🥸Advanced Computer Architecture Unit 7 – Memory Hierarchy and Cache Design

Memory Hierarchy Basics

• Memory hierarchy organizes storage devices based on their speed, capacity, and cost
• Faster memory technologies (registers, cache) are located closer to the processor for quick access
• Slower memory technologies (main memory, secondary storage) have higher capacities and are further from the processor
• Principle of locality exploits the idea that recently accessed data and nearby data are likely to be accessed again soon
  • Temporal locality refers to the reuse of specific data within a relatively small time duration
  • Spatial locality refers to the use of data elements within relatively close storage locations
• Memory hierarchy aims to provide the illusion of a large, fast, and inexpensive memory system
• Hit rate represents the percentage of memory accesses satisfied by a particular level of the memory hierarchy
• Miss rate is the percentage of memory accesses that are not satisfied by a particular level and require accessing a lower level

Cache Fundamentals

• Cache is a small, fast memory located close to the processor that stores frequently accessed data
• Acts as a buffer between the processor and main memory to reduce the average access time
• Exploits the principle of locality to improve performance
• Cache hit occurs when the requested data is found in the cache
• Cache miss occurs when the requested data is not found in the cache and must be fetched from a lower level of the memory hierarchy
• Cache size, block size, associativity, and replacement policy are key design parameters that affect cache performance
• Larger cache sizes can store more data but have higher access latency and power consumption
• Smaller block sizes reduce the amount of data transferred per cache miss but may increase the miss rate

Cache Design Principles

• Principle of locality is the foundation for cache design and optimization
• Temporal locality suggests keeping recently accessed data in the cache for future use
• Spatial locality suggests bringing nearby data into the cache along with the requested data
• Larger cache sizes can capture more locality but have diminishing returns and increased cost and complexity
• Higher associativity allows more flexible placement of data in the cache but increases the hit time and power consumption
• Write policies (write-through, write-back) determine how the cache handles write operations and affects performance and data consistency
• Block size balances the amount of data transferred per cache miss with the potential for increased miss rate
• Replacement policies (LRU, random) determine which cache block to evict when a miss occurs and a new block needs to be brought in

Cache Mapping Techniques

• Cache mapping determines how memory addresses are mapped to cache locations
• Direct mapping maps each memory block to a specific cache location based on its address
  • Simple and fast but can lead to conflicts and underutilization of cache space
• Fully associative mapping allows any memory block to be placed in any cache location
  • Flexible but requires complex and expensive hardware for tag comparison
• Set-associative mapping is a compromise between direct and fully associative mapping
  • Memory blocks are mapped to a specific set based on their address but can be placed anywhere within that set
  • Provides a balance between flexibility and hardware complexity
• N-way set-associative mapping divides the cache into N sets, each containing multiple blocks
  • Increases the chances of finding a free block for a new memory block
• Skewed associative mapping uses different hash functions for each way to reduce conflicts and improve performance

Replacement and Write Policies

• Replacement policies determine which cache block to evict when a miss occurs and a new block needs to be brought in
• Least Recently Used (LRU) policy replaces the block that has been accessed least recently
  • Exploits temporal locality by keeping recently accessed blocks in the cache
• Random replacement policy randomly selects a block to evict
  • Simple to implement but may not perform as well as LRU
• First-In-First-Out (FIFO) policy replaces the block that was brought into the cache earliest
  • Easy to implement but does not consider the recency of access
• Write policies determine how the cache handles write operations
• Write-through policy updates both the cache and main memory on every write operation
  • Ensures data consistency but may increase memory traffic and latency
• Write-back policy updates only the cache on a write operation and marks the block as dirty
  • Reduces memory traffic but requires writing back dirty blocks to main memory when evicted
• Write-allocate policy brings a block into the cache on a write miss
  • Ensures future reads to the block can be served from the cache
• No-write-allocate policy does not bring a block into the cache on a write miss
  • Avoids polluting the cache with blocks that may not be read again

Cache Performance Metrics

• Hit rate is the percentage of memory accesses that are satisfied by the cache
  • Higher hit rates indicate better cache performance and reduced access latency
• Miss rate is the percentage of memory accesses that are not satisfied by the cache and require accessing a lower level of the memory hierarchy
  • Lower miss rates indicate better cache performance and fewer long-latency memory accesses
• Average memory access time (AMAT) is the average time to access memory considering both cache hits and misses
  • Calculated as: AMAT = Hit time + (Miss rate × Miss penalty)
• Miss penalty is the additional time required to fetch data from a lower level of the memory hierarchy on a cache miss
  • Depends on the latency and bandwidth of the lower level memory
• Cache throughput measures the number of memory accesses that can be processed by the cache per unit of time
  • Affected by cache size, associativity, and access latency
• Cache efficiency is the ratio of the number of useful memory accesses to the total number of memory accesses
  • Measures the effectiveness of the cache in satisfying memory requests

Advanced Cache Architectures

• Multi-level cache hierarchies use multiple levels of caches (L1, L2, L3) to balance performance and capacity
  • Lower-level caches (L1) are smaller and faster, while higher-level caches (L2, L3) are larger and slower
• Inclusive cache hierarchies enforce the inclusion property, where all data in a lower-level cache must also be present in higher-level caches
  • Simplifies cache coherence and data consistency but may lead to duplication of data
• Exclusive cache hierarchies do not allow data to be present in multiple levels of the cache hierarchy simultaneously
  • Increases the effective cache capacity but may increase the miss rate and data movement between levels
• Non-inclusive cache hierarchies do not enforce any inclusion property and allow data to be present in any level of the cache hierarchy
  • Provides flexibility but may complicate cache coherence and data consistency
• Victim cache is a small, fully associative cache that stores recently evicted blocks from the main cache
  • Reduces the miss penalty by providing a second chance for recently evicted blocks
• Trace cache is a specialized cache that stores decoded instructions in the order they were executed
  • Improves instruction fetch performance and reduces the impact of control flow changes
• Prefetching techniques proactively bring data into the cache before it is requested by the processor
  • Hardware prefetching uses access patterns and heuristics to predict future memory accesses
  • Software prefetching uses explicit instructions to initiate data transfer from memory to cache

Memory Hierarchy Optimization

• Cache optimization techniques aim to improve cache performance and reduce the impact of memory latency
• Cache blocking (tiling) reorganizes data access patterns to maximize spatial locality and minimize cache misses
  • Divides large data structures into smaller blocks that fit in the cache
• Loop interchange changes the order of nested loops to improve spatial locality and reduce cache misses
  • Accesses data in the order that maximizes cache utilization
• Data prefetching brings data into the cache before it is needed, hiding memory latency
  • Hardware prefetching uses access patterns and heuristics to predict future memory accesses
  • Software prefetching inserts explicit prefetch instructions to initiate data transfer
• Cache-conscious data structures are designed to minimize cache misses and maximize cache utilization
  • Examples include cache-oblivious algorithms, cache-aware trees, and cache-friendly layouts
• Memory access scheduling reorders memory instructions to hide latency and improve performance
  • Out-of-order execution allows the processor to execute independent instructions while waiting for memory accesses
• Cache compression techniques reduce the size of cached data to increase the effective cache capacity
  • Compressed caches store compressed data blocks and decompress them on cache hits
• Cache partitioning allocates cache resources to different applications or threads to minimize interference and improve performance isolation
  • Prevents one application from monopolizing the cache and affecting the performance of others
• Cache bypassing selectively bypasses the cache for certain memory accesses to avoid polluting the cache with data that is not reused
  • Avoids bringing in data that is only used once or has poor temporal locality