Clustered time warp and logic simulation

Abstract

In this research, the feasibility of using parallel discrete-event simulation techniques to run logic-level circuit simulations on general purpose distributed memory architectures is investigated. After studying the characteristics of VLSI models, we introduce Clustered Time Warp (CTW), a novel approach to parallel discrete event simulation. In CTW, the logical gates of a circuit are partitioned into clusters and the synchronization algorithm makes use of an optimistic approach between the clusters and a sequential approach within the clusters. We also present a new family of three space-based checkpointing algorithms for use with CTW. Results show that each checkpointing algorithm developed for CTW occupies a different point in the spectrum of possible trade-offs between memory usage and execution time. We also present a dynamic load balancing algorithm developed for Clustered Time Warp which focuses on distributing the load of the simulation evenly among the processors and then tries to reduce inter-processor communications. We make use of a triggering technique based on the throughput of the simulation system. Performance results show that by dynamically balancing the load, the throughput of the simulation system could be improved by more than a 100%. No substantial improvement was observed on the overall simulation time when trying to minimize inter-processor communications, suggesting that load distribution is the most important factor to be taken into consideration in speeding up the simulation of digital circuits. Furthermore, we examine the impact of partitioning and mapping on the performance and behavior of the Clustered Time Warp algorithm. We show that partitioning algorithms which try to minimize the number of cutsets between the partitions do not necessarily succeed in minimizing inter-processor communications. We also show that in our environment, load imbalance has a stronger effect than rollback overhead. Finally, we study the problem of scalability encountered when using optimistic techniques. We show that the performance of Time Warp can greatly suffer from rollback explosions or when the dog chasing its tail phenomenon is observed. We also show that Clustered Time Warp is less sensitive to these phenomenons and as such, is more scalable than Time Warp.

Cite this work

No public copy is stored for this record, so only the metadata is cited.

Stored only when both papers are in the corpus — 47 citations exist upstream.

Papers that cite the same references — the relational fallback while embeddings are still backfilling.