Leveraging Parallel Bit Stream and Dynamic Compilation Technology for Big Data Applications

点击翻译按钮获取中文摘要

The proposed research continues the investigation of parallel bit stream technology as a vehicle for dramatically accelerating text processing applications, while incorporating additional technologies to broaden the scope and scale of applications that can be tackled. In particular, extensions to the research along two dimensions are anticipated. The first is to incorporate dynamic***compilation technologies building on the highly regarded LLVM compiler infrastructure. One important research goal of this extension is to investigate the extent to which parallel bit stream and dynamic compilation technology can be applied to develop accelerated text matching and searching algorithms for problems expressed using dynamic instances of text patterns. A second research goal in using dynamic compilation technology is to investigate the development of generic distributable software binaries that can dynamically adapt to take advantage of particular processor architectural features available at runtime. For example, can a generic x86-64 binary be employed in cloud computing deployments to take advantage of Intel AVX2 instructions when available on a particular execution processor, falling back to default SSE2 code in other cases?******The second dimension of the proposed research is to investigate how big data applications involving text processing can be accelerated by combining the bitwise data parallelism of parallel bit stream technology with other forms of chip-, cluster- and grid-level parallelism. In bringing parallel processors to bear in solving big data problems, the holy grail is to be able to achieve an N-fold acceleration of processing using N processors. In general, getting close to this goal is difficult, except perhaps for problems that are known as "embarrassingly parallel" (fundamentally parallel in nature). Limits on scalability of solutions often mean that there are diminishing returns with the addition of processors. However, parallel bit stream and dynamic compilation technologies offer the possibility of considerably increasing the work achieved per processor in N-processor systems. Furthermore, the bitwise data parallelism of parallel bit stream technology presents opportunities for both pipeline and coarse-grained data parallelism on chip multicore processors, GPUs and clusters. The proposed research then is to investigate the techniques for leveraging parallel bit stream technology and assessing the corresponding benefits that may be achieved.**