Prefetch 論文閱讀

contributed by<zhanyangch>

tags:zhanyangch sysprog2017 week3

論文閱讀

When Prefetching Works, When It Doesn’t, and Why

• hardware/software prefetching 的合作與衝突

1. INTRODUCTION
   • prefetching : tolerating cache miss latency exist
   • compiler 只能自動做到簡單的 prefetching，需插入 prefetching intrinsics，但很少嚴謹的規則，以及缺少對 software and hardware prefetching 的複雜組合的瞭解
   • intrinsics：看起來像函數，會被 compiler 直接替換成組語
     
     
   • 圖：sw software prefetching，GHB　STR：hardware，比較有 sw prefetching 跟只有hw prefetching 的 speed up，以5% 劃分 positive netual negative
   • software prefetching targets short array streams, irregular memory address patterns, and L1 cache miss reduction, there is an overall positive impact with code examples
   • software prefetching can interfere with the training of the hardware prefetcher, resulting in strong negative effects, especially when using software prefetch instructions for a part of streams.
   • stream prefetcher(STR):prefectch 連續的資料
     stride prefetcher(GHB):prefectch 距離 stride 的指令或資料，stride 由先前的紀錄得知
2. BACKGROUND ON SOFTWARE AND HARDWARE PREFETCHING
   
   
   • 不同的 data structures 的 access patterns 影響 prefectch 的方式
   • Recursive Data Structures (RDS)
   • x86 SSE SIMD extensions 的 instrinsic 會被轉換成 2道指令(direct addr)或4道指令(indirect addr)
     
     
   • Prefetch Classification:跟時間有關 Timely Late Early，重複 Redundant_dc Redundant_mshr，錯誤　Incorrect
   • Software Prefetch Distance
     D:prefetch distance
     l:prefetch latency
     s:length of the shortest path through the loop body
     D 必須大於 memory latency，但太大會造成 cache 內的資料被提早逐出，導致 cache miss 提高
     
     $$D\ge \frac{l}{s}.$$
   • direct memory indexing 易用 hardware prefetch
     indrect memory indexing，需特殊的硬體，易用 software prefetch
3. POSITIVE AND NEGATIVE IMPACTS OF SOFTWARE PREFETCHING

• software pefecting 的優點
  • hardware 的資源有限，Stream 數目多會有困難
  • stream detectors and book-keeping mechanisms
  • hardware prefetcher 需要訓練，如果 stream 長度太短 cache misses 不足以訓練
  • Hardware prefetchers 主要放 lower level cache，降低 L1 cache pollution
  • 在 software 中迴圈邊界可以被簡單的計算並且可藉由loop unrolling, software pipelining, and using　branch instructions 避免 prefetch requests out of array bounds
• software pefecting 的缺點
  • 增加指令數
  • prefetch 那些 data 事先決定，無法依 runtime 的情形決定
  • 對於指令少的迴圈很難插入 prefetching instruction，需要 loop splitting 的技巧
• software + hardware 優點
  • Handling Multiple Streams
  • Positive Training
• software + hardware 缺點
  • Negative Training : Software較慢或 hide streams
  • Harmful Software Prefetching : stress on cache, memory bandwidth

4. EXPERIMENTAL METHODOLOGY
   
   $$Distance=\frac{K\times L\times {IPC}_{bench}}{W_{loop}}.$$
   • where K is a constant factor, L is an average memory latency, IPCbench is the profiledaverage IPC of each benchmark, and W is the average instruction count in one loop iteration
   • MacSim, a trace-driven and cycle-level simulator
5. EVALUATIONS: BASIC OBSERVATIONS ABOUT PREFETCHING
   • Instruction Overhead:prefetch instructions + indirect memory accesses，Overhead 高不等於效能差
   • Software Prefetching Overhead:cache pollution, bandwidth consumption, memory access latency,redundant prefetch overhead,instruction overhead:除去各別變因，觀察對效能的影響
     • small:cache pollution,bandwidth
     • larger:memory latency
     • negligible:redundant prefetch instructions
     • not high:instruction overhead
   • optimal distance might vary by machine configurations
   • 利用軟體訓練硬體的成效有限:因軟體會放 L1，而硬體在 L2，且 hardware prefetch requests can be too aggressive
   • profile-guided optimization
   • ASD hardware prefetcher:Short Streams
   • content directed prefetching :irregular data structures
   6. 個案探討

參考

cmu Computer Architecture ppt