Ricardo Koller

I/O Deduplication: Utilizing content similarity to improve I/O performance

Duplication of data in storage systems is becoming increasingly common. We introduce I/O Deduplication, a storage optimization that utilizes content similarity for improving I/O performance by eliminating I/O operations and reducing the mechanical delays during I/O operations. I/O Deduplication consists of three main techniques: content-based caching, dynamic replica retrieval, and selective duplication. Each of these techniques is motivated by our observations with I/O workload traces obtained from actively-used production storage systems, all of which revealed surprisingly high levels of content similarity for both stored and accessed data. Evaluation of a prototype implementation using these workloads showed an overall improvement in disk I/O performance of 28 to 47% across these workloads. Further breakdown also showed that each of the three techniques contributed significantly to the overall performance improvement.

WattApp: an application aware power meter for shared data centers

The increasing heterogeneity between applications in emerging virtualized data centers like clouds introduce significant challenges in estimating the power drawn by the data center. In this work, we presentWattApp: an application-aware power meter for shared data centers that addresses this challenge. In order to deal with heterogeneous applications, WattApp introduces application parameters (e.g, throughput) in the power modeling framework. WattApp is based on a carefully designed set of experiments on a mix of diverse applications: power benchmarks, web-transaction workloads, HPC workloads and I/O-intensive workloads. Given a set of N applications and M server types, WattApp runs in O(N) time, uses O(NxM) calibration runs, and predicts the power drawn by any arbitrary placement within 5%of the real power for the applications studied.

CosMig: Modeling the Impact of Reconfiguration in a Cloud

Clouds allow enterprises to increase or decrease their resource allocation on demand in response to changes in workload intensity. Virtualization is one of the building blocks for cloud computing and provides the mechanisms to implement the dynamic allocation of resources. These dynamic reconfiguration actions lead to performance impact during the reconfiguration duration. In this paper, we model the cost of reconfiguring a cloud-based IT infrastructure in response to workload variations. We show that maintaining a cloud requires frequent reconfigurations necessitating both VM resizing and VM live migration, with live migration dominating reconfiguration costs. We design the CosMig model to predict the duration of live migration and its impact on application performance. Our model is based on parameters that are typically monitored in enterprise data centers. Further, the model faithfully captures the impact of shared resources in a virtualized environment. We experimentally validate the accuracy and effectiveness of CosMig using micro benchmarks and representative applications.

The cost of reconfiguration in a cloud

Emerging clouds promise enterprises the ability to increase or decrease their resource allocation on demand using virtual machine resizing and migration. These dynamic reconfiguration actions lead to performance impact during the reconfiguration duration. In this paper, we study the cost of reconfiguring a cloud-based IT infrastructure in response to workload variations. We observe that live migration requires a significant amount of spare CPU on the source server (but not on the target server). If spare CPU is not available, it impacts both the duration of migration and the performance of the application being migrated. Further, the amount of CPU required for live migration varies with the active memory of the VM being migrated. Finally, we show that live migration may impact any co-located VMs based on the cache usage pattern of the co-located VM. We distill all our observations to present a list of practical recommendations to cloud providers for minimizing the impact of reconfiguration during dynamic resource allocation.

Anatomy of a Real-Time Intrusion Prevention System

Host intrusion prevention systems for both servers and end-hosts must address the dual challenges of accuracy and performance. Researchers have mostly focused on addressing the former challenge, suggesting solutions based either on exploit- based penetration detection or anomaly-based misbehavior detection, but yet stopping short of comprehensive solutions that leverage merits of both approaches. The second challenge, however, is rarely addressed; doing so comprehensively is important since these systems can introduce substantial overhead and cause system slowdown, more so when the system load is high. We present Rootsense, a holistic and real-time intrusion prevention system that combines the merits of misbehavior- based and anomaly-based detection. Four principles govern the design and implementation of Rootsense. First, Rootsense audits events within different subsystems of the host operating system and correlates them to comprehensively capture the global system state. Second, Rootsense restricts the detection domain to root compromises only; doing so reduces run-time overhead and increases detection accuracy (root behavior is more easily modeled than user behavior). Third, Rootsense adopts a dual approach to intrusion detection - a root penetration detector detects activities that exploit system vulnerabilities to penetrate the security perimeter, and a root misbehavior detector tracks misbehavior by root processes. Fourth, Rootsense is designed to be configurable for overhead management allowing the system administrator to tune the overhead characteristics of the intrusion prevention system that affect foreground task performance. A Linux implementation of Rootsense is analyzed for both accuracy and performance, using several real-world exploits and a range of end-host and server benchmarks.

