T. S. Eugene Ng

The Impact of Virtualization on Network Performance of Amazon EC2 Data Center

Cloud computing services allow users to lease computing resources from large scale data centers operated by service providers. Using cloud services, users can deploy a wide variety of applications dynamically and on-demand. Most cloud service providers use machine virtualization to provide flexible and cost-effective resource sharing. However, few studies have investigated the impact of machine virtualization in the cloud on networking performance. In this paper, we present a measurement study to characterize the impact of virtualization on the networking performance of the Amazon Elastic Cloud Computing (EC2) data center. We measure the processor sharing, packet delay, TCP/UDP throughput and packet loss among Amazon EC2 virtual machines. Our results show that even though the data center network is lightly utilized, virtualization can still cause significant throughput instability and abnormal delay variations. We discuss the implications of our findings on several classes of applications.

c-Through: part-time optics in data centers

Data-intensive applications that operate on large volumes of data have motivated a fresh look at the design of data center networks. The first wave of proposals focused on designing pure packet-switched networks that provide full bisection bandwidth. However, these proposals significantly increase network complexity in terms of the number of links and switches required and the restricted rules to wire them up. On the other hand, optical circuit switching technology holds a very large bandwidth advantage over packet switching technology. This fact motivates us to explore how optical circuit switching technology could benefit a data center network. In particular, we propose a hybrid packet and circuit switched data center network architecture (or HyPaC for short) which augments the traditional hierarchy of packet switches with a high speed, low complexity, rack-to-rack optical circuit-switched network to supply high bandwidth to applications. We discuss the fundamental requirements of this hybrid architecture and their design options. To demonstrate the potential benefits of the hybrid architecture, we have built a prototype system called c-Through. c-Through represents a design point where the responsibility for traffic demand estimation and traffic demultiplexing resides in end hosts, making it compatible with existing packet switches. Our emulation experiments show that the hybrid architecture can provide large benefits to unmodified popular data center applications at a modest scale. Furthermore, our experimental experience provides useful insights on the applicability of the hybrid architecture across a range of deployment scenarios.

A hierarchical fair service curve algorithm for link-sharing, real-time and priority services

In this paper, we study hierarchical resource management models and algorithms that support both link-sharing and guaranteed real-time services with decoupled delay (priority) and bandwidth allocation. We extend the service curve based QoS model, which defines both delay and bandwidth requirements of a class, to include fairness, which is important for the integration of real-time and hierarchical link-sharing services. The resulting Fair Service Curve link-sharing model formalizes the goals of link-sharing and real-time services and exposes the fundamental tradeoffs between these goals. In particular, with decoupled delay and band-width allocation, it is impossible to simultaneously provide guaranteed real-time service and achieve perfect link-sharing. We propose a novel scheduling algorithm called Hierarchical Fair Service Curve (H-FSC) that approximates the model closely and efficiently. The algorithm always guarantees the performance for leaf classes, thus ensures real-time services, while minimizing the discrepancy between the actual services provided to the interior classes and the services defined by the Fair Service Curve link-sharing model. We have implemented the H-FSC scheduler in the NetBSD environment. By performing simulation and measurement experiments, we evaluate the link-sharing and real-time performances of H-FSC, and determine the computation over-head.

Programming your network at run-time for big data applications

Recent advances of software defined networking and optical switching technology make it possible to program the network stack all the way from physical topology to flow level traffic control. In this paper, we leverage the combination of SDN controller with optical switching to explore the tight integration of application and network control. We particularly study the run-time network configuration for big data applications to jointly optimize application performance and network utilization. We use Hadoop as an example to discuss the integrated network control architecture, job scheduling, topology and routing configuration mechanisms for Hadoop jobs. Our analysis suggests that such an integrated control has great potential to improve application performance with relatively small configuration overhead. We believe our study shows early promise of achieving the long-term goal of tight network and application integration using SDN.

Towards global network positioning

We propose a new approach to predict Internet network distance called Global Network Positioning (GNP). This approach models the Internet as a geometric space and distributedly computes geometric coordinates to characterize the positions of hosts in the Internet. By conducting Internet experiments, we show that the geometric distances implied by the GNP hosts coordinates can accurately predict the Internet network distances.

