In Internet mapping, IP address space is divided into a set of client aggregation units, which are the finest-grained units for global load balancing. Choosing the proper level of aggregation is a complex problem, which determines the number of aggregation units that a mapping system has to maintain and client redirection. In this paper, using Internet-wide measurements provided by a commercial global load balancing service provider, we show that even for the best existing client aggregation, almost 17% of clients have latency more than 50 ms apart from the average latency of clients in the same aggregation unit. To address this, we propose a data-driven client aggregation, AP-atoms, which can trade off scalability for accuracy and adapt for changing network conditions. Since AP-atoms are obtained from the passive measurements of existing traffic between server providers and clients, no extra measurement overheads are incurred. Our experiments show that by using the same scale of client aggregations, AP-atoms can reduce the number of widely dispersed clients by almost $2\times $ and the 98th percentile difference in clients' latencies by almost 100 ms.

Internet information services replicate their servers to improve availability, response time, and fault tolerance. Traditional replication algorithms do not address the scale and administrative decentralization of today's internetworks. We have proposed and implemented a scalable and efficient tool to replicate wide-area, autonomously managed services. We target replication degrees of thousands of weakly consistent replicas. The main goal of our replication tool is to make traditional replication services scale in today's exponentially growing, autonomously managed internetworks. Our tool, which we call flood-d, allows servers to be organized in multiple replication groups. For each replication group, flood-d builds a logical update topology that is resilient to server failure, and tries to minimize the communication cost and propagation time needed to transmit updates. Flood-d's logical topologies are computed based on communication latency and available network bandwidth. This paper describes flood-d, and presents simulation results obtained when using flood-d to extend existing replication algorithms. Our results show the gains of organizing service replicas into multiple, smaller replication groups, and using network-cognizant logical topologies to propagate updates. We argue that existing as well as emerging Internet information services can benefit from flood-d's services.

We compare the performance of four admission control algorithms-one parameter-based and three measurement-based-for controlled-load service. The parameter-based admission control ensures that the sum of reserved resources is bounded by the capacity. The three measurement-based algorithms are based on measured bandwidth, acceptance region and equivalent bandwidth. We use simulation on several network scenarios to evaluate the link utilization and adherence to service commitment achieved by these four algorithms.

As the number of Internet servers increases rapidly, it becomes difficult to determine the relevant servers when searching for information. The authors develop a new method to rank Internet servers for Boolean queries. Their method reduces time and space complexity from exponential to polynomial in the number of Boolean terms. They contrast it with other known methods and describe its implementation.

Many designs for integrated services networks offer a bounded delay packet delivery service to support real-time applications. To provide a bounded delay service, networks must use admission control to regulate their load. Previous work on admission control mainly focused on algorithms that compute the worst case theoretical queueing delay to guarantee an absolute delay bound for all packets. In this paper, we describe a measurement-based admission control algorithm (ACA) for predictive service, which allows occasional delay violations. We have tested our algorithm through simulations on a wide variety of network topologies and driven with various source models, including some that exhibit long-range dependence, both in themselves and in their aggregation. Our simulation results suggest that measurement-based approach combined with the relaxed service commitment of predictive service enables us to achieve a high level of network utilization while still reliably meeting the delay bound.

This paper reports the design, implementation, and performance of a scalable and efficient tool to replicate Internet information services. Our tool targets replication degrees of tens of thousands of weakly-consistent replicas scattered throughout the Internet's thousands of autonomously administered domains. The main goal of our replication tool is to make existing replication algorithms scale in today's exponentially-growing, autonomously-managed internetworks.

This paper describes a new technique that can speedup simulation of high-speed, wide-area packet networks by one to two orders of magnitude. Speedup is achieved by coarsening the representation of network traffic from packet-by-packet to train-by-train, where a train represents a cluster of closely spaced packets. Coarsening the timing granularity creates longer trains and makes the simulation proceed more quickly since the cost of processing trains is independent of train size. Coarsening the timing granularity introduces, of course, a degree of approximation. This paper presents experiments that evaluate our coarse time-grain simulation technique for first in/first out (FIFO) switched, TCP/IP, and asynchronous transfer mode (ATM) networks carrying a mix of data and streaming traffic. We show that delay, throughput, and loss rate can frequently be estimated within a few percent via coarse time-grain simulation. This paper also describes how to apply coarse time-grain simulation to other switch disciplines. Finally, this paper introduces three more simulation techniques which together can double the performance of well written packet simulators without trading with the simulation accuracy. These techniques reduce the number of outstanding simulation events and reduce the cost of manipulating the event list.

