Rodrigo Fonseca

A methodology for workload characterization of E-commerce sites

Performance analysis and capacity planning for e-commerce sites poses an interesting problem: how to best characterize the workload of these sites. Tradition al workload characterization methods, based on hits/set, page views/set, or visits/set, are not appropriate for e-commerce sites. In these environments, customers interact with the site through a series of consecutive and related requests, called sessions. Different navigational patterns can be observed for different groups of customers. In this paper, we propose a methodology for characterizing and generating e-commerce workload models. First, we introduce a state transition graph called Customer Behavior Model Graph (CBMG), that is used to describe the behavior of groups of customers who exhibit similar navigational patterns. A set of useful metrics, analytically derived from the analysis of the CBMG, is presented. Next, we define a workload model and show the steps required to obtain its parameters. We then propose a clustering algorithm to characterize workloads of e-commerce sites in terms of CBMGs. Finally, we present and discuss experimental results of the use of proposed methodology.

Flush: a reliable bulk transport protocol for multihop wireless networks

We present Flush, a reliable, high goodput bulk data transport protocol for wireless sensor networks. Flush provides end-to-end reliability, reduces transfer time, and adapts to time-varying network conditions. It achieves these properties using end-to-end acknowledgments, implicit snooping of control information, and a rate-control algorithm that operates at each hop along a flow. Using several real network topologies, we show that Flush closely tracks or exceeds the maximum goodput achievable by a hand-tuned but fixed rate for each hop over a wide range of path lengths and varying network conditions. Flush is scalable; its effective bandwidth over a 48-hop wireless network is approximately one-third of the rate achievable over one hop. The design of Flush is simplified by assuming that different flows do not interfere with each other, a reasonable restriction for many sensornet applications that collect bulk data in a coordinated fashion, like structural health monitoring, volcanic activity monitoring, or protocol evaluation. We collected all of the performance data presented in this paper using Flush itself.

Reliable transfer on wireless sensor networks

Many applications in wireless sensor networks, including structure monitoring, require collecting all data without loss from the nodes. End-to-end retransmission, which is used in the Internet for reliable transport, becomes very inefficient in wireless sensor networks, since wireless communication, and constrained resources pose new challenges. We look at factors affecting reliability, and search for efficient combinations of the possible options. Information redundancy like retransmission, and erasure codes, can be used. Route fix, which tries alternative next hop after some failures, also reduces packet loss. We implemented and evaluated these options on a real test bed of Berkeley Mica2Dot motes. Our experimental results show that each option overcomes different kinds of failures. Link-level retransmission is efficient but limited in achieving reliability. Erasure code enables very high reliability by tolerating packet losses. Route fix responds to link failures quickly. Previous work had found it difficult to increase reliability past a certain threshold. We show that the right combination of primitives can yield more than 99% reliability with low overhead, providing a viable alternative to end-to-end retransmission over multiple hops.

Participatory networking: an API for application control of SDNs

We present the design, implementation, and evaluation of an API for applications to control a software-defined network (SDN). Our API is implemented by an OpenFlow controller that delegates read and write authority from the networks administrators to end users, or applications and devices acting on their behalf. Users can then work with the network, rather than around it, to achieve better performance, security, or predictable behavior. Our API serves well as the next layer atop current SDN stacks. Our design addresses the two key challenges: how to safely decompose control and visibility of the network, and how to resolve conflicts between untrusted users and across requests, while maintaining baseline levels of fairness and security. Using a real OpenFlow testbed, we demonstrate our APIs feasibility through microbenchmarks, and its usefulness by experiments with four real applications modified to take advantage of it.

Jockey: guaranteed job latency in data parallel clusters

Data processing frameworks such as MapReduce [8] and Dryad [11] are used today in business environments where customers expect guaranteed performance. To date, however, these systems are not capable of providing guarantees on job latency because scheduling policies are based on fair-sharing, and operators seek high cluster use through statistical multiplexing and over-subscription. With Jockey, we provide latency SLOs for data parallel jobs written in SCOPE. Jockey precomputes statistics using a simulator that captures the jobs complex internal dependencies, accurately and efficiently predicting the remaining run time at different resource allocations and in different stages of the job. Our control policy monitors a jobs performance, and dynamically adjusts resource allocation in the shared cluster in order to maximize the jobs economic utility while minimizing its impact on the rest of the cluster. In our experiments in Microsofts production Cosmos clusters, Jockey meets the specified job latency SLOs and responds to changes in cluster conditions.

