Yiping Yao

Experimental analysis of logical process simulation algorithms in JAMES II

The notion of logical processes is a widely used modeling paradigm in parallel and distributed discrete-event simulation. Yet, the comparison among different simulation algorithms for LP models still remains difficult. Most simulation systems only provide a small subset of available algorithms, which are usually selected and tuned towards specific applications. Furthermore, many modeling and simulation frameworks blur the boundary between model logic and simulation algorithm, which hampers the extensibility and the comparability. Based on the general-purpose modeling and simulation framework JAMES II, which has already been used for experiments with algorithms several times, we present an environment for the experimental analysis of simulation algorithms for logical processes. It separates model from simulator concepts, is extensible (in regards to the benchmark models, the algorithms used, etc.), and facilitates a fair comparison of algorithms.

An efficient algorithm in the HLA time management

The HLA time management is an important factor that limits the scalability of distributed simulations. An efficient algorithm of greatest available logical time (GALT) is thus much critical for the time management in an RTI to support large-scale simulations. The concept of GALT in IEEE 1516 was also called lower bound time stamp (LBTS) in HLA 1.3. The computation of GALT in the HLA time management is different from that of LBTS in traditional parallel discrete event simulation (PDES). In this paper, an algorithm about GALT is presented and its correctness is proved. Its efficiency is also explained by applying it to RTI1.3-NG. In fact, the algorithm has been implemented in our RTI to support thousands of federates in our cluster systems. In addition, a real-world example is introduced to explain the correctness of the algorithm proving, and the reason of our RTI supporting large-scale simulations.

Development of a runtime infrastructure for large-scale distributed simulations

With the development of distributed modeling and simulation, it is necessary for the RTI to support large-scale applications. However, many RTIs can not support large-scale distributed simulations with more than 100 federates very well nowadays. StarLink+ is an RTI developed according to the IEEE 1516 standard, which can be used for large-scale simulations with thousands of federates. Great innovations are made in StarLink+, such as its architecture and inner implementation technologies. This paper presents the two-level architecture in StarLink+. The unique architecture has the advantages of both central architecture and distributed architecture. To improve the performance much more for large-scale simulations, two important technologies, i.e. multiple threads and data packing, are adopted in StarLink+. In addition, this paper explains the efficient advancing mechanism in time management and discusses the large-scale experiments with thousands of federates in StarLink+.

Implementation of time management in a runtime infrastructure

The high level architecture (HLA) time management is concerned with mechanisms for guaranteeing message order, process synchronization and execution correctness in distributed simulations. Time management greatly influences on the scales of applications, especially for the computation of greatest available logical time (GALT) and the implementation of optimistic services. StarLink is an RTI with central architecture, which is compliant with the IEEE 1516 standard. This paper systematically describes the implementation algorithms for main time management services in StarLink. Two smart and efficient algorithms about GALT computation and optimistic services are also introduced, which are suitable for many RTIs such as RTI1.3-NG, pRTI and DRTI. For the GALT algorithm, it is not necessary for an RTI to resolve the recursion nor any deadlock. For optimistic services, a simple mechanism without rollback in an RTI is also introduced; therefore, it can greatly simplify the development of an RTI.

Data consistency in a large-scale runtime infrastructure

In order to support large scale distributed simulation, we have developed a RTI called StarLink+ with particular architecture which is compliant with IEEE 1516. StarLink+ is composed of a central RTI server and multiple local RTI servers. Each Local RTI server manages multiple federates. Data consistency has great influence on RTIs performance and scale. In StarLink+, only a small portion of data must be globally consistent for all local RTI servers. However, a great amount of data is not consistent for different local RTI servers. This paper focuses on the research of data consistency about a variety of data in StarLink+. On the one hand, we introduce the fully consistent data such as object name designation and handle assignment; on the other hand, we also study the partly consistent data such as publication and subscription, ownership transfer, and time management.

