RHODA Topology Configuration Using Bayesian Optimization
RHODA Topology Configuration Using Bayesian Optimization
复制标题
DOI：
10.1007/978-3-030-38085-4_12
复制
发表时间：
2019-05
影响因子：
3.7
通讯作者：
Maotong Xu;Min Tian;E. Modiano;S. Subramaniam
中科院分区：
文献类型：
--
作者：
Maotong Xu;Min Tian;E. Modiano;S. Subramaniam
数据中心流量的快速增长要求数据中心网络具有可扩展性、高能效和低延迟性。光波分复用（WDM）是一种很有前途的数据中心技术，可用于构建包含数百万台服务器的数据中心。在[24]中，提出了一种基于wdm的可重构分层光学DCN架构（RHODA），该架构可容纳多达1000多万台服务器和各种流量模式。RHODA还通过广泛使用无源光学器件，最大限度地减少耗电和昂贵器件的使用，节省了大量的功率和成本。RHODA通过服务器机架的可重构集群实现高吞吐量。在本文中，我们着重于集群拓扑（也称为集群间网络）的设计。给定成对集群流量，我们对集群拓扑的目标是最小化平均跳长。在[24]中，使用了匈牙利算法的一种简单变体，它可以最大化集群之间的单跳流量或直接流量。在本文中，我们利用贝叶斯优化（BO）框架，提出了一种快速算法来最小化RHODA集群间网络的平均跳数。据我们所知，这是第一篇使用BO来优化光学DCN性能的论文。我们提出了基于网络约束的BO的设计决策和修改。结果表明，BO可以获得最优或接近最优的结果，并且分别优于著名的规则拓扑（Gemnet）和基于匈牙利的方法。
The rapid growth of data center traffic requires data center networks (DCNs) to be scalable, energy-efficient, and provide low latencies. Optical Wavelength Division Multiplexing (WDM) is a promising technique to build data centers comprising millions of servers. In [24], a WDM-based Reconfigurable Hierarchical Optical DCN Architecture (RHODA) was presented, which can accommodate up to 10+ million of servers and a variety of traffic patterns. RHODA also saves tremendous amounts of power and cost through its extensive use of passive optical devices, and minimal use of power-hungry and costly devices. RHODA achieves high throughput through reconfigurable clustering of racks of servers. In this paper, we focus on the design of the cluster topology (also called inter-cluster network). Given the pair-wise cluster traffic, our objective for the cluster topology is to minimize the average hop length. In [24], a simple variant of the Hungarian algorithm that maximizes the one-hop or direct traffic among clusters was used. In this paper, we leverage the Bayesian Optimization (BO) framework and propose a fast algorithm to minimize the average number of hops in the inter-cluster network of RHODA. To the best of our knowledge, this is the first paper that employs BO to optimize optical DCN performance. We present our design decisions and modifications to BO based on the network constraints. Results show that BO can achieve optimal or near-optimal results, and outperforms a well-known regular topology (Gemnet) and the Hungarian-based method by up toand, respectively.