Business-oriented resource management policies for e-commerce servers

Abstract Quality of service of e-commerce sites has been usually managed by the allocation of resources such as processors, disks, and network bandwidth, and by tracking conventional performance metrics such as response time, throughput, and availability. However, the metrics that are of utmost importance to the management and shareholders of a Web store are revenue and profit. Thus, resource management schemes for e-commerce servers should be geared towards optimizing business metrics as opposed to conventional performance metrics. This paper uses a state transition graph called customer behavior model graph (CBMG) to describe a customer session. It then presents a family of priority based resource management policies for e-commerce servers. Priorities change dynamically as a function of the state a customer is in and as a function of the amount of money the customer has accumulated in his/her shopping cart. A detailed simulation model was developed to assess the gain of these dynamic policies with respect to policies that are oblivious to economic considerations. Simulation results show that the multilevel dynamic priority scheme suggested here can significantly improve the values of business-oriented metrics, such as revenue per second, during peak periods. E-commerce sites that use this approach will be able to improve revenue at peak times with the same server capacity.

In search of invariants for e-business workloads

ABSTRACT Understanding the nature and hara teristi s of e-business workloads is a ru ial step to improve the quality of servi e o ered to ustomers in ele troni business environments. However, the variety and omplexity of the intera tions between ustomers and sites make the hara terization of ebusiness workloads a hallenging problem. Using a multilayer hierar hi al model, this paper presents a detailed hara terization of the workload of two a tual e-business sites: an online bookstore and an ele troni au tion site. Through the hara terization pro ess, we found the presen e of autonomous agents, or robots, in the workload and used the hierar hi al stru ture to determine their hara teristi s. We also found that sear h terms follow a Zipf distribution.

Rank-preserving two-level caching for scalable search engines

We present an e ective caching scheme that reduces the computing and I/O requirements of a Web search engine without altering its ranking characteristics. The novelty is a two-level caching scheme that simultaneously combines cached query results and cached inverted lists on a real case search engine. A set of log queries are used to measure and compare the performance and the scalability of the search engine with no cache, with the cache for query results, with the cache for inverted lists, and with the two-level cache. Experimental results show that the two-level cache is superior, and that it allows increasing the maximum number of queries processed per second by a factor of three, while preserving the response time. These results are new, have not been reported before, and demonstrate the importance of advanced caching schemes for real case search engines.

Characterizing selfishly constructed overlay routing networks

We analyze the characteristics of overlay routing networks generated by selfish nodes playing competitive network construction games. We explore several networking scenarios - some simplistic, others more realistic - and analyze the resulting Nash equilibrium graphs with respect to topology, performance, and resilience. We find a fundamental tradeoff between performance and resilience, and show that limiting the degree of nodes is of great importance in controlling this balance. Further, by varying the cost function, the game produces widely different topologies; one parameter in particular - the relative cost between maintaining an overlay link and increasing the path length to other nodes - can generate topologies with node-degree distributions whose tails vary from exponential to power-law. We conclude that competitive games can create overlay routing networks satisfying very diverse goals.

CTP: An efficient, robust, and reliable collection tree protocol for wireless sensor networks

We describe CTP, a collection routing protocol for wireless sensor networks. CTP uses three techniques to provide efficient, robust, and reliable routing in highly dynamic network conditions. CTPs link estimator accurately estimates link qualities by using feedback from both the data and control planes, using information from multiple layers through narrow, platform-independent interfaces. Second, CTP uses the Trickle algorithm to time the control traffic, sending few beacons in stable topologies yet quickly adapting to changes. Finally, CTP actively probes the topology with data traffic, quickly discovering and fixing routing failures. Through experiments on 13 different testbeds, encompassing seven platforms, six link layers, and multiple densities and frequencies, and detailed observations of a long-running sensor network application that uses CTP, we study how these three techniques contribute to CTPs overall performance.