GPRM: a high performance programming framework for manycore processors

Processors with large numbers of cores are becoming commonplace. In order to utilise the available resources in such systems, the programming paradigm has to move towards increased parallelism. However, increased parallelism does not necessarily lead to better performance. Parallel programming models have to provide not only flexible ways of defining parallel tasks, but also efficient methods to manage the created tasks. Moreover, in a general-purpose system, applications residing in the system compete for the shared resources. Thread and task scheduling in such a multiprogrammed multithreaded environment is a significant challenge. In this thesis, we introduce a new task-based parallel reduction model, called the Glasgow Parallel Reduction Machine (GPRM). Our main objective is to provide high performance while maintaining ease of programming. GPRM supports native parallelism; it provides a modular way of expressing parallel tasks and the communication patterns between them. Compiling a GPRM program results in an Intermediate Representation (IR) containing useful information about tasks, their dependencies, as well as the initial mapping information. This compile-time information helps reduce the overhead of runtime task scheduling and is key to high performance. Generally speaking, the granularity and the number of tasks are major factors in achieving high performance. These factors are even more important in the case of GPRM, as it is highly dependent on tasks, rather than threads. We use three basic benchmarks to provide a detailed comparison of GPRM with Intel OpenMP, Cilk Plus, and Threading Building Blocks (TBB) on the Intel Xeon Phi, and with GNU OpenMP on the Tilera TILEPro64. GPRM shows superior performance in almost all cases, only by controlling the number of tasks. GPRM also provides a low-overhead mechanism, called “Global Sharing”, which improves performance in multiprogramming situations. We use OpenMP, as the most popular model for shared-memory parallel programming as the main GPRM competitor for solving three well-known problems on both platforms: LU factorisation of Sparse Matrices, Image Convolution, and Linked List Processing. We focus on proposing solutions that best fit into the GPRM’s model of execution. GPRM outperforms OpenMP in all cases on the TILEPro64. On the Xeon Phi, our solution for the LU Factorisation results in notable performance improvement for sparse matrices with large numbers of small blocks. We investigate the overhead of GPRM’s task creation and distribution for very short computations using the Image Convolution benchmark. We show that this overhead can be mitigated by combining smaller tasks into larger ones. As a result, GPRM can outperform OpenMP for convolving large 2D matrices on the Xeon Phi. Finally, we demonstrate that our parallel worksharing construct provides an efficient solution for Linked List processing and performs better than OpenMP implementations on the Xeon Phi. The results are very promising, as they verify that our parallel programming framework for manycore processors is flexible and scalable, and can provide high performance without sacrificing productivity.

Data structures for SIMD logic simulation

Due to the growth of design size and complexity, design verification is an important aspect of the Logic Circuit development process. The purpose of verification is to validate that the design meets the system requirements and specification. This is done by either functional or formal verification. The most popular approach to functional verification is the use of simulation based techniques. Using models to replicate the behaviour of an actual system is called simulation. In this thesis, a software/data structure architecture without explicit locks is proposed to accelerate logic gate circuit simulation. We call thus system ZSIM. The ZSIM software architecture simulator targets low cost SIMD multi-core machines. Its performance is evaluated on the Intel Xeon Phi and 2 other machines (Intel Xeon and AMD Opteron). The aim of these experiments is to: • Verify that the data structure used allows SIMD acceleration, particularly on machines with gather instructions ( section 5.3.1). • Verify that, on sufficiently large circuits, substantial gains could be made from multicore parallelism ( section 5.3.2 ). • Show that a simulator using this approach out-performs an existing commercial simulator on a standard workstation ( section 5.3.3 ). • Show that the performance on a cheap Xeon Phi card is competitive with results reported elsewhere on much more expensive super-computers ( section 5.3.5 ). To evaluate the ZSIM, two types of test circuits were used: 1. Circuits from the IWLS benchmark suit [1] which allow direct comparison with other published studies of parallel simulators.2. Circuits generated by a parametrised circuit synthesizer. The synthesizer used an algorithm that has been shown to generate circuits that are statistically representative of real logic circuits. The synthesizer allowed testing of a range of very large circuits, larger than the ones for which it was possible to obtain open source files. The experimental results show that with SIMD acceleration and multicore, ZSIM gained a peak parallelisation factor of 300 on Intel Xeon Phi and 11 on Intel Xeon. With only SIMD enabled, ZSIM achieved a maximum parallelistion gain of 10 on Intel Xeon Phi and 4 on Intel Xeon. Furthermore, it was shown that this software architecture simulator running on a SIMD machine is much faster than, and can handle much bigger circuits than a widely used commercial simulator (Xilinx) running on a workstation. The performance achieved by ZSIM was also compared with similar pre-existing work on logic simulation targeting GPUs and supercomputers. It was shown that ZSIM simulator running on a Xeon Phi machine gives comparable simulation performance to the IBM Blue Gene supercomputer at very much lower cost. The experimental results have shown that the Xeon Phi is competitive with simulation on GPUs and allows the handling of much larger circuits than have been reported for GPU simulation. When targeting Xeon Phi architecture, the automatic cache management of the Xeon Phi, handles and manages the on-chip local store without any explicit mention of the local store being made in the architecture of the simulator itself. However, targeting GPUs, explicit cache management in program increases the complexity of the software architecture. Furthermore, one of the strongest points of the ZSIM simulator is its portability. Note that the same code was tested on both AMD and Xeon Phi machines. The same architecture that efficiently performs on Xeon Phi, was ported into a 64 core NUMA AMD Opteron. To conclude, the two main achievements are restated as following: The primary achievement of this work was proving that the ZSIM architecture was faster than previously published logic simulators on low cost platforms. The secondary achievement was the development of a synthetic testing suite that went beyond the scale range that was previously publicly available, based on prior work that showed the synthesis technique is valid.