Composable memory transactions-英文文献.pdf

Optimizing Memory Transactions Tim Harris† Mark Plesko† Avraham Shinnar‡ David Tarditi† Microsoft Research† Harvard University‡ tharris@ markples@ shinnar@ dtarditi@ Abstract processor machines [12] because they are parallelism preserving : Atomic blocks allow programmers to delimit sections of code as atomic blocks can execute concurrently so long as a location being ‘atomic’, leaving the language’s implementation to enforce atomic- updated in one block is not being accessed in any of the others. This ity. Existing work has shown how to implement atomic blocks over preserves the kind of sharing allowed in a conventional data cache. word-based transactional memory that provides scalable multi- Although they scale well, current implementations of atomic processor performance without requiring changes to the basic blocks introduce substantial runtime overhead [12]. They are built structure of objects in the heap. However, these implementations using word-based software transactional memory (STM) which perform poorly because they interpose on all accesses to shared allows a series of memory accesses made via the STM library to be memory in the atomic block, redirecting updates to a thread-private performed atomically. There are three main reasons for the runtime overhead, which we discuss in more detail in Section 2: log which must be searched by reads in the block and later recon- ciled w