Note_Book_Multi-Core_CacheHierarchies-Ch2

2.1 Data Management for a Large shared NUCA Cache 

• NUCA

• Static : long acces time
• Dynamic : time a little less, increased complexity, more power

• local region(of CPU) : private data

• center region : shared data

• Block placement: random (based on tag)

• Problem : locating a block

• sol: multicast partially - > if miss :multicast remained bank

•     fast but with heavy network and bank load

• Another sol: partial tag by kim

• Problem of above:

• high overhead of partial tag
• block is hit only in right bank

• Sharing Degree

• Completely Shared : SD = 16, coherence only in L1
• No Shared : SD = 1, coherence for every private L2 

• Varying SD to get better performance for different applications
• SD increased, hit rate increase
• SD of 2 or 4 is often better

• Hybrid shared/private LLC
• ways of set are distributed across slices
• Some are private while others are shared
• The blocks accessed in shared will move to private
• The evicted block will move to shared cache -> victim cache

• NuRAPID : Non-Uniform access with Replacement And placement usIng Distance associativity
• tag-lookup in centralized structure near CPU
•  direct to the bank contains block
• accessed blocks are swapped with not recently touched blocks
•  blocks near the core are often hot
• reduce inner-bank traffic
• Data pointer is needed for block to save entry
• Replacement policy
• may need to independently track LRU among banks
• Random selection is actually good enough
• CMP-NuRAPID
• Organized private L2
• Decouples Tag and Data Array
• Tag Arrays are private while Data Array is shared.
• Each core maintains its own private tag array
• May lower the capacity
• Multicore read/write the same block in private L2 may leads to coherence miss
• L2