Jerome Rolia

The Method of Layers

Distributed applications are being developed that contain one or more layers of software servers. Software processes within such systems suffer contention delays both for shared hardware and at the software servers. The responsiveness of these systems is affected by the software design, the threading level and number of instances of software processes, and the allocation of processes to processors. The Method of Layers (MOL) is proposed to provide performance estimates for such systems. The MOL uses the mean value analysis (MVA) linearizer algorithm as a subprogram to assist in predicting model performance measures. >

SkewTune: mitigating skew in mapreduce applications

We present an automatic skew mitigation approach for user-defined MapReduce programs and present SkewTune, a system that implements this approach as a drop-in replacement for an existing MapReduce implementation. There are three key challenges: (a) require no extra input from the user yet work for all MapReduce applications, (b) be completely transparent, and (c) impose minimal overhead if there is no skew. The SkewTune approach addresses these challenges and works as follows: When a node in the cluster becomes idle, SkewTune identifies the task with the greatest expected remaining processing time. The unprocessed input data of this straggling task is then proactively repartitioned in a way that fully utilizes the nodes in the cluster and preserves the ordering of the input data so that the original output can be reconstructed by concatenation. We implement SkewTune as an extension to Hadoop and evaluate its effectiveness using several real applications. The results show that SkewTune can significantly reduce job runtime in the presence of skew and adds little to no overhead in the absence of skew.

A toolset for performance engineering and software design of client-server systems

Abstract TimeBench/SRVN is a prototype toolset for computer-aided design and performance analysis of software, with an emphasis on distributed client-server systems. The performance behaviour of such systems may defy intuition because it involves factors in the software design (such as the partitioning of the functionality and the frequency with which requests will be made to each server) and in the configuration of the distributed system (including replication of services, the distribution of data, and the speed of network access). The novelty of the tool consists in providing support both for developing design specifications and also for performance analysis. The integrated approach avoids the semantic gap between a designers domain and the performance modeling domain, and assists the designer to explore factors that impact the performance of a design. The performance models are based on the Stochastic Rendezvous Network (SRVN) formalism for client-server systems with synchronous service requests. The distinctive features of SRVNs are nested services (since servers can also act as clients to other servers) and the existence of two or more phases of service (the first executed while the client is blocked, and the others executed in parallel with the client). TimeBench/SRVN is intended as a demonstration of the concept of an integrated designer/performance interface, and as a research environment for fast analytic solvers for the models. Besides a simulation solver, it offers three approximate analytic solvers based on recent research, a Markovian solver, a technique for finding bounds on the throughput without too many assumptions, and a tool for rapidly exploring the space of possible parameter values.

Skew-resistant parallel processing of feature-extracting scientific user-defined functions

Scientists today have the ability to generate data at an unprecedented scale and rate and, as a result, they must increasingly turn to parallel data processing engines to perform their analyses. However, the simple execution model of these engines can make it difficult to implement efficient algorithms for scientific analytics. In particular, many scientific analytics require the extraction of features from data represented as either a multidimensional array or points in a multidimensional space. These applications exhibit significant computational skew, where the runtime of different partitions depends on more than just input size and can therefore vary dramatically and unpredictably. In this paper, we present SkewReduce, a new system implemented on top of Hadoop that enables users to easily express feature extraction analyses and execute them efficiently. At the heart of the SkewReduce system is an optimizer, parameterized by user-defined cost functions, that determines how best to partition the input data to minimize computational skew. Experiments on real data from two different science domains demonstrate that our approach can improve execution times by a factor of up to 8 compared to a naive implementation.

A Synthetic Workload Generation Technique for Stress Testing Session-Based Systems

Enterprise applications are often business critical but lack effective synthetic workload generation techniques to evaluate performance. These workloads are characterized by sessions of interdependent requests that often cause and exploit dynamically generated responses. Interrequest dependencies must be reflected in synthetic workloads for these systems to exercise application functions correctly. This poses significant challenges for automating the construction of representative synthetic workloads and manipulating workload characteristics for sensitivity analyses. This paper presents a technique to overcome these problems. Given request logs for a system under study, the technique automatically creates a synthetic workload that has specified characteristics and maintains the correct interrequest dependencies. The technique is demonstrated through a case study involving a TPC-W e-commerce system. Results show that incorrect performance results can be obtained by neglecting interrequest dependencies, thereby highlighting the value of our technique. The study also exploits our technique to investigate the impact of several workload characteristics on system performance. Results establish that high variability in the distributions of session length, session idle times, and request service times can cause increased contention among sessions, leading to poor system responsiveness. To the best of our knowledge, these are the first results of this kind for a session-based system. We believe our technique is of value for studies where fine control over workload is essential

