摘要 :
Stellar datacenter networks are a recent generic construction designed to transform a base-graph into a dual-port, server-centric datacenter network. We prove that the S-bisection width of any stellar datacenter network can be obt...
展开
Stellar datacenter networks are a recent generic construction designed to transform a base-graph into a dual-port, server-centric datacenter network. We prove that the S-bisection width of any stellar datacenter network can be obtained from the solution of isoperimetric problems on the base-graph, provided that the base-graph is regular. We extend previous research on the stellar datacenter networks GQ*, instantiated with generalized hypercubes, and show that with respect to S-bisection width, GQ* performs well in comparison with the dual-port datacenter network FiConn. Our work develops a strong combinatorial link between graph bisection width and throughput metrics for stellar datacenter networks.
收起
摘要 :
Most datacenter network designs overwhelmingly use expensive and power-consuming electronic switches or expensive active optical switches with long reconfiguration time. In this paper, we explore architectural solutions to leverag...
展开
Most datacenter network designs overwhelmingly use expensive and power-consuming electronic switches or expensive active optical switches with long reconfiguration time. In this paper, we explore architectural solutions to leverage the design elements of Passive Optical Cross-Connection Networks with Multiple Planes (POXN/MPs) and Passive Optical Cross-Connection Networks with Multiple Planes and Bundled Ports (PDXN/MP-BPs), both of which consist primarily of passive optical fabrics and optical transceivers that replace groups of switches in hierarchical networks. Through simple physical interconnections, our proposed architectures allow datacenter network (DCN) to incrementally scale out in network capacity. From developed formulas for calculating cost and power consumption, we demonstrate that POWMP-BPs can significantly reduce the cost and power consumption of datacenter networks compared to the traditional DCNs. To lower overhead and adapt to the types of real datacenter scenarios that are possible with PDXN/MP-BPs, we propose the new Multiple Channels with Bundled Ports Distributed Access Protocol (MCBDAP), which outperforms the Multiple Channels Distributed Access Protocol (MCDAP) for PDXN/MP in terms of bandwidth efficiency, especially for those applications involving higher proportions of inter-rack traffic than intra-rack traffic.
收起
摘要 :
In an optical inter-datacenter (inter-DC) network, for preventing data loss, a cloud system usually leverages multiple DCs for obtaining sufficient data redundancy. In order to improve the data-transfer efficiency of the regular D...
展开
In an optical inter-datacenter (inter-DC) network, for preventing data loss, a cloud system usually leverages multiple DCs for obtaining sufficient data redundancy. In order to improve the data-transfer efficiency of the regular DC backup, this paper investigates fast and coordinated data backup in geographically distributed (geo-distributed) optical inter-DC networks. By considering a mutual backup model, in which DCs can serve as the backup sites of each other, we study how to finish the regular DC backup within the shortest time duration (i.e., DC backup window (DC-B-Wnd)). Specifically, we try to minimize DC-B-Wnd with joint optimization of the backup site selection and the data-transfer paths. An integer linear programming (ILP) model is first formulated, and then we propose several heuristics to reduce the time complexity. Moreover, in order to explore the tradeoff between DC-B-Wnd and operational complexity, we propose heuristics based on adaptive reconfiguration (AR). Extensive simulations indicate that among all the proposed heuristics, AR-TwoStep-ILP achieves the best tradeoff between DC-B-Wnd and operational complexity and it is also the most time-efficient one.
收起
摘要 :
Datacenter federations are able to manage appropriately the green energy resources available in each datacenter (DC) thanks to their geographically distributed infrastructure, thus reducing energy expenditure. Scheduling algorithm...
