Arif Merchant

Adaptive control of virtualized resources in utility computing environments

Data centers are often under-utilized due to over-provisioning as well as time-varying resource demands of typical enterprise applications. One approach to increase resource utilization is to consolidate applications in a shared infrastructure using virtualization. Meeting application-level quality of service (QoS) goals becomes a challenge in a consolidated environment as application resource needs differ. Furthermore, for multi-tier applications, the amount of resources needed to achieve their QoS goals might be different at each tier and may also depend on availability of resources in other tiers. In this paper, we develop an adaptive resource control system that dynamically adjusts the resource shares to individual tiers in order to meet application-level QoS goals while achieving high resource utilization in the data center. Our control system is developed using classical control theory, and we used a black-box system modeling approach to overcome the absence of first principle models for complex enterprise applications and systems. To evaluate our controllers, we built a testbed simulating a virtual data center using Xen virtual machines. We experimented with two multi-tier applications in this virtual data center: a two-tier implementation of RUBiS, an online auction site, and a two-tier Java implementation of TPC-W. Our results indicate that the proposed control system is able to maintain high resource utilization and meets QoS goals in spite of varying resource demands from the applications.

Automated control of multiple virtualized resources

Virtualized data centers enable sharing of resources among hosted applications. However, it is difficult to satisfy service-level objectives(SLOs) of applications on shared infrastructure, as application workloads and resource consumption patterns change over time. In this paper, we present AutoControl, a resource control system that automatically adapts to dynamic workload changes to achieve application SLOs. AutoControl is a combination of an online model estimator and a novel multi-input, multi-output (MIMO) resource controller. The model estimator captures the complex relationship between application performance and resource allocations, while the MIMO controller allocates the right amount of multiple virtualized resources to achieve application SLOs. Our experimental evaluation with RUBiS and TPC-W benchmarks along with production-trace-driven workloads indicates that AutoControl can detect and mitigate CPU and disk I/O bottlenecks that occur over time and across multiple nodes by allocating each resource accordingly. We also show that AutoControl can be used to provide service differentiation according to the application priorities during resource contention.

Sinfonia: a new paradigm for building scalable distributed systems

We propose a new paradigm for building scalable distributed systems. Our approach does not require dealing with message-passing protocols -- a major complication in existing distributed systems. Instead, developers just design and manipulate data structures within our service called Sinfonia. Sinfonia keeps data for applications on a set of memory nodes, each exporting a linear address space. At the core of Sinfonia is a novel minitransaction primitive that enables efficient and consistent access to data, while hiding the complexities that arise from concurrency and failures. Using Sinfonia, we implemented two very different and complex applications in a few months: a cluster file system and a group communication service. Our implementations perform well and scale to hundreds of machines.

Minerva: An automated resource provisioning tool for large-scale storage systems

Enterprise-scale storage systems, which can contain hundreds of host computers and storage devices and up to tens of thousands of disks and logical volumes, are difficult to design. The volume of choices that need to be made is massive, and many choices have unforeseen interactions. Storage system design is tedious and complicated to do by hand, usually leading to solutions that are grossly over-provisioned, substantially under-performing or, in the worst case, both.To solve the configuration nightmare, we present minerva: a suite of tools for designing storage systems automatically. Minerva uses declarative specifications of application requirements and device capabilities; constraint-based formulations of the various sub-problems; and optimization techniques to explore the search space of possible solutions.This paper also explores and evaluates the design decisions that went into Minerva, using specialized micro- and macro-benchmarks. We show that Minerva can successfully handle a workload with substantial complexity (a decision-support database benchmark). Minerva created a 16-disk design in only a few minutes that achieved the same performance as a 30-disk system manually designed by human experts. Of equal importance, Minerva was able to predict the resulting systems performance before it was built.

FAB: building distributed enterprise disk arrays from commodity components

This paper describes the design, implementation, and evaluation of a Federated Array of Bricks (FAB), a distributed disk array that provides the reliability of traditional enterprise arrays with lower cost and better scalability. FAB is built from a collection of bricks, small storage appliances containing commodity disks, CPU, NVRAM, and network interface cards. FAB deploys a new majority-voting-based algorithm to replicate or erasure-code logical blocks across bricks and a reconfiguration algorithm to move data in the background when bricks are added or decommissioned. We argue that voting is practical and necessary for reliable, high-throughput storage systems such as FAB. We have implemented a FAB prototype on a 22-node Linux cluster. This prototype sustains 85MB/second of throughput for a database workload, and 270MB/second for a bulk-read workload. In addition, it can outperform traditional master-slave replication through performance decoupling and can handle brick failures and recoveries smoothly without disturbing client requests.

