Towards exascale computing systems

拟议的研究计划将侧重于与Exascale计算有关的问题，即计算系统的性能达到目前PFlop/s系统性能的1000倍。 具体来说，我们将专注于方法，这将有助于实现这样的系统，通过性能增强技术适用于互连网络，调度和低功耗设计所有的关键问题，并在国际上深入研究。 在集群系统中，互连在将数据传递到计算中起着重要作用。我们的研究计划将开发方法，以最小的延迟将数据传输到系统的多个核心进行的计算。我们将研究通信信道的发送端和接收端。在发送端，预测将允许隐藏建立信道的开销，而在接收端，直接将有效载荷注入到存储层次结构中最接近（并且最快）将消耗它的计算的部分，以确保最小的延迟。 伴随着数据到计算的传递，是这些计算的调度，特别是在众核处理器的时代。为了获得最高效率，需要确保在特定核心上调度的计算不等待数据或指令。我们的目标是开发方法，确保计算是最佳的可用资源。 就功率而言，我们的目标是开发一个设计环境，利用计算描述中的信息自动生成低功耗系统。我们正在开发技术，以识别在计算过程中不活跃的系统部分，因此它们可以被关闭。我们的目标是这种环境的嵌入式系统，但也专门的计算引擎，将成为混合节点的一部分，因为我们采取行动，以解决的挑战exascale系统。

The proposed research program will focus on issues pertaining to Exascale Computing that is, computing systems attaining performance that is 1000x the performance of the present day PFlop/s systems. Specifically, we shall focus on methods that will contribute in the realization of such systems through performance enhancing techniques applicable to interconnection networks, scheduling, and low power design All are critical issues and are researched intensively internationally. In a cluster system, the interconnect plays an important role in delivering data to the computations. Our research program will develop methods that deliver data to the computations carried out by the many cores of the system with a minimum of latency. We shall investigate both the send and the receive sides of the communication channel. At the send side, prediction will allow the hiding of the setting-up-the-channel overhead while at the receiving end, directly injecting the payload to a section of the storage hierarchy that is the closest (and the fastest) to the computation that will consume it ensures minimal delays. Concomitant to the delivering of data to the computation, is the scheduling of these computations especially in the era of many-core processors. For maximum efficiency, one needs to ensure that a computation scheduled on a particular core is not waiting for data, or instructions. Our objective is to develop methods ensuring that computations are optimally scheduled on the available resources. As far as power is concerned, our objective is to develop a design environment that will automatically produce low power systems utilizing the information that is present within the description of the computation. We are developing techniques that identify the portions of the system that are not active during portions of the computation, so they can be switched-off. We are targeting this environment for embedded systems, but also to specialized computation engines that will be part of hybrid nodes as we move to address the challenge of exascale systems.