展开
Datacenter federations are able to manage appropriately the green energy resources available in each datacenter (DC) thanks to their geographically distributed infrastructure, thus reducing energy expenditure. Scheduling algorithms can compute virtual machine migration, transferring huge amounts of raw data from one DC to another to minimize operational costs and ensuring a certain Quality of Experience. Because green energy availability greatly depends on weather conditions, in this work we present a statistical model to improve green solar energy availability estimation accuracy and we use it in a mixed integer linear programming formulation to compute optimal virtual machine placement. Optical connections can be used to provide connectivity services of enough capacity to support those migrations. In particular, elastic optical networks can provide connections with multi-granular bitrate, which can be adapted on demand. DC resource managers can request optical connections and control their capacity. However, that scheme involves the resource managers to implement algorithms and interfaces to deal with network specifics and complexity. To solve that issue, in this paper we propose coordinating transfer-based inter-DC connectivity services; inter-DC connectivity is requested in terms of volume of data and completion time. We analyze cost savings when each connectivity model is applied in a DC federation. For the sake of a compelling analysis, exhaustive simulation experiments are carried out considering realistic scenarios. Results show that the notification-based model can save up to 20 % of energy costs and more than 40 % of communication costs in the evaluated scenarios.
收起
摘要 :
The huge energy consumption of datacenters providing cloud services over the Internet has motivated different studies regarding cost savings in datacenters. Since energy expenditure is a predominant part of the total operational e...
展开
The huge energy consumption of datacenters providing cloud services over the Internet has motivated different studies regarding cost savings in datacenters. Since energy expenditure is a predominant part of the total operational expenditures for datacenter operators, energy aware policies for minimizing datacenters' energy consumption try to minimize energy costs while guaranteeing a certain quality of experience (QpE). Federated datacenters can take advantage of its geographically distributed infrastructure by managing appropriately the green energy resources available in each datacenter at a given time, in combination with workload consolidation and virtual machine migration policies. In this scenario, inter-datacenter networks play an important role and communications cost must be considered when minimizing operational expenditures. In this work we tackle the Elastic Operations in Federated Datacenter for Performance and Cost Optimization (ELFADO) problem for scheduling workload orchestrating federated datacenters. Two approaches, distributed and centralized, are studied and integer linear programming (ILP) formulations and heuristics are provided. Using those heuristics, we analyze cost savings with respect to a fixed workload placement. For the sake of a compelling analysis, exhaustive simulation experiments are carried out considering realistic scenarios. Results show that the centralized ELFADO approach can save up to 52% of energy cost and more than 44% when communication costs are also considered.
收起
摘要 :
The soaring demands for always-on and fast-response online services have driven modern datacenter networks to undergo tremendous growth. These networks often rely on scale-out designs with large numbers of commodity switches to re...
展开
The soaring demands for always-on and fast-response online services have driven modern datacenter networks to undergo tremendous growth. These networks often rely on scale-out designs with large numbers of commodity switches to reach immense capacity while keeping capital expenses under check. The downside is more devices means more failures, raising a formidable challenge for network operators to promptly handle these failures with minimal disruptions to the hosted services. Recent research efforts have focused on automatic failure localization. Yet, resolving failures still requires significant human interventions, resulting in prolonged failure recovery time. Unlike previous work, NetPilot aims to quickly mitigate rather than resolve failures. NetPilot mitigates failures in much the same way operators do - by deactivating or restarting suspected offending components. NetPilot circumvents the need for knowing the exact root cause of a failure by taking an intelligent trial-and-error approach. The core of NetPilot is comprised of an Impact Estimator that helps guard against overly disruptive mitigation actions and a failure-specific mitigation planner that minimizes the number of trials. We demonstrate that NetPilot can effectively mitigate several types of critical failures commonly encountered in production datacenter networks.
收起
摘要 :
Emerging optical communication technologies support the dynamic reconfiguration of datacenter network topologies depending on the traffic they serve. However, to reap the benefits of such demand-aware networks, control logic that ...
