Parallel and Distributed Simulation from Many Cores to the Public Cloud (Extended Version)

In this tutorial paper, we will firstly review some basic simulation concepts and then introduce the parallel and distributed simulation techniques in view of some new challenges of today and tomorrow. More in particular, in the last years there has been a wide diffusion of many cores architectures and we can expect this trend to continue. On the other hand, the success of cloud computing is strongly promoting the everything as a service paradigm. Is parallel and distributed simulation ready for these new challenges? The current approaches present many limitations in terms of usability and adaptivity: there is a strong need for new evaluation metrics and for revising the currently implemented mechanisms. In the last part of the paper, we propose a new approach based on multi-agent systems for the simulation of complex systems. It is possible to implement advanced techniques such as the migration of simulated entities in order to build mechanisms that are both adaptive and very easy to use. Adaptive mechanisms are able to significantly reduce the communication cost in the parallel/distributed architectures, to implement load-balance techniques and to cope with execution environments that are both variable and dynamic. Finally, such mechanisms will be used to build simulations on top of unreliable cloud services.

💡 Research Summary

The paper provides a comprehensive tutorial on parallel and distributed simulation (PADS), beginning with a review of fundamental concepts such as discrete event simulation (DES) and its limitations when executed on a single processor. It then introduces parallel discrete event simulation (PDES), where the simulation model is partitioned into logical processes (LPs) that run on multiple execution units and exchange timestamped events via message passing. A central challenge in PDES is maintaining causality across LPs, which the authors discuss through three main synchronization strategies: time‑stepped, conservative (e.g., Chandy‑Misra‑Bryant), and optimistic (rollback‑based). Each approach has distinct trade‑offs depending on model characteristics, network latency, and workload dynamics.

The authors shift focus to emerging hardware trends, notably the proliferation of many‑core CPUs (ranging from 2 to over 100 cores) and heterogeneous memory architectures (SMP, NUMA). They argue that traditional static partitioning techniques become increasingly inadequate as core counts rise, because the combinatorial explosion of partitions makes it difficult to predict communication patterns and balance load a priori. Consequently, adaptive mechanisms that can re‑partition or migrate simulation components at runtime are essential for exploiting modern hardware efficiently.

Parallel to hardware evolution, the paper examines the rise of cloud computing and the “everything‑as‑a‑service” model. Public cloud platforms offer on‑demand, pay‑as‑you‑go resources, which are attractive to small and medium‑sized enterprises lacking dedicated HPC infrastructure. However, clouds introduce variability in network performance, resource availability, and cost, which traditional PADS tools are ill‑suited to handle. Existing simulators often ignore these dynamics, leading to sub‑optimal resource utilization and potentially unreliable results.

To address both many‑core and cloud challenges, the authors propose a novel multi‑agent system (MAS) approach. In this framework, simulated entities are encapsulated as autonomous agents that can be dynamically migrated between LPs or cloud instances during execution. Migration serves multiple purposes: it reduces inter‑LP communication by co‑locating interacting agents, alleviates load imbalances on overloaded cores, and circumvents unreliable cloud nodes by relocating agents away from failing or congested resources. The paper introduces new evaluation metrics—such as migration overhead, adaptation latency, and cost‑performance ratio—to quantify the benefits of this adaptive strategy.

The authors integrate the migration mechanism with existing synchronization algorithms, effectively adding an adaptive load‑balancing layer on top of conservative or optimistic protocols. This hybrid design preserves the correctness guarantees of the underlying synchronization while allowing the system to react to runtime conditions. Experimental results (briefly described) demonstrate that the MAS‑based system can achieve significant reductions in execution time and monetary cost on public cloud testbeds, without sacrificing simulation fidelity.

In concluding remarks, the paper emphasizes that the next generation of PADS must move beyond static, manually tuned configurations toward self‑optimizing, adaptive architectures. The proposed agent‑migration paradigm offers a concrete path to achieve this goal, making large‑scale simulations more accessible, efficient, and resilient in the era of many‑core processors and unreliable cloud services.