XMTSim: A Simulator of the XMT Many-core Architecture

This paper documents the features and the design of XMTSim, the cycle-accurate simulator of the Explicit Multi-Threading (XMT) computer architecture. The Explicit Multi-Threading (XMT) is a general-purpose many-core computing platform, with the vision of a 1000-core chip that is easy to program but does not compromise on performance. XMTSim is a primary component in its publicly available toolchain along with an optimizing compiler. Research and experimentation enabled by the toolchain played a central role in supporting the ease-of-programming and performance aspects of the XMT architecture. The compiler and the simulator are also important milestones for an efficient programmer’s workflow from PRAM algorithms to programs that run on the shared memory XMT hardware. This workflow is a key component in accomplishing the goal of ease-of-programming and performance. The applicability of the XMT simulator extends beyond specific XMT choices. It can be used to explore the much greater design space of shared memory many-cores by system researchers or by programmers. As the toolchain can practically run on any computer, it provides a supportive environment for teaching parallel algorithmic thinking with a programming component. XMTSim is the highly-configurable cycle-accurate simulator of the XMT computer architecture [38, 39, 48, 49]. It is tuned to approximate the behavior of major on-die components of XMT, such as the cores, interconnect and on-chip caches. Additionally XMTSim features a power model and a thermal model, and it provides means to simulate dynamic power and thermal management algorithms. We made XMTSim publicly available as a part of the XMT programming toolchain [7,9], which also includes an optimizing compiler [45]. Detailed information on XMT architecture and the programming model can be found in [30]. XMT envisions bringing efficient on-chip parallel programming to the mainstream, and the toolchain is instrumental in obtaining results to validate these claims, as well as making a simulated XMT platform accessible from any personal computer. XMTSim is useful to a range of communities such as system architects, teachers of parallel programming and algorithm developers due to the following four reasons: 1. Opportunity to evaluate alternative system components. XMTSim allows users to change the parameters of the simulated architecture including the number of functional units and organization of the parallel cores. It is also easy to add new functionality to the simulator, making it the ideal platform for evaluating both architectural extensions and algorithmic improvements that depend on the availability of hardware resources. For example, Caragea, et. al [8] searches for the optimal size and replacement policy for prefetch buffers given limited transistor resources. Furthermore, to our knowledge, XMTSim is the only publicly available many-core simulator that allows evaluation of architectural mechanisms/features, such as dynamic power and thermal management. Finally, the capabilities of our toolchain extend beyond specific XMT choices: system architects can use it to explore a much greater design-space of shared memory many-cores. 2. Performance advantages of XMT and PRAM algorithms. A number of publications [6, 10, 15, 16, 17, 18, 19, 41] list the performance advantages of XMT compared to exiting parallel architectures, and also document the interest of the academic community in such results. XMTSim was the enabling factor for the publications that investigate planned/future configurations. Moreover, despite past doubts in the practical relevance of PRAM algorithms, results facilitated by the toolchain showed not only that theory-based algorithms can provide good speedups in practice, but that sometimes they are the only ones to do so. 3. Teaching and experimenting with on-chip parallel programming. As a part of the XMT toolchain, XMTSim contributed to the experiments that established the ease-of-programming of XMT. These experiments were presented in publications [23, 40, 44, 46] and conducted in courses taught to graduate, undergraduate, high-school and middle-school students including at Thomas Jefferson High School, Alexandria, VA. The curriculum at Thomas Jefferson High School has featured XMT programming since 2008; more than two hundred of its students have already programmed XMT and in 2012 forty of these high-school students demonstrated Ph.D. level parallel programming [14]. In addition, the XMT toolchain provides convenient platform for teaching parallel algorithms and programming, because students can install and use it on any personal computer to work on their assignments.