展开
Emerging optical communication technologies support the dynamic reconfiguration of datacenter network topologies depending on the traffic they serve. However, to reap the benefits of such demand-aware networks, control logic that quickly learns and adapts to traffic patterns is required. This paper presents CacheNet, a novel approach to efficiently control demand-aware networks. CacheNet consists of two components, a demand-aware links-cache, and a demand-oblivious topology. CacheNet leverages temporal and spatial locality in the traffic by managing the reconfigurable links of the optical switches as a links-cache. Network traffic, in turn, can be served either by a link from the links-cache component or by a demand-oblivious topology component. We study several classic caching algorithms like online LFU and LRU as our caching algorithms, as well as offline optimal caching as a benchmark, and provide an analytical model which captures their performance benefits compared to an all demand-oblivious topology. Our analytical results show that based on the hit ratios and the links-cache size, when considering the average packet delay, our hybrid design outperforms a design that is based only on demand-oblivious topology. We also evaluate CacheNet empirically, using both synthetic and real-world traffic traces, confirming the potential of our approach to consider reconfigurable links as a network of links-cache.
收起
摘要 :
Several application domains are collecting data using Internet of Things sensing devices and shipping it to remote cloud datacenters for analysis (fusion, storage, and processing). Data analytics activities raise a new set of tech...
展开
Several application domains are collecting data using Internet of Things sensing devices and shipping it to remote cloud datacenters for analysis (fusion, storage, and processing). Data analytics activities raise a new set of technical challenges from the perspective of ensuring end-to-end security and privacy of data as it travels from an edge datacenter (EDC) to a cloud datacenter (CDC) (or vice versa). This article discusses the security threats in EDCs and CDCs by dividing the complete network structure into three layers: perception layer, network layer, and application layer.
收起
摘要 :
We present Statesman, a network-state management service that allows multiple network management applications to operate independently, while maintaining network-wide safety and performance invariants. Network state captures vario...
展开
We present Statesman, a network-state management service that allows multiple network management applications to operate independently, while maintaining network-wide safety and performance invariants. Network state captures various aspects of the network such as which links are alive and how switches are forwarding traffic. Statesman uses three views of the network state. In observed state, it maintains an up-to-date view of the actual network state. Applications read this state and propose state changes based on their individual goals. Using a model of dependencies among state variables, Statesman merges these proposed states into a target state that is guaranteed to maintain the safety and performance invariants. It then updates the network to the target state. Statesman has been deployed in ten Microsoft Azure datacenters for several months, and three distinct applications have been built on it. We use the experience from this deployment to demonstrate how Statesman enables each application to meet its goals, while maintaining network-wide invariants.
收起
摘要 :
Numerous studies have shown that datacenter computers rarely operate at full utilization, leading to a number of proposals for creating servers that are energy proportional with respect to the computation that they are performing....
展开
Numerous studies have shown that datacenter computers rarely operate at full utilization, leading to a number of proposals for creating servers that are energy proportional with respect to the computation that they are performing. In this paper, we show that as servers themselves become more energy proportional, the data-center network can become a significant fraction (up to 50%) of cluster power. In this paper we propose several ways to design a high-performance datacenter network whose power consumption is more proportional to the amount of traffic it is moving- that is, we propose energy proportional datacenter networks.
We first show that a flattened butterfly topology itself is inherently more power efficient than the other commonly proposed topology for high-performance datacenter networks. We then exploit the characteristics of modern plesiochronous links to adjust their power and performance envelopes dynamically. Using a network simulator, driven by both synthetic workloads and production data-center traces, we characterize and understand design tradeoffs, and demonstrate an 85% reduction in power - which approaches the ideal energy-proportionality of the network.
Our results also demonstrate two challenges for the designers of future network switches: 1) We show that there is a significant power advantage to having independent control of each unidirectional channel comprising a network link, since many traffic patterns show very asymmetric use, and 2) system designers should work to optimize the high-speed channel designs to be more energy efficient by choosing optimal data rate and equalization technology. Given these assumptions, we demonstrate that energy proportional datacenter communication is indeed possible.
收起