Transaction-Oriented Simulation In Ad Hoc Grids: Design and Experience

In this paper we analyse the requirements of performing parallel transaction-oriented simulations within loosely coupled systems like ad hoc grids. We focus especially on the space-parallel approach to parallel simulation and on discrete event synchronisation algorithms that are suitable for transaction-oriented simulation and the target environment of ad hoc grids. To demonstrate our findings, a Java-based parallel simulator for the transaction-oriented language GPSS/H is implemented on the basis of the most promising shock-resistant Time Warp (SRTW) synchronisation algorithm and using the grid framework ProActive. The analysis of our parallel simulator, based on experiments using the Grid5000 platform, shows that the SRTW algorithm can successfully reduce the number of rolled back transaction moves but it also reveals circumstances in which the SRTW algorithm can be outperformed by the normal Time Warp algorithm. Finally, possible improvements to the SRTW algorithm are proposed in order to avoid such problems.

💡 Research Summary

The paper investigates how to run transaction‑oriented discrete‑event simulations efficiently on loosely coupled, dynamic infrastructures known as ad‑hoc grids. After outlining the specific challenges of such environments—namely, volatile resource availability, heterogeneous network latency, and limited central control—the authors focus on two technical pillars: (1) space‑parallel decomposition of the simulation model, and (2) synchronization algorithms that can maintain a consistent Global Virtual Time (GVT) across distributed logical processes.

Transaction‑oriented simulation, exemplified by the GPSS/H language, treats a “transaction” as a mobile entity that traverses resources and changes state. This mobility makes it natural to partition the model spatially, assigning each partition to a separate logical process. However, the asynchronous movement of transactions across partitions creates causality violations that must be resolved without sacrificing parallelism. Traditional conservative synchronization avoids violations by waiting for safe events, but this induces excessive idle time. Optimistic approaches such as Time Warp allow processes to proceed speculatively and roll back when a causality error is detected, offering higher parallelism at the cost of potentially large rollback overhead.

To mitigate rollback costs, the authors adopt the Shock‑Resistant Time Warp (SRTW) algorithm. SRTW augments classic Time Warp with a predictive “shock” detector that estimates the likelihood of future rollbacks and dynamically throttles speculative execution to keep rollback frequency low. The paper’s contribution is threefold: (1) a detailed analysis of the requirements for transaction‑oriented simulation on ad‑hoc grids, (2) the design and implementation of a Java‑based GPSS/H parallel simulator built on the ProActive grid middleware, and (3) an empirical evaluation on the Grid5000 testbed that compares SRTW with standard Time Warp under various workload conditions.

Implementation details reveal that ProActive’s active‑object model maps each GPSS/H partition to an active object, allowing transaction moves to be expressed as remote method invocations. This abstraction simplifies dynamic node allocation and fault tolerance, but introduces Java‑specific overheads such as garbage‑collection pauses and serialization costs. The simulator therefore includes custom buffering and lightweight serialization to reduce communication latency.

The experimental campaign consists of three scenarios: (a) a high‑transaction‑density model with frequent inter‑partition interactions, (b) a low‑density model where transactions rarely cross partition boundaries, and (c) a mixed‑interaction model with varying communication patterns. Results show that in scenario (a) SRTW reduces the number of rolled‑back transaction moves by roughly 30 % and improves overall execution time by about 12 % compared with classic Time Warp. In scenario (b), however, the predictive throttling of SRTW adds unnecessary overhead, leading to an 8 % slowdown relative to Time Warp. Scenario (c) demonstrates that SRTW’s adaptive control can alleviate network congestion when communication spikes, but the benefit is modest and depends on the accuracy of the shock estimator.

From these findings the authors conclude that SRTW is not a universal replacement for Time Warp; its advantages manifest only when rollback probability is significant. For workloads with sparse cross‑partition traffic, a hybrid approach that switches between optimistic and conservative modes—or that disables the shock‑resistance component—yields better performance.

The paper also proposes several enhancements to SRTW tailored to ad‑hoc grid characteristics. First, an adaptive thread‑pool mechanism that continuously monitors rollback statistics and adjusts speculative execution parameters on a per‑partition basis. Second, a variable GVT computation interval that aligns GVT updates with the observed event generation rate, thereby reducing unnecessary synchronization. Third, a pre‑execution profiling step that identifies “hot‑spot” partitions and redistributes load dynamically to balance transaction traffic.

Finally, the authors discuss future work, including porting the simulator to a lower‑level language (e.g., C++) to eliminate Java‑related latency, integrating more sophisticated network‑aware scheduling policies, and extending the evaluation to larger, more heterogeneous ad‑hoc grid deployments. The overall contribution is a comprehensive blueprint for building scalable, transaction‑oriented simulation platforms that can exploit the opportunistic resources of ad‑hoc grids while managing the inherent trade‑offs between optimism, rollback overhead, and dynamic resource variability.