We describe work in progress on the Spectre system which aims to provide high performance computing over distributed shared memory, targeting workloads such as graph algorithms for which functional or dataflow decompositions are inefficient. We exploit aggressive speculation to hide the latency of remote memory accesses and synchronization, and execute all code transactionally so that mis-speculations can be discovered and reverted.

Unlike previous speculative transactional systems, Spectre makes side effects visible beyond transaction boundaries before the transactions have committed, tracking dependencies to ensure correctness on abort: we call this property transgression. We outline the Spectre design and provide preliminary results from a microbenchmark to motivate the approach.

PDF file

In  2nd ACM SIGOPS Asia-Pacific Workshop on Systems

Publisher  ACM
Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy otherwise, to republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee.

Details