what's a NUMA box

[Tanner]

sorry if this Q. is not technical enough...but this is where I saw the term, so here's where I ask

 

[Nothinman]

Check out this page.

 

[ProviaFan]

Non-Uniform Memory Access for the lazy

 

[Nothinman]

Little vague don't you think?

 

[SuperTool]

NUMA is Non-Uniform Memory Access.
Basically, memory is distributed between nodes, each node having CPU, cache, Memory. If CPU tries to access memory at it's own node, it's faster then when it tries to access memory at some other nodes, hence non-uniform.
And in case you want to know what CC-NUMA is, CC stands for cache coherent, which means that cache lines at all nodes are updated or invalidated if some other node writes to the same location that is stored in cache.
That is if I recall things correctly.
Actually for those who really want to know, I recommend "Parallel Computer Architecture a hardware/software approach" book by DE Culler and JP Singh. It goes into these things in very gory detail (1000+ p)

 

[Tanner]

Nothinman thanks for the linky!

SuperTool Thanks for the summary d00d.

I ckd out that link...man...that's some seriously difficult stuff... looks like it actually requires a degree in the field to do anything with it! Unlike some of the hardware/software/networking that goes on around here and in small circles around the globe...

 

[lifeguard1999]

Just 2 additions to what has already been said.

1. A node may contain more than one CPU. For example, the SGI Origin 3800 has 4 CPUs on a node.

2. Problems may occur if the process is not bound to a node or CPU. For example, suppose a program starts on node 0 and gets the memory local to the node. If the operating system moves the process to node 64 and the memory is still local to node 0, you can see large variations in the run time of your code. We have experienced an increase in execution time up to 5X (though 2X is more common). Binding processes to nodes solves this problem (and may create others).

 

[Nothinman]

1. A node may contain more than one CPU. For example, the SGI Origin 3800 has 4 CPUs on a node.

I believe the SGI site states that, also mentioned is CC-NUMA where the cache on all the CPUs is kept coherent, although I would think that would be a big performance hit.\

2. Problems may occur if the process is not bound to a node or CPU

What's the point then? Isn't the main point of NUMA to have a pretty much transparent cluster where CPUs and memory are 'readily available' from every node?