high performance computing on graphics processing units: hgpu.org

Power and thermal constraints gained critical importance in the design of microprocessors over the past decade. Chipmakers failed to keep power at bay while sustaining the performance growth of serial computers at the rate expected by consumers. As an alternative, they turned to fitting an increasing number of simpler cores on a single die. While this is a step forward for relaxing the constraints, the issue of power is far from resolved and it is joined by new challenges which we explain next. As we move into the era of many-cores, processors consisting of 100s, even 1000s of cores, single-task parallelism is the natural path for building faster general-purpose computers. Alas, the introduction of parallelism to the mainstream general-purpose domain brings another long elusive problem to focus: ease of parallel programming. The result is the dual challenge where power efficiency and ease-of-programming are vital for the prevalence of up and coming many-core architectures. The observations above led to the lead goal of this dissertation: a first order validation of the claim that even under power/thermal constraints, ease-of-programming and competitive performance need not be conflicting objectives for a massively-parallel general-purpose processor. As our platform, we choose the eXplicit Multi-Threading (XMT) many-core architecture for fine grained parallel programs developed at the University of Maryland. We hope that our findings will be a trailblazer for future commercial products. XMT scales up to thousand or more lightweight cores and aims at improving single task execution time while making the task for the programmer as easy as possible. Performance advantages and ease-of-programming of XMT have been shown in a number of publications, including a study that we present in this dissertation. Feasibility of the hardware concept has been exhibited via FPGA and ASIC (per our partial involvement) prototypes. Our contributions target the study of power and thermal envelopes of an envisioned 1024-core XMT chip (XMT1024) under programs that exist in popular parallel benchmark suites. First, we compare XMT against an area and power equivalent commercial high-end many-core GPU. We demonstrate that XMT can provide an average speedup of 8.8x in irregular parallel programs that are common and important in general purpose computing. Even under the worst-case power estimation assumptions for XMT, average speedup is only reduced by half. We further this study by experimentally evaluating the performance advantages of Dynamic Thermal Management (DTM), when applied to XMT1024. DTM techniques are frequently used in current single and multi-core processors, however until now their effects on single-tasked many-cores have not been examined in detail. It is our purpose to explore how existing techniques can be tailored for XMT to improve performance. Performance improvements up to 46% over a generic global management technique has been demonstrated. The insights we provide can guide designers of other similar many-core architectures. A significant infrastructure contribution of this dissertation is a highly configurable cycle-accurate simulator, XMTSim. To our knowledge, XMTSim is currently the only publicly-available shared-memory many-core simulator with extensive capabilities for estimating power and temperature, as well as evaluating dynamic power and thermal management algorithms. As a major component of the XMT programming toolchain, it is not only used as the infrastructure in this work but also contributed to other publications and dissertations.