This paper explores the claims that TCP Vegas [2] both uses network bandwidth more efficiently and achieves higher network throughput than TCP Reno [6]. It explores how link bandwidth, network buffer capacity, TCP receiver acknowledgment algorithm, and degree of network congestion affect the relative performance of Vegas and Reno.

Many designs for integrated service networks offer a bounded delay packet delivery service to support real-time applications. To provide bounded delay service, networks must use admission control to regulate their load. Previous work on admission control mainly focused on algorithms that compute the worst case theoretical queueing delay to guarantee an absolute delay bound for all packets. In this paper we describe a measurement-based admission control algorithm for predictive service, which allows occasional delay violations. We have tested our algorithm through simulations on a wide variety of network topologies and driven with various source models, including some that exhibit long-range dependence, both in themselves and in their aggregation. Our simulation results suggest that, at least for the scenarios studied here, the measurement-based approach combined with the relaxed service commitment of predictive service enables us to achieve a high level of network utilization while still reliably meeting the delay bound.

Analyses a multiround flow control algorithm that attempts to minimize the time required to multicast a message to a group of recipients and receive responses directly from each group member. Such a flow control algorithm may be necessary because the hurry of responses to the multicast can overflow the buffer space of the process that issued the multicast. The condition that each recipient directly respond to the multicast prevents the use of reliable multicast protocols based on software combining trees or negative-acknowledgments. The flow control analysed algorithm directs the responding processes to hold their responses for some period of time, called the backoff time, before sending them to the originator. The backoff time depends on the number of recipients that respond, the originator's available buffer space and buffer service time distribution, and the number of times that the originator is willing to retransmit its message. This paper develops an approximate analysis of the service time distribution of the limited-buffer preemptive queuing process that occurs within the protocol processing layers of a multiprogrammed operating system. It then uses this model to calculate multicast backoff times. The paper reports experimental verification of the accuracy of this service time model and discusses its application to the multicast flow control problem.>

This paper presents evidence that several, judiciously placed file caches could reduce the volume of FTP traffic by 42%, and hence the volume of all NSFNET backbone traffic by 21%. In addition, if FTP client and server software automatically compressed data, this savings could increase to 27%. We believe that a hierarchical architecture of whole file caches, modeled after the existing name server's caching architecture, could become a valuable part of any internet.We derived these conclusions by performing trace driven simulations of various file caching architectures, cache sizes, and replacement policies. We collected the traces of file transfer traffic employed in our simulations on a network that connects the NSFNET backbone to a large, regional network. This particular regional network is responsible for about 5 to 7% of NSFNET traffic.While this paper's analysis and discussion focus on caching for FTP file transfer, the proposed caching architecture applies to caching objects from other internetwork services.

Researchers evaluate and contrast new network routing, admission control, congestion control and flow control algorithms through simulation. Analytically de-rived arguments justifiably lack credibility because, in the attempt to model the underlying physical system, the analyst is forced to make compromising approximations. However, unlike analytical techniques like Jackson Queueing Networks, simulations require significant computation and a simulation's state can consume a great deal of memory.This paper describes a technique that we developed to reduce the memory consumption of communication network simulators. Reduced memory makes simulations of larger and higher bandwidth-delay networks possible, but introduces an adjustable degree of approximation in the simulation. The higher the memory savings, the less accurate the computed measures. We call our technique Flowsim. The paper motivates the need to simulate computer networks rather than model them analytically, motivates why a simulator's state can grow quickly, and explains why analytical techniques have failed to model modern communication networks.

