A load-sharing architecture for high performance optimistic simulations on multi-core machines

Abstract

In Parallel Discrete Event Simulation (PDES), the simulation model is partitioned into a set of distinct Logical Processes (LPs) which are allowed to concurrently execute simulation events. In this work we present an innovative approach to load-sharing on multi-core/multiprocessor machines, targeted at the optimistic PDES paradigm, where LPs are speculatively allowed to process simulation events with no preventive verification of causal consistency, and actual consistency violations (if any) are recovered via rollback techniques. In our approach, each simulation kernel instance, in charge of hosting and executing a specific set of LPs, runs a set of worker threads, which can be dynamically activated/deactivated on the basis of a distributed algorithm. The latter relies in turn on an analytical model that provides indications on how to reassign processor/core usage across the kernels in order to handle the simulation workload as efficiently as possible. We also present a real implementation of our load-sharing architecture within the ROme OpTimistic Simulator (ROOT-Sim), namely an open-source C-based simulation platform implemented according to the PDES paradigm and the optimistic synchronization approach. Experimental results for an assessment of the validity of our proposal are presented as well

Transparent multi-core speculative parallelization of DES models with event and cross-state dependencies

“…As an additional preliminary note, in our design we targeted PDES platforms relying on the multi-threading paradigm. These have been shown to provide a set of benefits and to support optimized resource usage policies (see, e.g., [4,11,27,26]) when compared to the traditional counterpart where parallelization is achieved by running a set of single-threaded processes within the simulation platform. Overall, we designed a memory management architecture allowing not to loose the benefits from multi-threading.…”

“…It models interference across different channels within a same cell, and power management upon call setup/handoff in a high fidelity fashion. This same model has been already used to assess the multi-threaded version of ROOT-Sim where we have integrated ECS, and its detailed description can be found in [26]. Two specific aspects are relevant for this study: 1) each simulation object models an individual cell, and the interactions between objects exclusively take place via handoff events of mobile devices across different cells (hence memory accesses by the different simulation objects are intrinsically disjoint); 2) the average granularity (CPU requirement) of the events is directly proportional to the wireless channel utilization factor, since the more channels are busy, the more complex is the calculation of interference and Signal-to-Interference Ratio (SIR) while simulating power regulation.…”

NUMA Time Warp

“…In Figures 7-9 we report results related to a different configuration of the traffic application, where the volume of cars managed is greater (particularly, the car inter-arrival frequency is derived from [5]), and where we consequently get the possibility of accidents to occur according to the aforementioned statistical distribution. This gives rise to dynamic unbalance of the workload across the simulation objects, such that the dynamic re-balance of the actual workload on the computing platform is actuated by ROOT-Sim according to the simulation-object-to-worker-thread migration policy presented in [35]. Also, for this scenario we improve the level of granularity of the representation of the Italian highway by increasing the total number of simulation objects to 1024 (hence each object simulates a shorter potion of the highway, which allows for a finer grain representation of the -relative-movement of vehicles).…”

Cross-state events: A new approach to parallel discrete event simulation and its speculative runtime support

“…As hinted, in our approach we target optimistic PDES platforms relying on the multithreading paradigm, which have been shown to provide a set of benefits and to support optimized resource usage policies (see, e.g., [4,[16][17][18]) when compared to the traditional counterpart where parallelization is achieved by running a set of single-threaded processes within the simulation platform. Hence, we target the scenario where multiple threads can take care of dispatching whichever simulation object for execution which takes place by simply calling an event handler, with proper input parameters, along that thread.…”