Measurement based analysis, modeling, and synthesis of the internet delay space

Understanding the characteristics of the Internet delay space (i.e., the all-pairs set of static round-trip propagation delays among edge networks in the Internet) is important for the design of global-scale distributed systems. For instance, algorithms used in overlay networks are often sensitive to violations of the triangle inequality and to the growth properties within the Internet delay space. Since designers of distributed systems often rely on simulation and emulation to study design alternatives, they need a realistic model of the Internet delay space. In this paper, we analyze measured delay spaces among thousands of Internet edge networks and quantify key properties that are important for distributed system design. Our analysis shows that existing delay space models do not adequately capture these important properties of the Internet delay space. Furthermore, we derive a simple model of the Internet delay space based on our analytical findings. This model preserves the relevant metrics far better than existing models, allows for a compact representation, and can be used to synthesize delay data for simulations and emulations at a scale where direct measurement and storage are impractical. We present the design of a publicly available delay space synthesizer tool called DS 2 and demonstrate its effectiveness.

A hierarchical fair service curve algorithm for link-sharing, real-time, and priority services

We study hierarchical resource management models and algorithms that support both link-sharing and guaranteed real-time services with priority (decoupled delay and bandwidth allocation). We extend the service curve based quality of service (QoS) model, which defines both delay and bandwidth requirements of a class in a hierarchy, to include fairness, which is important for the integration of real-time and hierarchical link-sharing services. The resulting fair service curve (FSC) link-sharing model formalizes the goals of link-sharing, real-time and priority services and exposes the fundamental trade-offs between these goals. In particular, with decoupled delay and bandwidth allocation, it is impossible to simultaneously provide guaranteed real-time service and achieve perfect link-sharing. We propose a novel scheduling algorithm called hierarchical fair service curve (H-FSC) that approximates the model closely and efficiently. The algorithm always guarantees the service curves of leaf classes, thus ensures real-time and priority services, while trying to minimize the discrepancy between the actual services provided to and the services defined by the FSC link-sharing model for the interior classes. We have implemented the H-FSC scheduler in NetBSD. By performing analyzes, simulations and measurement experiments, we evaluate the link-sharing and real-time performances of H-FSC, and determine the computation overhead.

Towards network triangle inequality violation aware distributed systems

Many distributed systems rely on neighbor selection mechanisms to create overlay structures that have good network performance. These neighbor selection mechanisms often assume the triangle inequality holds for Internet delays. However, the reality is that the triangle inequality is violated by Internet delays. This phenomenon creates astrange environment that confuses neighbor selection mechanisms. This paper investigates the properties of triangle inequality violation (TIV) in Internet delays, the impacts of TIV on representative neighbor selection mechanisms, specifically Vivaldi and Meridian, and avenues to reduce these impacts. We propose a TIV alert mechanism that can inform neighbor selection mechanisms to avoid the pitfalls caused by TIVs and improve their effectiveness.

Understanding the effects and implications of compute node related failures in hadoop

Hadoop has become a critical component in todays cloud environment. Ensuring good performance for Hadoop is paramount for the wide-range of applications built on top of it. In this paper we analyze Hadoops behavior under failures involving compute nodes. We find that even a single failure can result in inflated, variable and unpredictable job running times, all undesirable properties in a distributed system. We systematically track the causes underlying this distressing behavior. First, we find that Hadoop makes unrealistic assumptions about task progress rates. These assumptions can be easily invalidated by the cloud environment and, more surprisingly, by Hadoops own design decisions. The result are significant inefficiencies in Hadoops speculative execution algorithm. Second, failures are re-discovered individually by each task at the cost of great degradation in job running time. The reason is that Hadoop focuses on extreme scalability and thus trades off possible improvements resulting from sharing failure information between tasks. Third, Hadoop does not consider the causes of connection failures between its tasks. We show that the resulting overloading of connection failure semantics unnecessarily causes an otherwise localized failure to propagate to healthy tasks. We also discuss the implications of our findings and draw attention to new ways of improving Hadoop-like frameworks.

Rethinking the physical layer of data center networks of the next decade: using optics to enable efficient *-cast connectivity

Not only do big data applications impose heavy bandwidth demands, they also have diverse communication patterns denoted as *-cast) that mix together unicast, multicast, incast, and all-to-all-cast. Effectively supporting such traffic demands remains an open problem in data center networking. We propose an unconventional approach that leverages physical layer photonic technologies to build custom communication devices for accelerating each *-cast pattern, and integrates such devices into an application-driven, dynamically configurable photonics accelerated data center network. We present preliminary results from a multicast case study to highlight the potential benefits of this approach.