Over a million computers implement the Internet's Domain Name System or DNS, making it the world's most distributed database and the Internet's most significant source of wide-area RPC-like traffic. Last year, over eight percent of the packets and four percent of the bytes that traversed the NSFnet were due to DNS. We estimate that a third of this wide-area DNS traffic was destined to seven root name servers. This paper explores the performance of DNS based on two 24-hour traces of traffic destined to one of these root name servers. It considers the effectiveness of name caching and retransmission timeout calculation, shows how algorithms to increase DNS's resiliency lead to disastrous behavior when servers fail or when certain implementation faults are triggered, explains the paradoxically high fraction of wide-area DNS packets, and evaluates the impact of flaws in various implementations of DNS. It shows that negative caching would improve DNS performance only marginally in an intern...

In this paper, we characterize wide-area network applications that use the TCP transport protocol.We also describe a new way to model the wide-area traffic generated by a stub network.We believe the fmffic model presented here will be useful in studying congestion control, routing algorithms, and other resource management schemes for existing and future networks.Our model is based on trace analysis of TCP/IP widearea intemetwork traffic.We collected the TCP/IP packet headers of USC, UCB, and Bellcore networks at the point they connect with their respective regional access networks.We then wrote a handful of programs to analyze the traces.Our model characterizes individual TCP conversations by the distributions ofi number of bytes transfemed, duration, number of packets transferred, packet size, and packet interarrival time.Our trace analysis shows that both interactive and bulk transfer traffic from all sites reflect a large number of short conversations.Similarly, it shows that a very large percentage of traffic is bidirectional, even for bulk transfer.We observed that interactive applications send significantly different amounts of data in each direction of a conversation, and that interarrival times for interactive applications closely follow a constant plus exponential model.Half of the conversations are directed to a handful of networks, but the other half are directed to hundreds of networks.Many of these observations contradict commonly held beliefs regarding wide-area traffic.

We model the limited buffer queueing process that occurs within the UNIX operating system's protocol processing layers. Our model accounts for the effects of user process multiprogramming and preemptive, priority scheduling of interrupt, operating system, and user tasks. After developing the model, we use it to predict message loss that occurs during local area network (LAN) multicast. Our service time model can be applied to window-and rate-based stream flow control.

When many or all of the recipients of a multicast message respond to the multicast's sender, their responses may overflow the sender's available buffer space. Buffer overflow is a serious, known problem of broadcast-based protocols, and can be troublesome when as few as three or four recipients respond. We develop analytical models that calculate the expected number of buffer overflows that can be used to estimate the number of buffers necessary for an application. The common cure for buffer overflow requires that recipients delay their responses by some random amount of time in order to increase the minimum spacing between response messages, eliminate collisions on the network, and decrease the peak processing demand at the sender. In our table driven algorithm, the sender tries to minimize the multicast's latency, the elapsed time between its initial transmission of the multicast and its reception of the final response, given the number of times (rounds) it is willing to retransmit the multicast. It includes in the multicast the time interval over which it anticipates receiving the response, the round timeout. We demonstrate that the latency of single round multicasts exceeds the latency of multiple round multicasts. We show how recipients minimize the sender's buffer overflows by independently choosing their response times as a function of the round's timeout, sender's buffer size, and the number of other recipients.

This paper compares the performance of executing compiled Prolog code on two different architectures under development at U. C. Berkeley. The first is the PLM, a special-purpose CISC architecture intended as a coprocessor for a host machine. The second is SPUR, a general-purpose RISC architecture that supports tagged data. Fourteen standard benchmark programs were run on both the PLM and SPUR simulators. The compiled code for SPUR was obtained by simple macro-expansion of PLM code generated by the PLM Prolog compiler. The two implementations are compared with regard to static and dynamic program size, execution speed, and memory system performance. On average, the macrocoded SPUR implementation has a static code size 14 times larger than the PLM, executes 16 times more instructions, yet requires only 2.3 times the number of machine cycles (or has the performance of 0.43 PLMs). When memory system performance is taken into account, SPUR is equivalent to 0.29 PLMs. Optimizations of the macro-expanded code and minor architectural changes to SPUR would increase this ratio to 0.53, or 0.60 for the largest benchmarks. Thus a tagged RISC architecture can execute Prolog at least half as fast as a special-purpose CISC architecture for Prolog.