Policy-based management provides a means for IT systems to operate according to business needs. Unfortunately, there is often an “impedance mismatch ” between the policies administrators want and the controls they are given. Consider the Apache web server. Administrators want to control CPU and memory utilizations, but this must be done indirectly by manipulating tuning parameters such as MaxClients and KeepAlive. There has been much interest in using feedback control to bridge the impedance mismatch. However, these efforts have focused on a single metric that is manipulated by a single control and hence have not considered interactions between controls such as those that are common in computing systems. This paper shows how multiple-input, multiple-output (MIMO) control theory can be used to enforce policies for interrelated metrics. MIMO is used both to model the target system, Apache in our case, and to design feedback controllers. The MIMO model captures the interactions between KA and MC, and can be used to identify infeasible metric policies. In addition, MIMO control techniques can provide considerable benefit in handling trade-offs between speed of metric convergence and sensitivity to random fluctuations while enforcing the desired policies.

Controlling the timing performance of a network server is a challenging problem. This paper presents a Queueing Model Based Feedback Control approach to keep the timing performance of a network server close to the service level specification. We show that in an instrumented Apache server, combining feedback control with a queueing model leads to better tracking of QoS specifications than with feedback control alone or queueing model based feed forward control alone. Network based server systems, e.g., Web servers, have now become an integral part of our information services infrastructure. Controlling the timing performance of each individual connection to a network server is a challenging

Workload generators may be classified as based on a closed system model, where new job arrivals are only triggered by job completions (followed by think time), or an open system model, where new jobs arrive independently of job completions. In general, system designers pay little attention to whether a workload generator is closed or open. Using a combination of implementation and simulation experiments, we illustrate that there is a vast difference in behavior between open and closed models in realworld settings. We synthesize these differences into eight simple guiding principles, which serve three purposes. First, the principles specify how scheduling policies are impacted by closed and open models, and explain the differences in user level performance. Second, the principles motivate the use of partly open system models, whose behavior we show to lie between that of closed and open models. Finally, the principles provide guidelines to system designers for determining which system model is most appropriate for a given workload. 1

Analysis and modeling of computer network traffic is a daunting task considering the amount of available data. This is quite obvious when considering the spatial dimension of the problem, since the number of interacting computers, gateways and switches can easily reach several thousands, even in a Local Area Network (LAN) setting. This is also true for the time dimension: W. Willinger and V. Paxson in [42] cite the figures of 439 million packets and 89 gigabytes of data for a single week record of the activity of a university gateway in 1995. The complexity of the problem further increases when considering Wide Area Network (WAN) data [28]. In light of the above, it is clear that a notion of importance for modern network engineering is that of invariants, i.e. characteristics that are observed with some reproducibility and independently of the precise settings of the network under consideration. In this tutorial paper, we focus on two such invariants related to the time d...

We present architectures and algorithms for efficiently serving dynamic data at highly accessed Web sites together with the results of an analysis motivating our design and quantifying its performance benefits. This includes algorithms for keeping cached data consistent so that dynamic pages can be cached at the Web server and dynamic content can be served at the performance level of static content. We show that our system design is able to achieve cache hit ratios close to 100% for cached data which is almost never obsolete by more than a few seconds, if at all. Our architectures and algorithms provide more than an order of magnitude improvement in performance using an order of magnitude fewer servers over that obtained under conventional methods.

We investigate the distribution of the waiting time V in a stable M/G/1 processor sharing queue with trac intensity < 1. When the distribution of a customer service request B belongs to a large class of subexponential distributions with tails heavier than e , it is shown that P[V > x] = P[B > (1 )x](1 + o(1)) as x !1: Furthermore, we demonstrate that the preceding relationship does not hold if the service distribution has a lighter tail than e .

There has been tremendous progress in understanding how bandwidth is shared by TCP-like connections. By associating each TCPlike connection with a utility function, the bandwidth sharing problem of TCP-like connections can be modelled as a distributed optimization problem for utility functions. However, little is known on how bandwidth is shared by HTTP-like connections through their utility functions at the TCP level. One of the main objectives of this paper is to provide a theory for bandwidth sharing of a large number of HTTP-like connections. Based on certain technical assumptions, we show that there is a utility function at the HTTP level for an HTTP-like connection and such a utility function can be derived from the utility function at the TCP level. The bandwidth is then shared by HTTP-like connections through utility functions at the HTTP level. Moreover, there is a probabilistic interpretation for how the utility function at the HTTP level is related to the utility function at the TCP level. This is done by relating utility functions to large deviation rate functions.

Abstract. Properly optimizing the setting of configuration parameters can greatly improve performance, especially in the presence of changing workloads. This paper explores approaches to online optimization of the Apache web server, focusing on the MaxClients parameter (which controls the maximum number of workers). Using both empirical and analytic techniques, we show that MaxClients has a concave upward effect on response time and hence hill climbing techniques can be used to find the optimal value of MaxClients. We investigate two optimizers that employ hill climbing—one based on Newton’s Method and the second based on fuzzy control. A third technique is a heuristic that exploits relationships between bottleneck utilizations and response time minimization. In all cases, online optimization reduces response times by a factor of 10 or more compared to using a static, default value. The trade-offs between the online schemes are as follows. Newton’s method is well known but does not produce consistent results for highly variable data such as response times. Fuzzy control is more robust, but converges slowly. The heuristic works well in our prototype system, but it may be difficult to generalize because it requires knowledge of bottleneck resources and an ability to measure their utilizations. 1

... workload generator used to evaluate Web server performance in the computer industry. In this paper, we examine how well SPECweb99 captures the server performance characteristics that have been identified by the research community, such as URL popularity and the distributions of request methods, response codes, and transfer sizes. We compare these characteristics generated by SPECweb99 both with those reported in the literature and against a set of sample Web server logs, and find that the workload generator has a varying record depending on the characteristic. In particular, SPECweb99 suffers from failing to capture conditional GET requests, which significantly affects both the response code statistics and the transfer size distributions. We conclude with some recommendations to improve the accuracy of the benchmark in the future.

We design and implement an efficient on-line approach, FlowMate, for clustering flows (connections) emanating from a busy server, according to shared bottlenecks. Clusters can be periodically input to load balancing, congestion coordination, aggregation, admission control, or pricing modules. FlowMate uses in-band (passive) end-to-end delay measurements to infer shared bottlenecks. Delay information is piggybacked on feedback from the receivers, or, if impossible, TCP or application round trip time estimates are used. We simulate FlowMate and examine the effects of network load, traffic burstiness, network buffer sizes, and packet drop policies on clustering correctness, evaluated via a novel accuracy metric. We find that coordinated congestion management techniques are more fair when integrated with FlowMate. We also implement FlowMate in the Linux kernel v2.4.17 and evaluate its performance on the Emulab testbed, using both synthetic and tcplib-generated traffic. Our results demonstrate that clustering of medium to long-lived flows is accurate, even with bursty background traffic. Finally, we validate our results on the Internet Planetlab testbed.