Software DSM protocols that adapt between single writer and multiple writer

Abstract

We present t wo s o f t ware DSM protocols that dynamically adapt between a single writer (SW) and a multiple writer (MW) protocol based on the application's sharing patterns. The rst protocol (WFS) adapts based on write-write false sharing the second (WFS+WG) based on a combination of write-write false sharing and write granularity. The adaptation is automatic. No user or compiler information is needed. The choice between SW and MW is made on a perpage basis. We measured the performance of our adaptive protocols on an 8-node SPARC cluster connected by a 155 Mbps ATM network. We used eight applications, covering a broad spectrum in terms of write-write false sharing and write granularity. W e compare our adaptive protocols against the MW-only and the SWonly approach. Adaptation to write-write false sharing proves to be the critical performance factor, while adaptation to write granularity p l a ys only a secondary role in our environment and for the applications considered. Each of the two adaptive protocols matches or exceeds the performance of the best of MW and SW in seven out of the eight applications. For these applications, speedup improvements over SW range from 1.02 to 2.7. The largest improvements over SW occur for applications with high write-write false sharing. Compared to MW, speedups improve b y a factor of 1.02 to 1.6, with the largest improvements occurring for applications with little or no write-write false sharing. Both WFS and WFS+WG speedups fall below the best of MW and SW for one application, but only by a factor of 1.09 and 1.06. In addition, memory usage is reduced considerably compared to MW, in some cases making the memory overhead all but negligible.

A performance comparison of homeless and home-based lazy release consistency protocols in software shared memory

“…For migratory data, Tmk uses half as many messages, because it transfers the diff(s) directly from the last writer to next writer. (The current implementation of Tmk avoids the problem of diff accumulation through the methods described by Amza et al [2].) For producer/consumer data, the two protocols use roughly the same number of messages.…”

“…In Tmk, the bucket is transferred once from the last writer to the new writer. The diff accumulation problem in Tmk is avoided through the method described by Amza et al [2]. [8] is a molecular dynamics simulation.…”

Using prediction to accelerate coherence protocols

“…Predictions can be made by programmers [14,38,1,25], compilers [23,32,31,15], software [4], or hardware. Specialized predictors in hardware include read-modify-write operation prediction in the SGI Origin protocol [19], pair-wise sharing prediction in SCI [34], dynamic self-invalidation [20], and migratory protocols [12,35].…”

The effect of network total order, broadcast, and remote-write capability on network-based shared memory computing

“…Amza et al [3] describe adaptive extensions to the TreadMarks [2] protocol that avoid twin/diff operations on shared pages with only a single writer (pages with multiple writers still use twins and diffs). In Cashmere, if a page has only a single writer, the home always migrates to that writer, and so twin/diff operations are avoided.…”