high performance computing on graphics processing units: hgpu.org

This work focuses on compiler and run-time techniques for improving the productivity and the performance portability of general-purpose parallel programming. More specifically, we focus on shared-memory task-parallel languages, where the programmer explicitly exposes parallelism in the form of short tasks that may outnumber the cores by orders of magnitude. The compiler, the run-time, and the platform (henceforth the system) are responsible for harnessing this unpredictable amount of parallelism, which can vary from none to excessive, towards efficient execution. The challenge arises from the aspiration to support fine-grained irregular computations and nested parallelism. This work is even more ambitious by also aspiring to lay the foundations to efficiently support declarative code, where the programmer exposes all available parallelism, using high-level language constructs such as parallel loops, reducers or futures. The appeal of declarative code is twofold for general-purpose programming: it is often easier for the programmer who does not have to worry about the granularity of the exposed parallelism, and it achieves better performance portability by avoiding overfitting to a small range of platforms and inputs for which the programmer is coarsening. Furthermore, PRAM algorithms, an important classof parallel algorithms, naturally lend themselves to declarative programming, so supporting it is a necessary condition for capitalizing on the wealth of the PRAM theory. Unfortunately, declarative codes often expose such an overwhelming number of fine-grained tasks that existing systems fail to deliver performance. Our contributions can be partitioned into three components. First, we tackle the issue of coarsening, which declarative code leaves to the system. We identify two goals of coarsening and advocate tackling them separately, using static compiler transformations for one and dynamic run-time approaches for the other. Additionally, we present evidence that the current practice of burdening the programmer with coarsening either leads to codes with poor performance-portability, or to a significantly increased programming effort. This is a "show-stopper" for general-purpose programming. To compare the performance portability among approaches, we define an experimental framework and two metrics, and we demonstrate that our approaches are preferable. We close the chapter on coarsening by presenting compiler transformations that automatically coarsen some types of very fine-grained codes. Second, we propose Lazy Scheduling, an innovative run-time scheduling technique that infers the platform load at run-time, using information already maintained. Based on the inferred load, Lazy Scheduling adapts the amount of available parallelism it exposes for parallel execution and, thus, saves parallelism overheads that existing approaches pay. We implement Lazy Scheduling and present experimental results on four different platforms. The results show that Lazy Scheduling is vastly superior for declarative codes and competitive, if not better, for coarsened codes. Moreover, Lazy Scheduling is also superior in terms of performance-portability, supporting our thesis that it is possible to achieve reasonable efficiency and performance portability with declarative codes. Finally, we also implement Lazy Scheduling on XMT, an experimental many-core platform developed at the University of Maryland, which was designed to sup-port codes derived from PRAM algorithms. On XMT, we manage to harness the existing hardware support for scheduling at parallelism to compose it with Lazy Scheduling, which supports nested parallelism. In the resulting hybrid scheduler, the hardware and software work in synergy to overcome each other’s weaknesses. We show the performance composability of the hardware and software schedulers, both in an abstract cost model and experimentally, as the hybrid always performs better than the software scheduler alone. Furthermore, the cost model is validated by using it to predict if it is preferable to execute a code sequentially, with outer parallelism, or with nested parallelism, depending on the input, the available hardware parallelism and the calling context of the parallel code.