Very impressive increase in throughput, especially considering that they're now limited by the IO hubs.
One thing that puzzles me about this design, they use it to boost the memory bandwidth in NUMA systems but isn't that more usually done using switches or special cards with crossbars or point-to-point links rather than routers?
AFAIK there is no way to do DMA between NIC cards and GPU. However it is possible with current hardware. Having that would improve GPU performance slightly.
On the other hand, this research is limited by PCI-E bandwidth (quote "performance of our system is limited by the dual-IOH") so CPU/GPU does not seem to be bottleneck in most of the tests.
4 comments
[ 3.8 ms ] story [ 20.0 ms ] threadOne thing that puzzles me about this design, they use it to boost the memory bandwidth in NUMA systems but isn't that more usually done using switches or special cards with crossbars or point-to-point links rather than routers?
On the other hand, this research is limited by PCI-E bandwidth (quote "performance of our system is limited by the dual-IOH") so CPU/GPU does not seem to be bottleneck in most of the tests.