Lazy release consistency for software distributed shared memory

Abstract

Relaxed memory consistency models, such a s release consistency, w ere introduced in order to reduce the impact of remote memory access latency in both software and hardware distributed shared memory (DSM). However, in a software DSM, it is also important to reduce the number of messages and the amount o f d a t a e xchanged for remote memory access. Lazy release consistency is a new algorithm for implementing release consistency that lazily pulls modi cations across the interconnect only when necessary. T race-driven simulation using the SPLASH benchmarks indicates that lazy release consistency reduces both the number of messages and the amount of data transferred between processors. These reductions are especially signi cant f o r programs that exhibit false sharing and make extensive use of locks.

Lazy garbage collection of recovery state for fault-tolerant distributed shared memory

“…In [35], log-based recovery was first used in a Lazy Release Consistent (LRC) [20] DSM. Their system uses independent checkpointing along with pessimistic logging by a receiver at synchronization points and access misses.…”

“…Coherency data is propagated as page invalidations. The contents of a page is updated lazily at page fault time, either from its home [37] or from the last writer(s) [20]. If the protocol supports multiple writers, then updates are detected as a difference (diff) between the modified page and a reference copy created before the first write following a page fault [21].…”

“…Lazy Release Consistency (LRC) is a variant of release consistency in which propagation of writes is delayed to the time of an acquire [20]. Writes of a process are grouped into intervals delimited by synchronization operations.…”

“…Page invalidations are propagated in the form of write notices (wn), where a write notice is a pair ðp; tÞ that specifies that some page p was updated during some interval t. The local logical time of a process is defined as a counter of local intervals. The partial-order relation happened-before [25] (denoted 0 ) between intervals across nodes is captured as a vector of logical times called vector timestamp [20]. The vector timestamp T i keeps track of intervals for which write notices were received by process P i .…”

“…ae 1 Introduction I N the last decade, an impressive amount of research has been conducted in software distributed shared memory (DSM), mostly aiming for performance (e.g., relaxed consistency models, lazy protocols, and communication hardware support) [15], [1], [5], [20], [21], [31], [30]. Advances in performance are making it possible for DSM systems to use very large clusters as a cost-effective platform for data-intensive, long-running applications.…”