A Pattern Recognition Approach for Speculative Firing Prediction in Distributed Saturation State-Space Generation

Abstract

The saturation strategy for symbolic state-space generation is particularly effective for globallyasynchronous locally-synchronous systems. A distributed version of saturation, SaturationNOW, uses the overall memory available on a network of workstations to effectively spread the memory load, but its execution is essentially sequential. To achieve true parallelism, we explore a speculative firing prediction, where idle workstations work on predicted future event firing requests. A naïve approach where all possible firings may be explored a priori, given enough idle time, can result in excessive memory requirements. Thus, we introduce a history-based approach for firing prediction that recognizes firing patterns and explores only firings conforming to these patterns. Experiments show that our heuristic improves the runtime and has a small memory overhead.

A dynamic firing speculation to speedup distributed symbolic state-space generation

“…Table 1 shows runtimes, total memory requirements for W workstations, and the maximum memory requirements for any workstation, for sequential S m A r T (SEQ) [8] and the original S m A r T N ow (DISTR) [5], and the percentage change w.r.t. DISTR for the naïve (NAÏVE) [6], history-based (HIST) [6], and the new pattern-length-adjusted (LENGTH) and weighted-score-adjusted (SCORE) speculative firing predictions; "d" means that dynamic memory load balancing is triggered, "s" means that, in addition, memory swapping occurs. For the LENGTH heuristic, we initialize MaxDiff to 2 and increase/decrease it by 2 whenever the speculation hit rate increases/decreases 5%.…”

“…At the same time, we also seek to reduce the memory requirements to store this auxiliary information, which were already shown to be potentially substantial for the approach of [6]. The idea is to encode firing patterns implicitly, so that MDD nodes can share the encoding of the same patterns, while reducing overhead at same time.…”

“…Thus, we introduced the idea of using idle workstation time to fire events e with Top(e) > k on saturated MDD nodes at level k a priori, hoping to reduce the time required to saturate MDD nodes at levels above k [6]. An MDD node p at level k is saturated if all events e with Top(e) = k have been fired exhaustively on p. may still need to be fired on p when saturating some MDD node q at level l. To reduce the time to saturate such hypothetical MDD node q, we speculatively create the (possibly disconnected) MDD node p corresponding to the saturation of the result of firing f on p, and cache the result.…”

“…In [6], we tackle this drawback by using workstations idle time to speculatively fire events on MDD nodes, hoping that many of of these firings will be needed later in the computation. To prevent unrestrained specula-tion from squandering the overall NOW memory, we introduce the idea that workstations recognize event firing patterns, namely sequences of events that have been fired on MDD nodes, at runtime, then speculatively explore only firings conforming to these patterns.…”

Speculative Image Computation for Distributed Symbolic Reachability Analysis

“…Table 1 shows runtimes and total memory requirements using W workstations for sequential S m A r T (SEQ) [11] and the original S m A r T N ow (DISTR) [5], and the percentage change w.r.t. DISTR for the naïve (NAÏVE), the history-based (HIST) [6] and the pattern-length-adjusted (LENGTH) or weighted-score-adjusted (SCORE) speculative firing predictions [7]; 'd'means that dynamic memory load balancing is triggered and 's' means that, in addition, memory swapping occurs. The last four columns show the overall hit rate of the speculative caches for the NAÏVE, HIST, LENGTH and SCORE approaches.…”