An asynchronous GVT computing algorithm in neuron time warp-multi thread

Multi-threaded Parallel Discrete Event Simulation (PDES) is promising to achieve high performance. Generally employing a collection of multi-core nodes is necessary to accomplish large scale PDES, which makes it running in a hybrid of distributed and shared memory platform. Present Global Virtual Time (GVT) computing algorithms are suitable for pure distributed or shared memory platform. In this paper we present an asynchronous GVT computing algorithm in Neuron Time Warp-Multi Thread (NTW-MT) simulator for stochastic simulation in NEURON project. GVT is computed asynchronously both within and among processes, which is the first try in multi-threaded PDES as far as we know. Then we prove this algorithm can compute a valid GVT at any wall clock time, and conclude it has less computational cost through analysing the cost and delay. Finally we show results of simulating a calcium wave model in an unbranched apical dendrite of a hippocampal pyramidal neuron.

Three-level-parallelization support framework for large-scale analytic simulation

Fully exploiting the parallelism in large-scale analytic simulation is an essential way to meet the increasing demand for computing resources. This paper deconstructs large-scale analytic simulation using a hierarchical approach. Five computational characteristics that cause the huge computing requirements of analytic simulation are summarized: “Multi-sample”, “Multi-entity”, “Running-as-fast-as-possible”, “Synchronization for constraint of causality”, and “Complex model calculation”. According to these characteristics, a “Sample, Entity, Model” three-level-Parallelization support framework is proposed to exploit the parallelism on three levels. Under the guidance of this framework, a High-Performance Simulation Computer system which integrated software management and hardware support was designed, and then applied in realistic applications. The experimental results show that the designed system can effectively utilize the potential parallelism characteristics in analytic simulation. Consequently, the simulation performance can be improved dozens or even hundreds of times.

An experimental analysis environment for logical process simulation algorithms

The notion of logical processes (LPs) is a widely used modeling paradigm in parallel and distributed discrete-event simulation (PDES). Nevertheless the comparison among different simulation algorithms for LP models still remains difficult: there are too many combinations of algorithms to be explored, often simulation systems only provide a small subset of available algorithms, and many m&s frameworks blur the boundary between model logic and simulation algorithm, which hampers extensibility and comparability. We present an environment for the experimental analysis of simulation algorithms for logical processes. It separates between model and simulator, is extensible, and facilitates a fair comparison of algorithms. We illustrate the functioning of the environment by presenting experimental results for well-known simulation algorithms and a benchmark model.

A Binary Search Enhanced Sort-based Interest Matching Algorithm

In distributed simulation, communication based on publish/subscribe will generate large amount of irrelevant data transmissions, and thereby degrading the performance. To solve the problem, HLA standard defines data distribution management to filter unnecessary communication. Among several famous interest matching algorithms, the sort-based algorithm has been proven to be the most efficient method in most scenarios. However, the potential of existing sort-based algorithm has not been fully exploited, due to the overhead of sorting the bounds can be further reduced and a portion of unnecessary bit operations can be eliminated. In this paper, we propose a binary search enhanced sort-based interest matching algorithm (BSSIM). Based on a different sufficient and necessary condition to judge interval overlapping, the size of list to be sorted can be remarkably reduced. Moreover, unnecessary bit operations can be eliminated by binary searches. Experimental results show that BSSIM algorithm outperforms the sort-based algorithm, and approximately 64%-159% performance improvement can be achieved at different scenarios.

An Integration Architecture Based on Event Graph for PDES Application Development

PDES(Parallel discrete event simulation) applications usually cover multiple domains, consequently developing such an application often needs professional skills and cooperation from multiple domain experts, which adds the difficulty of development. In addition, in most case, models nested in an application couple with each other tightly and thus become dependent on the application or the platform, which hinders its development and debugging and limits the models reusability. To make matter worse, current visual development tools cannot support integrating models from different domains into applications efficiently. Focusing on this problem, this paper proposes an integration architecture based on event graph for PDES application development, which isolates the model from the application. This proposed method separates models development the application framework development and provides an approach to plug models into the application framework, which improves the flexibility and reusability of models and efficiency of the whole application development as a result. In the end, a tool based on this method is developed and the experiment shows the proposed method can improve the efficiency of the development and provides a convenient way of debugging the application.