A GPU-based discrete event simulation kernel

Abstract

The graphic processing unit (GPU) can perform some large-scale simulations in an economical way. However, harnessing the power of a GPU to discrete event simulation (DES) is difficult because of the mismatch between GPU’s synchronous execution mode and DES’s asynchronous time advance mechanism. In this paper, we present a GPU-based simulation kernel (gDES) to support DES and propose three algorithms to support high efficiency. Since both limited parallelism and redundant synchronization affect the performance of DES based on a GPU, we propose a breadth-expansion conservative time window algorithm to increase the degree of parallelism while retaining the number of synchronizations. By using the expansion method, it can import as many as possible ‘safe’ events. The irregular and dynamic requirement for storing the events leads to uneven and sparse memory usage, thereby causing waste of memory and unnecessary overhead. A memory management algorithm is proposed to store events in a balanced and compact way by using a lightweight stochastic method. When events processed by threads in a warp have different types, the performance of gDES decreases rapidly because of branch divergence. An event redistribution algorithm is proposed by reassigning events of the same type to neighboring threads to reduce the probability of branch divergence. We analyze the superiority of the proposed algorithms and gDES with a series of experiments. Compared to a CPU-based simulator on a multicore platform, the gDES can achieve up to 11×, 5×, and 8× speedup in PHOLD, QUEUING NETWORK, and epidemic simulation, respectively.

Distributed simulation: state-of-the-art and potential for operational research

“…Researchers have also sought to maximise performance from multi-core CPUs (Liu & Wainer, 2012;De Munck, Vanmechelen, & Broeckhove, 2014) and GPU systems (Tang & Yao, 2013;Li, Cai, & Turner, 2016). Some have proposed dedicated computer hardware for DS (Lynch & Riley, 2009).…”

A Survey on Agent-based Simulation Using Hardware Accelerators

“…The current time window is extended dynamically in work by Tang and Yao [155] to allow more events to be executed in parallel. After executing all events within the current window, their algorithm evaluates the first event in the event queue with a timestamp larger than the LBTS that can still safely be executed according to the lookahead.…”

Supporting efficient execution of continuous space agent‐based simulation on GPU

“…Besides, as reported in [20], allocations and deallocations fulfilled by CUDA's native memory allocator are prone to fault under a heavy load. The design of future event list is further improved in [23]. In [21] and [22], discrete event simulation for queuing network is performed on GPU.…”

“…This leads to a huge memory consumption. The design of future event list is further improved in . Taken the event generation rate of each thread into consideration, it proposes a balancing strategy that computes a proper segment to place new events.…”