CS 7810 Lecture 19 Coherence Decoupling: Making Use of Incoherence J.Huh, J. Chang, D. Burger, G. Sohi Proceedings of ASPLOS-XI October 2004 Coherence / Consistency • Coherence guarantees (i) that a write will eventually be seen by other processors, and (ii) write serialization (all processors see writes to the same location in the same order) • The consistency model defines the ordering of writes and reads to different memory locations – the hardware guarantees a certain consistency model and the programmer attempts to write correct programs with those assumptions Consistency Examples Initially, A = B = 0 P1 A=1 if (B == 0) critical section P2 B=1 if (A == 0) critical section Initially, A = B = 0 P1 A=1 P2 P3 if (A == 1) B=1 if (B == 1) register = A P1 Data = 2000 Head = 1 P2 while (Head == 0) {} … = Data Snooping-Based Cache Coherence • Caches share a bus; every cache sees each transaction in the same cycle; every cache manages itself • When one cache writes to a block, every other cache invalidates its copy of that block • When a cache has a read miss, the block is provided by memory or the last writer • Protocols are defined by states: MSI, MESI, MOESI Processor Processor Processor Processor Caches Caches Caches Caches Memory Directory-Based Cache Coherence • A directory keeps track of the sharing status of each block • Every request goes to the directory and the directory then sends directives to each cache – the directory is the point of serialization (just as the bus is, in a snooping protocol) • For example, on a write, the request reaches the directory, the directory sends invalidates to other sharers, and permissions are granted to the writer Processor Processor Processor Processor Caches Caches Caches Caches Network Memory Directory TLDS • A certain ordering of reads and writes is assumed – if that ordering is violated, the thread is re-executed • The coherence protocol is used to propagate writes Thread 1 Thread 2 Thread 3 Thread 4 Caches Caches Caches Caches Memory The Traditional Model • No thread is speculative – a parallel application with synchronization points and parallel regions and guaranteed to execute correctly with no need for re-execution • Threads wait at synchronization points and wait for the correct permissions for every block of data Thread 1 Thread 2 Thread 3 Thread 4 Caches Caches Caches Caches Memory Coherence Decoupling • A simple coherence protocol is often a slow protocol – for example, a simple protocol may not allow multiple outstanding requests • Coherence decoupling: maintain a fast and incorrect protocol; and a slow and correct backing protocol; incurs fewer stalls in the common case and occasional recoveries Coherence Decoupling • A coherence operation is broken into two components: (i) acquiring and using the value, (ii) receiving the correct set of permissions SCL Protocol • Why does speculative cache look-up work? False sharing: a line was invalidated, but a different word was written to Silent stores or value locality If there is spare bandwidth, updated values can be pushed out to sharers Implementation • The Miss Status Holding Register (MSHR) keeps track of outstanding requests – it can buffer the speculative value and ensure it matches the correct value – on a mis-speculation, that instruction is treated like a branch mis-predict • Speculation on a coherence operation is no different from traditional forms of speculation Coherence Decoupling Components Microbenchmark Behavior Results Results Summary • Arguments for coherence decoupling: Reduces protocol complexity Reduces programming complexity Marginal hardware overhead Coherence misses will emerge as greater bottlenecks? • What is the expected trend for CMPs? Title • Bullet