Motherboard advice. dual processor opteron

[Sxotty]

I am building a system and debating between these motherboards.

They are both Tyan boards and should be really stable.
$436
http://www.tyan.com/products/html/thunderk8w_spec.html

$236
http://www.tyan.com/products/html/tigerk8w_spec.html

You can see there is a large price difference. What I understand is that the second shares the memory with both processors, or more accurately one processor is in control and just passes info to the other. And the more expensive has seperate memory for each processor. Now what is the performance difference going to be between these? The computer will be used to run simulations and models and such mathematically heavy stuff.

Thanks for any info

The key difference I see is NUMA support what the heck is that anyway?

 

[Althornin]

NUMA=non-uniform memory access.
Basically, each processor can access its own memory banks individually - no shared access - provides much greater memory bandwidth.
There is a great review of them on TechReport:
HERE

 

[Sxotty]

Althornin, that is exactly where I looked myself I was still confused about what numa actually was. And it looks like one cannot access it in windows XP profeshional anyway... does XP even support multiple processors?

Anyway thanks for the info.

 

[{Sniping}Waste]

one is PCI-X, the other does not. One shares mem banks, the otherhas its own mem banks for each CPU.

 

[Gubbi]

Althornin, that is exactly where I looked myself I was still confused about what numa actually was. And it looks like one cannot access it in windows XP profeshional anyway... does XP even support multiple processors?

Anyway thanks for the info.

The problem is that Windows XP is not NUMA aware. On top of that it has poor default CPU affinity (ie. processes has zero CPU affinity until you manually set it )

This means that not only can you have your processes bounce between the two CPUs (trashing caches as they do so), but even if you pin down your processes to a given CPU (with affinity control) your proces will still end up allocating 50% of its pages on the other CPUs memory banks, -not good for performance.

I don't know if Microsoft plans to solve the NUMA memory allocation with a patch for the current version of Windows or if we have to wait for the next gen (Win XP 64?).

The board with memory banks connected on each CPU should be a definite win in the future.

Hope this helps

Cheers
Gubbi

 

[mczak]

I'm not sure that even Win XP 64 for AMD64 will support NUMA.
But according to that table, http://www.microsoft.com/windowsserver2003/evaluation/features/compareeditions.mspxWindows Server 2003 supports even today NUMA - only Enterprise and Datacenter Editions, which are "slightly" expensive unfortunately. They also support more than 2 cpus which is the limit of windows xp.
Or you could just install linux if that's an option (for "serious" number crunching stuff, there's usually no reason why you'd need windows).

 

[PiNkY]

NUMA=non-uniform memory access.
Basically, each processor can access its own memory banks individually - no shared access - provides much greater memory bandwidth.

That's actually not what it is. Both processors' memory forms a single address space while being physically separated. This implies that accesses to the other processor's memory are penalized by higher latency. A numa-aware os is "aware of this non-uniform access behaviour" and can optimize memory usage accordingly. So a non-numa aware os profits from the higher bandwidth as well but only profits from lower latency if the accessed datum resides within the processor's own physical memory (in the other case, the latency should be comparable to conventional memory-controller-in-the-northbridge type architectures).