Transient overload is a short term event characterized by a sudden and unexpected rise of the arrival rate in a Web server. Transient overload events in Web servers have been observed more frequently as the number of Internet users grows and the Web services become more popular. This paper deals with the non-trivial problem of generating overload for Web server testing. In this paper we present a new Web server workload generator called TORO, which is able to overload a Web server based on the maximum effective capacity of the server. The load generator TORO allows the usage of both real and synthetic loads. The results show that the load generator is able to deliver thousands of requests per second in a variety of workload patterns. Additionally, this paper presents an evaluation of the Apache HTTP Server under transient overload events. The results can be applied to the design of new Web servers.

As Internet computing gains speed, complexity, and becomes ubiquitous, the need for precise and accurate time synchronization increases. In this paper, we present a characterization of a clock synchronization network managed by Network Time Protocol (NTP), composed by thousands of nodes, including hundreds of Stratum 1 servers, based on data collected recently by a robot. NTP is the most common protocol for time synchronization in the Internet. Many aspects that define the quality of timekeeping are analyzed, as well as topological characteristics of the network. The results are compared to previous characterizations of the NTP network, showing the evolution of clock synchronization in the last fifteen years.

Networked systems are susceptible to transient overload, which can lead to system crash due to resource exhaustion. In the absence of appropriate mechanisms, a system can enter and persist in a state of livelock during the overload period. This paper presents an evaluation of two mechanisms for livelock control experienced in computers of different capacities in a gigabit network. The first mechanism is called interrupt moderation and is implemented in most of today's high-performance network interfaces. The second mechanism, called NAPI, is a new API for Linux network drivers, which implements a set of strategies to handle packet overload and avoid its consequences. We show that NAPI is more efficient than interrupt moderation in many aspects, including system stability in the case of huge traffic and high CPU usage.

The service time distribution is a fundamental parameter in analysis and simulation models. However, the distribution of the service times of Web servers and caches is not well understood. As a consequence, in many studies on Web server scheduling and task assignment, the service times have been assumed as directly proportional to the size of the file transferred. But Web transfers of the same size may experience different service times due to dynamic characteristics such as server load, network traffic and client connectivity. In this paper we present models for both the service time and the file size of Web cache transfers. Our goal is to compare both empirical distributions, exploring the correlation between them. Our results are based on recent logs from NLANR. We show that both variables may be modeled by the log-normal distribution but with different parameters.

Bandwidth estimation is an important task because the knowledge of the bandwidth of a path is useful in a wide variety of contexts. Clients, applications and servers benefit greatly from knowing the bandwidth of a route. In this paper we describe and discuss experiments carried out to provide bandwidth estimates using five tools: bprobe, clink, nettimer pathrate, and pchar The tools are evaluated and compared according to the criteria accuracy, statistical robustness and time to estimation. We show several results for short and long network paths. The topologies tested include links which capacities range from 2 Mb/s to 1 Gb/s.

We consider the problem of task assignment in a distributed system (such as a distributed Web server) in which task sizes are drawn from a heavy-tailed distribution. Many task assignment algorithms are based on the heuristic that balancing the load at the server hosts will result in optimal performance. We show this conventional wisdom is less true when the task size distribution is heavy-tailed (as is the case for Web file sizes). We introduce a new task assignment policy, called Size Interval Task Assignment with Variable Load (SITA-V). SITA-V purposely operates the server hosts at different loads, and directs smaller tasks to the lighter-loaded hosts. The result is that SITA-V provably decreases the mean task slowdown by significant factors (up to 1000 or more) where the more heavy-tailed the workload, the greater the improvement factor. We evaluate the tradeoff between improvement in slowdown and increase in waiting time in a system using SITA-V, and show conditions under which SITA-V represents a particularly appealing policy.