Design of supports for high performance optimistic simulations with special focus on multi-core hosting environments

Parallel Discrete Event Simulation (PDES) is a classical means to achieve high performance simulations. This is especially important on contexts entailing real-time constraints, as for the case of decision making tools, where target configurations have to be identified, via simulation, within bounded time. PDES is based on splitting the model to simulate in different simulation objects, that are mapped onto logical processes (LPs), which can process simulation events in parallel, thus potentially allowing for (large) speedups when compared to traditional serial execution of the simulation model. On the other hand, synchronization mechanisms need to be actuated in order to allow the parallel run to provide correct results, hence correct evolution of the objects’ state along the simulation time axis. In this dissertation we cope with performance of PDES systems, with special focus on deploys of the PDES platforms onto multi-core architectures. These architectures represent the current CPU trend and nowadays it is common to have chips with a significant number of cores, sometimes enough for effectively performing parallel computations within a single chip. By design, these architectures share most of the CPU internal components across different cores, and physical memory across the different CPUs. This implies several drawbacks, mainly due to the contention, but at the same time, such a sharing can become a resource, if well exploited through sector specific software design. The base building block for our work is the optimistic PDES synchronization paradigm, which has been shown to be highly promising in terms of potential for fruitful parallelism exploitation. As a first contribution, we will present the design and implementation of optimized supports for state log/restore operations, which are at the base of the rollback procedures used to guarantee consistency in the optimistic paradigm. System requirements for these supports only entail the x86 instruction set, with no particular additional facilities to be offered by the underlying computing platform. On the other hand, the provided solution has a further reflection on performance for deploys on multi-core systems since it aims at automatizing the optimization of a set of CPU/memory tradeoffs among which we also find virtual memory usage, which may impact physical memory (e.g. cache) contention for the case of multi-core architectures. Secondly, we address the issue of optimizing the behavior of the caching hierarchy by providing an innovative buffer delivery mechanism suited for memory demands in optimistic PDES systems. This is a quite relevant aspect to cope with, given the intrinsic high-demand of virtual memory by optimistic PDES runs. The proposed solution tends to reduce the impact of virtual memory demand on physical memory, which is achieved in a highly general manner, thus making the approach applicable to differentiated types of hardware architectures. However, by nature, it reveals particularly suited for the case of multi-core systems, given the aforementioned issue if physical memory contention across the cores. Finally, we provide an additional contribution, specifically tailored to the multi-core architecture, which introduces an innovative load-sharing paradigm for optimistic PDES platforms. It is aimed at improving performance and fruitful resource usage in a fully orthogonal manner with respect to traditional load-balancing. In fact, it allows the computing resources to be (dynamically) assigned to each simulation platform instance on the basis of its current workload, instead of migrating the workload from one kernel to another. This also has the advantage of not suffering from all the issues related to workload (namely LPs) remapping (i.e. data and metadata migration), leading to the simplification of the job of both application programmers and simulation-system developers. To assess the real benefits, all the above approaches have been implemented within ROOT-Sim, an open source C-based Optimistic Simulation platform targeting the optimistic synchronization paradigm, and several experimental studies have been conducted, whose outcomes are reported in the thesis.