believe 
> it shows how well the OS handles SMP.

Even if it were, and even if it did, how is that relevant to the 
question of whether M-Audio Dual Box Pre-amp can run or not on a DP 
450MHz Mac? If you see the relevance, I don't, so perhaps it's best to 
agree to disagree.

As to SMP, perhaps we are talking at cross-purposes.

Symmetric Multiprocessing occurs when all processors can access, and 
are equally close to, all RAM locations. For technical reasons, SMPs do 
not scale well over 16 processors, and significant increases in 
apparent processing power are not expected for average users beyond 4 
processors. Had the System X (I believe that's the "official" name of 
the Virginia Tech supercomputer) been "the largest working example of 
SMP" it would certainly have been a major theoretical breakthrough -- 
let alone a massive invalidation of Amdahl's law.

The ways to get around SMP limitations are the NUMA architecture and 
cluster computing. The latter is a group of interconnected computers 
(nodes) working together. The node itself may be SMP or not, but the 
whole cluster isn't -- it can't be. The operating system running the 
node is not the same thing as the software which implements the 
cluster. The crucial element is the parallel communication libraries.

Clustering is not new. For instance, Beowulf clusters (Linux- or 
FreeBSD-based networked PCs) have been in use for some time. The major 
problem has been the fact that TCP/IP was not designed to support this 
type of task, and the high latency of 100Mbit/s Ethernet.

So now we come to System X. It is not SMP, and it is not a theoretical 
novelty -- it uses a well-established clustering model. If my 
understanding is correct, the nodes (which are SMP, and which are Mac 
OS X-run, because it is based on FreeBSD) are connected by InfiniBand 
and clustering is achieved by MPI-based libraries. What is truly 
remarkable about System X is that a cluster of such size could be made 
to work on off-the-shelf consumer-level components for a budget that 
amounts to peanuts when talking supercomputing -- and Dr Varadarajan's 
Deja Vu software, which does the essential fault tolerance job.

But this is not really my cup of tea, and perhaps my understanding is 
wrong. If so, I should be grateful if it were corrected.

f




More information about the G4 mailing list