Analysis, verification and design for energy aware computation

千兆赫的竞争结束了。二十多年来，微处理器的时钟速度和性能每隔一两年就会翻一番。在过去的两年里，这种指数增长趋势已经被缓慢攀升所取代。热量是罪魁祸首——推动cpu提高速度的技术趋势也推动了它们的功耗上升。热量限制了台式电脑的中央处理器和图形处理器；电源限制了笔记本电脑、pda和手机的电池寿命；而像谷歌和e-bay这样的大型计算机公司，其装满计算机的仓库的空调也很有限。并行是最有希望的解决方案。使芯片工作的晶体管的物理特性意味着，如果一个任务可以在多个CPU之间进行分配，那么总能量就会比在单个CPU上完成要少得多。这就是为什么每个计算机制造商都推出多核处理器的原因。为了有效地利用并行处理器，程序员需要了解并行算法的权衡。虽然已经提出了许多并行计算模型，但我们知道没有一个模型考虑到上述能量和时间之间的基本权衡。我们已经开始研究这个问题，并发现了一些令人惊讶的结果。例如，我们可以展示并行算法的例子，它们执行更多的操作，但比顺序算法使用更少的时间和能量。我们建议发展这种能量感知算法设计理论。我们将看看它如何应用于生物学和网络搜索中的大规模计算问题。我们还将把这与我们正在进行的集成电路设计研究联系起来。通过将理论建模与实际设计相结合，我们期望产生对理论家有吸引力的结果，并对计算机系统设计人员有用。

The gigahertz race is over.  For over two decades, microprocessors have been doubling in clock speed and performance every one or two years.  In the past two years, this exponential trend has been replaced by a slow climb.  Heat is the culprit -- the technological trends that drove CPUs up in speed also drove them up in power consumption.  Heat limits the CPU and graphics processor in a desktop computer; power limits the battery life of laptop computers, PDAs, and cell phones; and large computing companies such as Google and e-bay are limited airconditioning for their warehouses full of computers.  Parallelism is the most promising solution.  The physics of the transistors that make the chips work mean that if a task can be divided between multiple CPUs, the total energy can be much less than if the task were done on a single CPU.  This is why every computer manufacturer is coming out with multi-core processors.To make effective use of parallel processors, programmers need to understand the trade-offs of parallel algorithms.  While there have been many models proposed for parallel computation, we are aware of none that take into account the fundamental trade-off between energy and time described above.  We have started looking at this problem and have found some surprising results.  For example, we can show examples of parallel algorithms that perform more operations but use less time and energy than their sequential counterparts.We propose to develop this theory of energy aware algorithm design.  We will look at how it applies to large-scale computational problems from biology and web-searching.  We will also connect this with our ongoing research into integrated circuit design.  By combining theoretical modeling with practical design, we expect to produce results that are appealing to the theoreticians and useful to computer system designers.