. We consider a distributed server system model and ask which policy should be used for assigning tasks to hosts. In our model each host processes tasks in First-Come-First-Serve order and the task's service demand is known in advance. We consider four task assignment policies commonly proposed for such distributed server systems: RoundRobin, Random, Size-Based, in which all tasks within a give size range are assigned to a particular host, and Dynamic-Least-Work-Remaining, in which a task is assigned to the host with the least outstanding work. Our goal is to understand the influence of task size variability on the decision of which task assignment policy is best. We find that no one of the above task assignment policies is best and that the answer depends critically on the variability in the task size distribution. In particular we find that when the task sizes are not highly variable, the Dynamic policy is preferable. However when task sizes show the degree of variability more charac...

Recent measurements of network traffic have shown that self-similarity is an ubiquitous phenomenon present in both local area and wide area traffic traces. In previous work, we have shown a simple, robust application layer causal mechanism of traffic self-similarity, namely, the transfer of files in a network system where the file size distributions are heavy-tailed. In this paper, we study the effect of scale-invariant burstiness on network performance when the functionality of the transport layer and the nteraction of traffic sources sharing bounded network resources is incorporated. First, we show that transport layer mechanisms are important factors in translating the application layer causality into link traffic self-similarity. Network performance as captured by throughput, packet loss rate, and packet retransmission rate degrades gradually with increased heavy-tailedness while queueing delay, response time, and fairness deteriorate more drastically. The degree to which heavy-tailedness affects self-similarity is determined by how well congestion control is able to shape a source traffic into an on-average constant output stream while conserving information. Second, we show that increasing network resources such as link bandwidth and buffer capacity results in a superlinear improvement in performance. When large file transfers occur with nonnegligible probability, the incremental

workload characterization, performance, proxy caching This paper presents a detailed workload characterization study of a World-Wide Web proxy. Measurements from a proxy within an Internet Service Provider (ISP) environment were collected. This ISP allows clients to access the Web using highspeed cable modems rather than traditional dial-up modems. By examining this site we are able to evaluate the effects that cable modems have on proxy workloads. This paper focuses on workload characteristics such as file type distribution, file size distribution, file referencing behaviour and turnover in the active set of files. We find that when presented with faster access speeds users are willing to download extremely large files. A widespread increase in the transfer of these large files would have a significant impact on the Web. This behaviour increases the importance of caching for ensuring the scalability of the Web.

Heavy-tailed distributions, a.k.a. power-law distributions, have been observed in many natural phenomena ranging from physical phenomena to sociological phenomena. Recently heavy-tailed distributions have been discovered in computer systems. In particular the sizes (service demands) of computing jobs have been found to exhibit a heavy-tailed (power-law) distribution. Most previous analytic work in the area of computer system design has assumed that job sizes (service demands) are exponentially distributed. Many of the policies, algorithms, and general rules-of-thumb which are currently used in computer systems originated from analyses which assumed an exponentiallydistributed workload. In this paper we argue that we need to reevaluate existing computer system algorithms in light of the discovery of heavy-tailed workloads. We argue that an algorithm which is optimal under an exponentially distributed workload may be very far from optimal when the workload is heavy-tailed. We ...

The arrival process of jobs submitted to a parallel system is bursty, leading to fluctuations in the load at many time scales. In particular, rare events of extreme load may occur. Such events lead to an increase in the standard deviation of performance metrics, and thus delay the convergence of simulations used to evaluate the scheduling. Different performance metrics have been proposed in an effort to reduce this variability, and indeed display different rates of convergence. However, there is no single metric that outperforms the others under all conditions. Rather, the convergence of different metrics depends on the system being studied. 1

