Here, we discuss the problem of scaling coherence to large systems, potentially with hundreds of cores, such as Blacklight. Blacklight is a cache-coherent system, and we might expect it to have the policy of notifying all processors of every cache miss from a read or write. Large-scale systems like these will attempt to utilize NUMA to reduce latency, but its benefits are defeated when every memory access requires a broadcast to all other processors. Snooping-based cache coherence will suffer from contention for the shared interconnect.
Here, we discuss the problem of scaling coherence to large systems, potentially with hundreds of cores, such as Blacklight. Blacklight is a cache-coherent system, and we might expect it to have the policy of notifying all processors of every cache miss from a read or write. Large-scale systems like these will attempt to utilize NUMA to reduce latency, but its benefits are defeated when every memory access requires a broadcast to all other processors. Snooping-based cache coherence will suffer from contention for the shared interconnect.
This comment was marked helpful 0 times.