Recent empirical studies [6] have shown that Internet topologies exhibit power laws of the form � for the following relationships: (P1) outdegree of node (domain or router) versus rank; (P2) number of nodes versus outdegree; (P3) number of node pairs within a neighborhood versus neighborhood size (in hops); and (P4) eigenvalues of the adjacency matrix versus rank. However, causes for the appearance of such power laws have not been convincingly given. In this paper, we examine four factors in the formation of Internet topologies. These factors are (F1) preferential connectivity of a new node to existing nodes; (F2) incremental growth of the network; (F3) distribution of nodes in space; and (F4) locality of edge connections. In synthetically generated network topologies, we study the relevance of each factor in causing the aforementioned power laws as well as other properties, namely

Recent measurement based studies reveal that most of the Internet connections are short in terms of the amount of traffic they carry (mice), while a small fraction of the connections are carrying a large portion of the traffic (elephants). A careful study of the TCP protocol shows that without help from an Active Queue Management (AQM) policy, short connections tend to lose to long connections in their competition for bandwidth. This is because short connections do not gain detailed knowledge of the network state, and therefore they are doomed to be less competitive due to the conservative nature of the TCP congestion control algorithm.

As satellite, wireless and Cable TV-based networks spread their reach, there is an increased infrastructure of high-bandwidth links into the home and on the road. Much of this enhanced infrastructure inherently relies on broadcast technology to deliver data to large user populations. This increase in broadcast capacity has been complemented by the growth of large-scale information-centric applications. Many of these applications such as wireless internets and traffic information systems are pull-based, that is, they respond to on-demand user requests. In this paper, we study the scheduling problems arising in such on-demand broadcast environments for applications with data requests of varying sizes, and the novel issues that arise therein. We study the problem in its generality while much of the previous work has focused on one special case or the other, such as, assuming identical-sized data requests, or static client access profiles known by the server a priori, etc. Traditionally,...

Popular Web sites can neither rely on a single powerful server nor on independent mirroredservers to support the ever increasing request load. Scalability and availability can be provided by distributed Web-server architectures that schedule client requests among the multiple server nodes in a user-transparent way. In this paper we will review the state of the art in load balancing techniques on distributed Web-server systems. We will analyze the efficiency and limitations of the various approaches and their tradeoff.

It is becoming increasingly common to construct network services using redundant resources geographically distributed across the Internet. Content Distribution Networks are a prime example. Such systems distribute client requests to an appropriate server based on a variety of factors---e.g., server load, network proximity, cache locality---in an effort to reduce response time and increase the system capacity under load. This paper explores the design space of strategies employed to redirect requests, and defines a class of new algorithms that carefully balance load, locality, and proximity. We use large-scale detailed simulations to evaluate the various strategies. These simulations clearly demonstrate the effectiveness of our new algorithms, which yield a 60-91% improvement in system capacity when compared with the best published CDN technology, yet user-perceived response latency remains low and the system scales well with the number of servers.

MapReduce is emerging as an important programming model for large-scale data-parallel applications such as web indexing, data mining, and scientific simulation. Hadoop is an open-source implementation of MapReduce enjoying wide adoption and is often used for short jobs where low response time is critical. Hadoop’s performance is closely tied to its task scheduler, which implicitly assumes that cluster nodes are homogeneous and tasks make progress linearly, and uses these assumptions to decide when to speculatively re-execute tasks that appear to be stragglers. In practice, the homogeneity assumptions do not always hold. An especially compelling setting where this occurs is a virtualized data center, such as Amazon’s Elastic Compute Cloud (EC2). We show that Hadoop’s scheduler can cause severe performance degradation in heterogeneous environments. We design a new scheduling algorithm, Longest Approximate Time to End (LATE), that is highly robust to heterogeneity. LATE can improve Hadoop response times by a factor of 2 in clusters of 200 virtual machines on EC2. 1

We consider a distributed server system and ask which policy should be used for assigning jobs (tasks) to hosts. In our server, jobs are ot preemptible. Also, the job's service demand is ot known a priori. We are particularly concerned with the case where the workload is heavy-tailed, as is characteristic of many empirically measured computer workloads. We analyze several natural task assignment policies and propose a new one T/IGS (Task Assignment based on Guessing Size). The T/IGS algorithm is counterintuitive in many respects, including load mbalancing, omwork-conserving, and faithless. We find that under heavy-tailed workloads, T/IGS can outperform all task assignment policies known to us by several orders of magnitude with respect to both mean response time and mean slowdown, provided the system load is not too high.

In this paper we examine the problem of balancing load in a large-scale distributed system when information about server loads may be stale. It is well known that sending each request to the machine with the apparent lowest load can behave badly in such systems, yet this technique is common in practice. Other systems use round-robin or random selection algorithms that entirely ignore load information or that only use a small subset of the load information. Rather than risk extremely bad performance on one hand or ignore the chance to use load information to improve performance on the other, we develop strategies that interpret load information based on its age. Through simulation, we examine several simple algorithms that use such load interpretation strategies under a range of workloads. Our experiments suggest that by properly interpreting load information, systems can (1) match the performance of the most aggressive algorithms when load information is fresh relative to the...

Over the last decade an important new...

This paper considers policies for distributing requests in clustered Web servers, wherein multiple server machines are configured to function as a single high(er) performance Web server. We evaluate various load distribution policies with respect to both their ability to achieve good load balance (the primary goal) and also to their impact on the effectiveness of per-machine caching. Trace-driven simulation is employed, with workload traces from two heavily-loaded (3-8 million requests per day) commercial Web servers. Our results show that use of current state information is necessary in achieving good load balance only when the achievable per-request bandwidth is not strongly network or client limited. Use of current state information is not found to be necessary with respect to achieving good cache behaviour. Load distribution based on a static hashed assignment of the URL space is found to yield very similar cache performance to load distribution based on current cache contents. We also find that it is possible to achieve both good cache behaviour and good load balance, but it requires use of policies that take both objectives into consideration and that make use of information concerning current server loads.