Thispaperpresentsadiscussionofsimulation-related issuesarisinginthestudyofself-similarnetwork tra#cwithrespecttoitse#ectandcontrol.Selfsimilartra #chasbeenshowntobeanubiquitous phenomenonarisingindiversenetworkingcontexts withpotentiallyadversee#ectsonnetworkperformance. Inmanyinstances,anexperimentalorempirical approachneedstobetakentoe#ectivelyevaluate theperformanceimpactofsophisticatedcontrolalgorithmsactingatvariouslayersintheprotocolstack underself-similartra#cconditions.Simulatingorexperimentallyimplementingsuchenvironmentsisnon - trivialduetothefactthat,ingeneral,thecharacteristicsoftheobservedtra #cisitselfinfluencedbythe actionsofthecontrolalgorithmsunderstudy.To whatdegreeself-similaritymanifestsitselfinnetwork tra#cmaydependonthepropertiesoftheprotocols employed,andtrace-basedsimulationsthatrelyon tra#cmeasurementstodrivesimulationsfailtocapturethisdynamicaspect. Wediscussanapproachtoevaluatingcontrolprotocolsunderself -similartra#cconditionsbasedon asimple,robustapplication-levelcausalmechanism oftra#cself-similaritywhichisgroundedinboth empiricalUNIXfilesystemresearchandanalytic tra#cmodelsinvolvingcertainrenewalprocesses. Wepresentahigh-leveldiscussionconcentratingon simulation-relatedissues,withspecificresearchresultssummarizedorpointedtointhereferences. 1

This paper considers the basic problem of "how accurate can we make Internet performance measurements". The answer is somewhat counter-intuitive in that there are bounds on the accuracy of such measurements, no matter how many probes we can use in a given time interval, and thus arises a type of Heisenberg inequality describing the bounds in our knowledge of the performance of a network. The results stem from the fact that we cannot make independent measurements of a system's performance: all such measures are correlated, and these correlations reduce the efficacy of measurements. The degree of correlation is also strongly dependent on system load. The result has important practical implications that reach beyond the design of Internet measurement experiments, into the design of network protocols.

The Least Attained Service (LAS) scheduling policy, when used for scheduling packets over the bottleneck link of an Internet path, can greatly reduce the average flow time for short flows while not significantly increasing the average flow time for the long flows that share the same bottleneck. No modification of the packet headers is required to implement the simple LAS policy. However, previous work has also shown that a drawback of the LAS scheduler is that, when link utilization is greater than 70%, long flows experience large jitter in their packet transfer times as compared to the conventional First-Come-First-Serve (FCFS) link scheduling. This paper proposes and evaluates new differentiated LAS scheduling policies that reduce the jitter for long flows that are identified as "priority" flows.

Measurements of network tra c have shown that self-similarity is a ubiquitous phenomenon spanning across diverse network environments. In previous work, we have explored the feasibility of exploiting long-range correlation structure in self-similar tra c for congestion control. We have advanced the framework of multiple time scale congestion control and shown its e ectiveness at enhancing performance for rate-based feedback control. In this paper, we extend the multiple time scale control framework to window-based congestion control, in particular, TCP. This is performed by interfacing TCP with a large time scale control module which adjusts the aggressiveness of bandwidth consumption behavior exhibited by TCP as a function of \large time scale " network state, i.e., information that exceeds the horizon of the feedback loop as determined by RTT. How to e ectively utilize such information|due to its probabilistic nature, dispersion over multiple time scales, and a ection on top of existing window-based congestion controls|is a nontrivial problem. Our contribution is threefold. First, we de ne a modular extension of TCP|a function call with a simple interface|that applies to various avours of TCP|e.g., Tahoe, Reno, Vegas|

It has now been demonstrated in many studies that network traffic exhibits properties consistent with Long Range Dependence (LRD) and self-similarity. While theoretical frameworks are currently being developed to estimate the performance of such systems, simulation will remain a valuable tool for validating these theoretical models, and providing insight into systems which are too complicated to effectively model. Furthermore, when testing real systems, it is desirable to have traffic sources which are realistic, and hence display self-similarity. The Fractal Renewal Process (FRP) and its variants (including On/Off processes and superpositions thereof) have been proposed as models for LRD processes, in particular for network traffic. The FRP is a simple renewal point process with heavy-tailed inter-renewal times. The long-range correlations in the process are directly introduced by the heavy tail of the renewal times. The FRP has the great advantage that the number of computations required to generate a time series is linear and the time series can be generated on-line, facilitating generation of real traffic. However, there are some problems which arise when using such processes to generate LRD traffic. Most notably undersampling of the heavy-tailed random variables used to generate FRPs can lead to a truncation of the sampled autocorrelation that is not consistent with LRD. This problem becomes clear when the processes are investigated using the wavelet based methods of Abry and Veitch which segregate behaviour at different scales. This paper will describe the problem of undersampling, and its effects, and methods for avoiding the problem.