III: Medium: Massively Parallel Data Analytics on Heterogeneous Architectures

Massachusetts Institute Of Technology, Cambridge MA

Investigators

Abstract

The rise of connected devices and the Internet has led to unprecedented growth in the volumes of data that computers must store and process. This enormous growth in data volumes coincides with a growing demand for immediate answers and interactive analytics. Increasingly companies need real-time reports of sales, network traffic, anomalies, and other business trends. Internet-connected devices, like cars and industrial plants, demand even more real-time analysis of data. These trends mean database and data analytics platforms must deliver ever-faster performance from machines, a fact that has driven the dramatic interest in scalable multi-node processing systems, like Hadoop and Spark, which distribute the processing of large data sets across clusters of machines. Unfortunately, because of the way these platforms are engineered, they provide shockingly poor utilization of the hardware resources on each node, often times yielding single-node throughput that is thousands of times lower than what the raw hardware is capable of. In this project, an orthogonal direction will be pursued; a system, called Proteus, will be built that will obtain performance that utilizes hardware to the fullest extent possible, focusing on yielding a scalable system that fully utilizes all available computing resources. If successful, this project will have broad impact because databases and data-intensive parallel computing systems are used by millions of enterprises around the world, both on-site and in computing clouds; optimized implementations of these systems that better exploit hardware will improve response times and reduce hardware and energy costs, resulting in billions of dollars of cost savings. Proteus will parallelize across many cores on a single processor, as well as take advantages of many-core systems such as GPUs and Intel's Xeon Phi. In addition, Proteus will also be able to exploit large diverse clusters of hardware, but the aim is to do that without giving up this efficiency, rather than accepting inefficiency as a given of distributed computing. To do this, research in the Proteus project will focus on four key areas: (1) Developing optimized implementations of individual database algorithms, such as top-k sorts, sequential scans, random lookups, graph and machine learning algorithms for GPUs and CPUs. (2) Building cost models that predict the performance of these algorithms on heterogeneous architectures. (3) Developing intermediate languages that abstract details of the underlying hardware, to hide the nuances of these different platforms to but without giving up performance. (4) Building an optimizer that uses cost models to place these plans onto a heterogeneous mix of hardware to obtain the best overall performance for each query plan. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.