Go backward to Distributed Shared Memory Go up to Top Go forward to COMA Architectures

NUMA Architectures

• Each processor has memory and cache.
• All memories form a global address space.
• Directory-based protocols for cache consistency (no snooping possible).
  • Each cache copy is registered in directory.
  • If all cache copies of a data item are shared, memory requests yield new copies.
  • If cache copy is owned, memory request is forwarded to owner.
• Example: Stanford DASH multiprocessor

  Directory Architecture for Shared Memory.

Problem: Number of cache copies may grow arbitrarily.