ORCS : An Oblivious Routing Congestion Simulator
ORCS : An Oblivious Routing Congestion Simulator
复制标题
ORCS：一个不经意的路由拥塞模拟器
DOI：
10.1109/cluster.2017.18
复制
发表时间：
2009
期刊：
影响因子：
--
通讯作者：
A. Lumsdaine
中科院分区：
文献类型：
--
作者：
Timo Schneider;T. Hoefler;A. Lumsdaine
Hennessy和Patterson在[4]中定义的二等分带宽是指在最坏情况下网络的两个大小相等的部分之间的带宽，被广泛用作网络性能的理论模型。该模型给出了网络应用所经历的最小二等分带宽的上限，因为它没有考虑所使用的路由方案。事实证明，对于不同的网络拓扑，不经意的静态路由对于各种网络拓扑来说是次优的，其中每个(源、目的)对都有一条通过网络的固定路径。然而，不经意路由易于实现并且提供低延迟，因为不需要计算来路由分组，因为可以离线确定路由。因此，它被几个高性能网络[1，6，10]所使用。InfiniBand是使用不经意静态路由的互连结构之一。在[5]中，我们证明了对于足够的通信模式可以测量的有效二等分带宽显著低于二等分带宽模型预测的带宽。在我们的实验中，由于网络拥塞，没有一个被检查的InfiniBand网络能够提供超过61%的对分带宽。由于没有考虑流量控制机制等影响网络性能的因素，我们只通过模拟交通模式来研究拥塞问题，并通过测量验证了模型的正确性。为了研究拥塞对大规模簇的影响，作为FASTOS II项目的一部分，我们开发了一个框架来模拟基于不经意的目的地的路由网络中的拥塞。在这项工作中，我们将解释我们的模拟器和我们使用的相关工具。我们的模拟器框架的设计是模块化的，因此它可以扩展到模拟不同的流量模式(如第2.4节所示，已经预定义了多种模式)，或者使用不同的方法来呈现在模拟运行期间收集的数据，如第2.5节所述。我们将继续简要解释InfiniBand使用的路由方案，以及为什么使用这种路由策略的网络可能无法为应用程序提供全等分带宽，即使网络拓扑理论上能够做到这一点。第2节介绍了如何使用我们的模拟器，并将举例说明如何使用该软件包进行分析。在第3节中，我们将解释工具的使用，这些工具使用户能够使用我们的模拟器研究InfiniBand子网管理器OpenSM支持的不同路由算法。最后，我们将对模拟器实现进行描述，这将使用户在第4节中更深入地了解模拟器是如何工作的。
Bisection Bandwidth, as defined by Hennessy and Patterson in [4] as the bandwidth between the two equal sized halves of the network for the worst case partition, is widely used as a theoretical model for network performance. This model gives an upper bound for the minimal bisection bandwidth, as experienced by applications, of a network, as it does not take the used routing scheme into account. It has been proven that oblivious static routing, where there is one fixed path through the network for each (source, destination) pair, is suboptimal for various network topologies [8]. However, oblivious routing is easy to implement and delivers low latencies because no computation is needed to route packets, since the routes can be determined off-line. Therefore, it is used by several high performance networks [1, 6, 10]. InfiniBand is one of the interconnection fabrics that use oblivious static routing. In [5] we showed that the effective bisection bandwidth, that can be measured for adequate communication patterns, is significantly lower than the bandwidth predicted by the bisection bandwidth model. In our experiments none of the examined InfiniBand networks was able to deliver more than 61% of the bisection bandwidth, due to network congestion. Other effects that would deteriorate performance, for example flow control mechanisms, have not been taken into account, we only studied congestion by simulating traffic patterns and verified the correctness of our model by measurements. To study the effect of congestion on large scale clusters, a part of the FASTOS II project, we developed a framework to simulate the congestion in oblivious destination based routed networks. In this work, we will explain our simulator and the related tools that we used. The design of our simulator framework is modular, so it could be extended to simulate different traffic patterns (a variety of them is already predefined as shown in Section 2.4), or using different approaches to present the data gathered during the simulation runs as those described in Section 2.5. We will continue with a brief explanation of the routing scheme used by InfiniBand and why networks using such routing strategies might not deliver full bisection bandwidth for applications, even if the network topology is theoretically capable of doing so. Section 2 documents how our simulator can be used, and will give examples for analysis that can be performed with this software package. In Section 3, we will explain the usage of tools that enable the user to study the different routing algorithms supported by OpenSM, the InfiniBand subnet manager, with our simulator. We conclude with a description of the simulator implementation that will enable users to gain a deeper insight on how the simulator works in Section 4.