NUMA
A Non Uniform Memory Architecture combines multiple pools of memory, each with different characteristics. An example of this architecture is a simple GLobal/Local memory design, which provides each Processor with a private, Local memory and also provides a Global Memory region which may be shared by all Processors. The OS resides in Global Memory, while processes/threads executing on each Processor reside in the Local Memory of that Processor. Global Memory is cache coherent among all Processors, while Local isn't.
I am aware that the AMD architecture provides multiple, independent memory pools for each Processor in a multi-processor design. However, to my knowledge, there are no mainstream open-architecture designs which incorporate NUMA.
Any thoughts?
- Login or register to post comments
- Printer friendly version
- 545 reads
Thanks for considering NUMA ?
This concept may have started with the 5000 node parallel thinking machine back in the 1980s. In those days memory is scarce and knowing locations and availability may offer some help. Today, Intel had a project to have people offer their computer processor to be used on the spare time for huge computational projects.
With hdd into terabytes of memory all dependent of PRML data compression and multilayered magnetic fields, memory is cheap. Then data compression ratio had improved along the way. We are memory happy.
But we are processor poor. Multicores are just the beginning. Wafer scale computers of DSPs, is the key to processor happiness. Combining Riscs and A/D, D/A converters for data transfer; compression and decompression is the future work load. Not processing calculations, but reading tables is much faster. What are we getting ourselves into these days?