Alex Zhang

Exploiting nonstationarity for performance prediction

Real production applications ranging from enterprise applications to large e-commerce sites share a crucial but seldom-noted characteristic: The relative frequencies of transaction types in their workloads are nonstationary, i.e., the transaction mix changes over time. Accurately predicting application-level performance in business-critical production applications is an increasingly important problem. However, transaction mix nonstationarity casts doubt on the practical usefulness of prediction methods that ignore this phenomenon. This paper demonstrates that transaction mix nonstationarity enables a new approach to predicting application-level performance as a function of transaction mix. We exploit nonstationarity to circumvent the need for invasive instrumentation and controlled benchmarking during model calibration; our approach relies solely on lightweight passive measurements that are routinely collected in todays production environments. We evaluate predictive accuracy on two real business-critical production applications. The accuracy of our response time predictions ranges from 10% to 16% on these applications, and our models generalize well to workloads very different from those used for calibration. We apply our technique to the challenging problem of predicting the impact of application consolidation on transaction response times. We calibrate models of two testbed applications running on dedicated machines, then use the models to predict their performance when they run together on a shared machine and serve very different workloads. Our predictions are accurate to within 4% to 14%. Existing approaches to consolidation decision support predict post-consolidation resource utilizations. Our method allows application-level performance to guide consolidation decisions.

Self-Adaptive SLA-Driven Capacity Management for Internet Services

This work considers the problem of hosting multiple third-party Internet services in a cost-effective manner so as to maximize a providers business objective. For this purpose, we present a dynamic capacity management framework based on an optimization model, which links a cost model based on SLA contracts with an analytical queuing-based performance model, in an attempt to adapt the platform to changing capacity needs in real time. In addition, we propose a two-level SLA specification for different operation modes, namely, normal and surge, which allows for per-use service accounting with respect to requirements of throughput and tail distribution response time. The cost model proposed is based on penalties, incurred by the provider due to SLA violation, and rewards, received when the service level expectations are exceeded. Finally, we evaluate approximations for predicting the performance of the hosted services under two different scheduling disciplines, namely FCFS and processor sharing. Through simulation, we assess the effectiveness of the proposed approach as well as the level of accuracy resulting from the performance model approximations

The finite multiple lot sizing problem with interrupted geometric yield and holding costs

Abstract We consider the multiple lot sizing problem in production systems with random process yield losses governed by the interrupted geometric (IG) distribution. Our model differs from those of previous researchers which focused on the IG yield in that we consider a finite number of setups and inventory holding costs. This model particularly arises in systems with large demand sizes. The resulting dynamic programming model contains a stage variable (remaining time till due) and a state variable (remaining demand to be filled) and therefore gives considerable difficulty in the derivation of the optimal policy structure and in numerical computation to solve real application problems. We shall investigate the properties of the optimal lot sizes. In particular, we shall show that the optimal lot size is bounded. Furthermore, a dynamic upper bound on the optimal lot size is derived. An O ( nD ) algorithm for solving the proposed model is provided, where n and D are the two-state variables. Numerical results show that the optimal lot size, as a function of the demand, is not necessarily monotone.

Style goods pricing with demand learning

For many industries (e.g., apparel retailing) managing demand through price adjustments is often the only tool left to companies once the replenishment decisions are made. A significant amount of uncertainty about the magnitude and price sensitivity of demand can be resolved using the early sales information. In this study, a Bayesian model is developed to summarize sales information and pricing history in an efficient way. This model is incorporated into a periodic pricing model to optimize revenues for a given stock of items over a finite horizon. A computational study is carried out in order to find out the circumstances under which learning is most beneficial. The model is extended to allow for replenishments within the season, in order to understand global sourcing decisions made by apparel retailers. Some of the findings are empirically validated using data from U.S. apparel industry.

