体系结构量化研究方法（第二章-2）

Chapter 2 Memory Hierarchy Design

Memory Technology and Optimizations

Performance metrics

• Latency
  • Access time
    • Time between read request and when desired word arrives
  • Cycle time
    • Minimum time between unrelated requests to memory
  • Bandwidth

High Bandwidth Memory (HBM)

• A packaging innovation rather than a circuit innovation
• Reduce access latency by shortening the delay between the DRAM and the processor
• Interposer stacking (2.5D) available, vertical stacking (3D) still under development due to heat constraint

Six Basic Cache Optimizations

Average memory access time:

• Average memory access time = Miss rate × Miss penalty + Hit Time

Six basic cache optimizations:

• Reduce the miss rate
  • Larger block size (reduce compulsory misses)
    • Increase conflict misses, Increases miss penalty
  • Larger total cache capacity (reduce capacity misses)
    • Increase hit time, Increase cost, Increase power
  • Higher associativity (reduce conflict misses)
    • Increase hit time
      • More tag comparisons
    • Increase power
• Reduce the miss penalty
  • Higher number of cache levels
    • L1 miss penalty can be reduced by introducing L2 cache
    • balance fast-hits and few-misses
  • Giving priority to read misses over writes
    • “read-after-write” data hazard
      • Let read wait until write buffer is empty
      • Or check the content of the write buffer on a read miss
        • conflict -> wait
        • non-conflict -> handle read miss first
      • Prioritizing read can reduce unnecessary stalling of read miss
• Reduce the time to hit in cache
  • Avoiding address translation in cache indexing
    • Virtual Cache avoids address translation
      • virtual address -> cached block
        • use virtual address in both set indexing & tag comparison
    • Obstacles to realize virtual cache
      • To ensure protection
        • copy protection bit of TLB into an added field on cache miss
      • To support process switching
        • adopt PID marker in cache
      • To ensure compatibility with multiple alias logic addresses (cache consistency)
        • antialiasing: compare physical addresses of existing blocks with the fetched one

Ten Advanced Cache Optimizations

Advanced Cache Optimizations

• Reducing the hit time
• Reducing the miss penalty
• Reducing the miss rate
• Increasing cache bandwidth
• Reducing th