On the Design of High Performance Data Center Networks.
详细信息
文摘
Massive modern data centers consisting of tens of thousands of servers have emerged to form the backbone of a variety of powerful distributed computing frameworks,in which efficient communication is often required among huge data sets stored in tens of thousands of servers across a data center. This renders the performance of the data center network DCN) essential to the successful operation of a data center. This dissertation focuses on several important issues in the design space of high performance data center networks. First,we present a systematic study on optical packet switch,which is a key component in next generation DCNs. The study can be divided into two parts. In the first part,we present an efficient analytical model for a novel hybrid optical/electrical packet switch called OpCut. In the second part,we study the multicast scheduling problem in all-optical packet switches. We propose a novel optical buffer structure,a Low Latency Multicast Scheduling LLMS) Algorithm that guarantees delay upper bound,and a pipeline and parallel architecture that enables line-rate scheduling. Second,we study how to deploy high performance multicast communication in data center networks. Multicast can tremendously benefit many cloud-based services that require one-to-many group communication,such as redirecting search queries to multiple indexing servers and replicating file chunks in distributed file systems,through releasing the sender from duplicated transmission tasks,thus,significantly improving network latency. Exploring the unique novel features and techniques in data centers,our research focuses on the following issues: 1) Achieving cost-efficient provisioning of nonblocking multicast fat-tree DCNs by exploring server redundancy in data centers； 2) Developing a practical multicast flow scheduling algorithm that ensures guaranteed flow bandwidth and high network throughput under volatile data center traffic. Third,we study the recently developed hybrid packet/circuit Hypac) switched DCN architecture,which arguments the traditional electrical packet switched EPS) network with a high-speed optical circuit switched OCS) network. Considering that the OCS/EPS networks have unique strengths and weaknesses,we propose a time-efficient Collaborative Bandwidth Allocation CBA) algorithm that configures both networks in a complementary manner. Also,we show that given sufficient bandwidth from both networks,a Hypac DCN can guarantee 100% throughput with a bounded delay using the proposed CBA algorithm. Through comprehensive evaluations,we demonstrate CBA significantly improves the performance of Hypac DCNs in many aspects. In summary,this dissertation combines algorithm design,mathematical modeling,network optimization,theoretical analysis and simulation techniques to provide a thorough investigation on the above issues. The outcome of this research would benefit many cloud applications that rely on group communication,and have a significant impact on the fundamental design principles for future DCNs.