Operational Analysis of Parallel Servers

Multicore processors promise continued hardware performance improvements even as single-core performance flattens out. However they also enable increasingly complex application software that threatens to obfuscate application-level performance. This paper applies operational analysis to the problem of understanding and predicting application-level performance in parallel servers. We present operational laws that offer both insight and actionable information based on lightweight passive external observations of black-box applications. One law accurately infers queuing delays; others predict the performance implications of expanding or reducing capacity. The former enables improved monitoring and system management; the latter enable capacity planning and dynamic resource provisioning to incorporate application-level performance in a principled way. Our laws rest upon a handful of weak assumptions that are easy to test and widely satisfied in practice. We show that the laws are broadly applicable across many practical CPU scheduling policies. Experimental results on a multicore network server in an enterprise data center demonstrate the usefulness of our laws.

A game tree strategy for automated negotiation

Chess playing programs running on small computers, such as PocketPCs,can beat most human players. This paper reports a feasibility studyto determine if the techniques programs use to play chess can beapplied to the more economically interesting problem of negotiation.This study allowed us to identify the essential differences betweenplaying chess and negotiating and to demonstrate possible solutions tothe problems we encountered.

Web transaction analysis and optimization (TAO)

In the TAO project we develop metrics, models, and infrastructure to effectively manage the performance of Web applications. We use WebMon, a novel instrumentation tool to obtain profile data for web interactions, from end-user and system component perspectives. Our analysis techniques help determine important classes of web users and their transactions. The analysis is embedded in visualization and optimization modules, enabling efficient reporting for system and business administrators, and automated resource scheduling and planning. In this paper we present an overview of TAO, and highlight some of its novel aspects, e.g., use of pixel-bar charts, web request classification, and integrated demand and capacity planning.

Operational analysis of processor speed scaling

This brief announcement presents a pair of performance laws that bound the change in aggregate job queueing time that results when the processor speed changes in a parallel computing system. Our laws require only lightweight passive external observations of a black-box system and they apply to many commonly employed scheduling policies. By predicting the application-level performance impact of processing speed adjustments in parallel processors, including traditional SMPs and now increasingly ubiquitous multicore processors, our laws address problems ranging from capacity planning to dynamic resource allocation. Finally, our results show that operational analysis---an approach to performance analysis traditionally associated with commercial transaction processing systems---usefully complements existing parallel performance analysis techniques.

DP+IP = design of efficient backup scheduling

Many industries experience an explosion in digital content. This explosion of electronic documents, along with new regulations and document retention rules, sets new requirements for performance efficiency of traditional data protection and archival tools. During a backup session a predefined set of objects (client filesystems) should be backed up. Traditionally, no information on the expected duration and throughput requirements of different backup jobs is provided. This may lead to a suboptimal job schedule that results in the increased backup session time. In this work, we characterize each backup job via two metrics, called job duration and job throughput. These metrics are derived from collected historic information about backup jobs during previous backup sessions. Our goal is to automate the design of a backup schedule that minimizes the overall completion time for a given set of backup jobs. This problem can be formulated as a resource constrained scheduling problem where a set of n jobs should be scheduled on m machines with given capacities. We provide an integer programming (IP) formulation of this problem and use available IP-solvers for finding an optimized schedule, called binpacking schedule. Performance benefits of the new bin-packing schedule are evaluated via a broad variety of realistic experiments using backup processing data from six backup servers in HP Labs. The new bin-packing job schedule significantly optimizes the backup session time (20%–60% of backup time reduction). HP Data Protector (DP) is HPs enterprise backup offering and it can directly benefit from the designed technique. Moreover, significantly reduced backup session times guarantee an improved resource/power usage of the overall backup solution.

HP Enterprise Services Uses Optimization for Resource Planning

The main responsibility of resource and delivery managers at Hewlett-Packard HP Enterprise Services HPES involves matching resources skilled professionals with jobs that project opportunities require. The previous Solution Opportunity Approval and Review SOAR process at HPES addressed uncertainty by producing decentralized project staffing decisions. This often led to many last-minute subjective, sometimes costly, resource allocation decisions. Based on our research, we developed a decision support tool for resource planning RP to enhance the SOAR process. It optimizes matching professionals who have diverse delivery roles and skills to jobs and projects across geographical locations while explicitly accounting for both demand and supply uncertainties. It also embeds capabilities for managers to incorporate tacit human knowledge and judgment information into the decision-making process. With its 2009 deployment in Best Shore, Bangalore operations of HPES, the RP tool’s significant benefits include reduced service delivery costs, increased workforce utilization, and profitability.