Towards a fully non-blocking share-everything PDES platform

The Ultimate Share-Everything PDES System

“…In these scenarios, threads still block each other because of a critical section implementing the processing stage of the events at the destination object, which limits scalability especially with workloads entailing event bursts at subsets of objects. This problem has been tackled in [12], where non-blocking event-pool management is combined with CPU-dispatching rules that avoid collisions of multiple threads on the state of a same simulation object. However, differently from our proposal, none of these solutions guarantees non-blocking coordination in virtual time.…”

“…In this section we present performance results for a comparison of our Ultimate Share-Everything PDES system-which we refer to as USE-and the last generation Share-Everything solution presented in [12]-which we refer to as SE. The latter does not entail Time-Warp style processing of the events since, for each LP, at most one event is executed speculatively, and is eventually committed or aborted before any other event for the same LP can be CPUdispatched.…”

“…Overall, an ideal share-everything PDES platform should guarantee scalability along the following two dimensions in a combined manner: 1) wall-clock-time coordination across threads and 2) virtual-time coordination across simulation objects, ultimately managed by threads. While point 1) has recently been tackled by a few works [12,16]-they provided non-blocking algorithms for managing the fully shared event pool and share-everything-suited event-dispatching rules that avoid collisions across threads in the access to the state of the simulation objects-a holistic design coping with both the above points in a combined manner is still lacking. Such a design is the objective of this article, where we present a share-everything PDES system that entails speculative execution capabilities of the simulation objects-guaranteeing scalability of virtual-time coordination-and fully non-blocking wall-clock-time coordination across threads.…”

“…The share-everything paradigm that we adopt considers events as fully-shared workload units, thus being able to concentrate the computing power, say threads, on any burst of events that materializes among subsets of objects-any thread can in fact take care of processing whatever event in these bursts, thus contributing to promptly advance the currently hot portions of the simulation model. Furthermore, the speculative processing capabilities we include in our PDES system enable threads to process these bursts with no blocking phase along virtual time, as instead it occurs in previous share-everything proposals like [12].…”

Optimizing Simulation on Shared-Memory Platforms: The Smart Cities Case

“…The share-everything PDES paradigm aims at guaranteeing scalability along two orthogonal dimensions in a combined manner, namely: i) wall-clock time coordination across the different Worker Threads (WTs) which carry on the execution of events, and ii) virtual-time coordination across different simulation objects. As for point i), this is achieved by relying on non-blocking algorithms (Marotta et al 2017;Ianni et al 2017) which allow concurrent access on a shared event pool, without the need to rely on, e.g., spinlockprotected critical sections. These non-blocking algorithms implement an access strategy to avoid collisions across threads in the access both to the state of the simulation objects-this allow to reduce contention on private simulation states-and to simulation platform's metadata.…”