Web server performance measurement and modeling techniques

Abstract The popularity of the Internet, and the usage of the world wide web in particular, has grown rapidly in recent years. Thousands of companies have deployed Web servers and their usage rates have increased dramatically. Our research has focused on measuring, analyzing and evaluating the performance of Internet and Intranet Web servers with a goal of creating capacity planning models. We have created layered queuing models (LQMs) and demonstrated their superiority to traditional queuing network models since they incorporate layered resource demands. Along the way we built a tool framework that enables us to collect and analyze the empirical data necessary to accomplish our goals. This paper describes the custom instrumentation we developed and deployed to collect workload metrics and model parameters from large-scale, commercial Internet and Intranet Web servers. We discuss the measurement issues pertaining to model parametrization and validation. We describe an object-oriented tool framework that significantly improves the productivity of analyzing the nearly 100 GBs of measurements collected during this workload study interval. Finally, we describe the LQM we developed to estimate client response time at a Web server. The model predicts the impact on server and client response times as a function of network topology and Web server pool size. We also use it to consider the consequences of server system configuration changes such as decreasing the HTTP object cache size.

Trace-based load characterization for generating performance software models

Performance models of software designs can give early warnings of problems such as resource saturation or excessive delays. However models are seldom used because of the considerable effort needed to construct them. The ANGIOTRACE/sup TM/ was developed to gather the necessary information from an executable design and develop a model in an automated fashion. It applies to distributed and concurrent software with synchronous (send-reply or RPC) communications, developing a layered queuing network model. The trace-based load characterization (TLC) technique presented here extends the ANGIOTRACE/sup TM/ to handle software with both synchronous and asynchronous interactions. TLC also detects interactions which are effectively synchronous or partly-synchronous (forwarding) but are built up from asynchronous messages. These patterns occur in telephony software and in other systems. The TLC technique can be applied throughout the software life-cycle, even after deployment.

Correlating resource demand information with ARM data for application services

Recognizing and solving a distributed application’s performance problems can require information about the application’s structure, the application-level services offered by each of its processes, and the resource demands and response times of each of these services. The Application Response Measurement package (ARM) Version 2.0 offers a de facto standard instrumentation paradigm for distributed applications that enables the monitoring of services and their relationships. The data that can be captured is adequate for deducing the structure of distributed applications and for gathering response time measures. However for most platforms it is not possible to measure a service’s resource demands directly. This last step is necessary for sizing studies and the development of predictive performance models. In this paper we present the results of experiments that illustrate the conditions under which common statistical techniques, instead of direct measurement, can be used to accurately correlate the use of application level services with resource demand information. Both simulation and the measurement of a CORBA application are used to assess the accuracy of the techniques and the relative importance of various experimental factors. The results of the experiments are used to provide a set of requirements for ARM monitoring infrastructures and strategies for ARM instrumentation plans.

Designing process replication and activation: a quantitative approach

Distributed application systems are composed of classes of objects with instances that interact to accomplish common goals. Such systems can have many classes of users with many types of requests. Furthermore, the relative load of these classes can shift throughout the day, causing changes to system behavior and bottlenecks. When designing and deploying such systems, it is necessary to determine a process replication and threading policy for the server processes that contain the objects, as well as process activation policies. To avoid bottlenecks, the policy must support all possible workload conditions. Licensing, implementation or resource constraints can limit the number of permitted replicas or threads of a server process. Process activation policies determine whether a server is persistent or should be created and terminated with each call. This paper describes quantitative techniques for choosing process replication or threading levels and process activation policies. Inappropriate policies can lead to unnecessary queuing delays for callers or unnecessarily high consumption of memory resources. The algorithms presented consider all workload conditions, are iterative in nature and are hybrid mathematical programming and analytic performance evaluation methods. An example is given to demonstrate the technique and describe how the results can be applied during software design and deployment.

Measurement Tools and Modeling Techniques for Evaluating Web Server Performance

The past few years have seen a rapid growth in the popularity of the Internet and the World Wide Web. Many companies are deploying Web servers and seeing their usage rates climb rapidly over time. Our research focused on analyzing and evaluating the performance of Internet and intranet Web servers with a goal of creating a Layered Queueing Model to allow capacity planning and performance prediction of next generation server designs. To achieve this we built a tool framework that enables us to collect and analyze empirical data necessary to accomplish our goals.