An analytic behavior model for disk drives with readahead caches and request reordering

Modern disk drives read-ahead data and reorder incoming requests in a workload-dependent fashion. This improves their performance, but makes simple analytical models of them inadequate for performance prediction, capacity planning, workload balancing, and so on. To address this problem we have developed a new analytic model for disk drives that do readahead and request reordering. We did so by developing performance models of the disk drive components (queues, caches, and the disk mechanism) and a workload transformation technique for composing them. Our model includes the effects of workload-specific parameters such as request size and spatial locality. The result is capable of predicting the behavior of a variety of real-world devices to within 17% across a variety of workloads and disk drives.

pClock: an arrival curve based approach for QoS guarantees in shared storage systems

Storage consolidation is becoming an attractive paradigm for data organization because of the economies of sharing and the ease of centralized management. However, sharing of resources is viable only if applications can be isolated from each other. This work targets the problem of providing performance guarantees to an application irrespective of the behavior of other workloads. Application requirements are represented in terms of the average throughput, latency and maximum burst size. Most earlier schemes only do weighted bandwidth allocation; schemes that provide control of latency either cannot handle bursts or penalize applications for their own prior behavior, such as using spare capacity. Our algorithm pClock is based on arrival curves that intuitively capture the bandwidth and burst requirements of applications. We show analytically that an application following its arrival curve never misses its deadline. We have implemented pClock both in DiskSim and as a module in the Linux kernel 2.6. Our evaluation shows three important features of pClock: (1) benefits over existing algorithms; (2) efficient performance isolation and burst handling; and (3) the ability to allocate spare capacity to either speed up some applications or to a background utility, such as backup. pClock can be efficiently implemented in a system without much overhead.

Issues and challenges in the performance analysis of real disk arrays

The performance modeling and analysis of disk arrays is challenging due to the presence of multiple disks, large array caches, and sophisticated array controllers. Moreover, storage manufacturers may not reveal the internal algorithms implemented in their devices, so real disk arrays are effectively black-boxes. We use standard performance techniques to develop an integrated performance model that incorporates some of the complexities of real disk arrays. We show how measurement data and baseline performance models can be used to extract information about the various features implemented in a disk array. In this process, we identify areas for future research in the performance analysis of real disk arrays.

A modular, analytical throughput model for modern disk arrays

Enterprise storage systems depend on disk arrays for their capacity and availability needs. To design and maintain storage systems that efficiently satisfy evolving requirements, it is critical to be able to evaluate configuration alternatives without having to physically implement them. In this paper, we describe an analytical model to predict disk array throughput, based on a hierarchical decomposition of the internal array architecture. We validate the model against a state-of-the-art disk array for a variety of synthetic workloads and array configurations. To our knowledge, no previously published analytical model has either incorporated the combined effects of the complex optimizations present in modern disk arrays, or been validated against a real, commercial array. Our results are quite encouraging for an analytical model: predictions are accurate in most cases within 32% of the observed array performance (15% on the average) for our set of experiments.

Analytic modeling of clustered RAID with mapping based on nearly random permutation

A Redundant Array of Independent Disks (RAID) of G disks provides protection against single disk failures by adding one parity block for each G-1 data blocks. In a clustered RAID, the G data/parity blocks are distributed over a cluster of C disks (C>G), thus reducing the additional load on each disk due to a single disk failure. However, most methods proposed for implementing such a mapping do not work for general C and G values. In this paper, we describe a fast mapping algorithm based on almost-random permutations. An analytical model is constructed, based on the queue with a permanent customer, to predict recovery time and read/write performance. The accuracy of the results derived from this model is validated by comparing with simulations. Our analysis shows that clustered RAID is significantly more tolerant of disk failure than the basic RAID scheme. Both recovery time and performance degradation during recovery are substantially reduced in clustered RAID; moreover, these gains can be achieved using fairly small C/G ratios.