Centaur: Host-Side SSD Caching for Storage Performance Control

Host-side SSD caches represent a powerful knob for improving and controlling storage performance and improve performance isolation. We present Centaur, as a host-side SSD caching solution that uses cache sizing as a control knob to achieve storage performance goals. Centaur implements dynamically partitioned per-VM caches with per-partition local replacement to provide both lower cache miss rate, better performance isolation and performance control for VM workloads. It uses SSD cache sizing as a universal knob for meeting a variety of workload-specific goals including per-VM latency and IOPS reservations, proportional share fairness, and aggregate optimizations such as minimizing the average latency across VMs. We implemented Centaur for the VMware ESX hyper visor. With Centaur, times for simultaneously booting 28 virtual desktops improve by 42% relative to a non-caching system and by 18% relative to a unified caching system. Centaur also implements per-VM shares for latency with less than 5% error when running micro benchmarks, and enforces latency and IOPS reservations on OLTP workloads with less than 10% error.

Generalized ERSS tree model: Revisiting working sets

Accurately characterizing the resource usage of an application at various levels in the memory hierarchy has been a long-standing research problem. Existing characterization studies are either motivated by specific allocation problems (e.g., memory page allocation) or they characterize a specific memory resource (e.g., L2 cache). The studies thus far have also implicitly assumed that there is no contention for the resource under consideration. The inevitable future of virtualization driven consolidation necessitates the sharing of physical resources at all levels of the memory hierarchy by multiple virtual machines (VMs). Given the lack of resource isolation mechanisms at several levels of the memory hierarchy within current commodity systems, provisioning resources for a virtualized application not only requires a precise characterization of its resource usage but must also account for the impact of resource contention due to other co-located applications during its lifetime. In this paper, we present a unifying Generalized ERSS Tree Model that characterizes the resource usage at all levels of the memory hierarchy during the entire lifetime of an application. Our model characterizes capacity requirements, the rate of use, and the impact of resource contention, at each level of memory. We present a methodology to build the model and demonstrate how it can be used for the accurate provisioning of the memory hierarchy in a consolidated environment. Empirical results suggest that the Generalized ERSS Tree Model is effective at characterizing applications with a wide variety of resource usage behaviors.

The case for active block layer extensions

Self-managing storage systems have recently received attention from the research community due to their promised ability of continuously adapting to best reflect high-level system goal specifications. However, this eventuality is currently being met by both conceptual and practical challenges that threaten to slow down the pace of innovation. We argue that two fundamental directions will help evolve the state of self-managing storage systems: (i) a standardized development environment for self-management extensions that also addresses ease of deployment, and (ii) a theoretical framework for reasoning about behavioral properties of individual and collective self-management extensions. We propose Active Block Layer Extensions (ABLE), an operating system infrastructure that aids the development and manages the deployed instances of self-management extensions within the storage stack. ABLE develops a theory behind block layer extensions that helps address key questions about overall storage stack behavior, data consistency, and reliability. We exemplify specific storage self-management solutions that can be built as stackable extensions using ABLE. Our initial experience with ABLE and few block layer extensions that we have been building, leads to believe that the ABLE infrastructure can substantially simplify the development and deployment of robust, self-managing, storage systems.

Estimating Application Cache Requirement for Provisioning Caches in Virtualized Systems

Miss rate curves (MRCs) are a fundamental concept in determining the impact of caches on an applications performance. In our research, we use MRCs to provision caches for applications in a consolidated environment. Current techniques for building MRCs at the CPU caches level require changes to the applications and are restricted to a few processor architectures [7], [22]. In this work, we investigate two techniques to partition shared L2 and L3 caches in a server and build MRCs for the VMs. These techniques make different trade-offs across accuracy, flexibility, and intrusiveness dimensions. The first technique is based on operating system (OS) page coloring and does not require change in commodity hardware or application. We improve upon existing page-coloring based approaches by identifying and overcoming a subtle but real problem of unequal associative cache sets loading to implement accurate cache allocation. Our second technique called Cache Grabber is even less intrusive and requires no changes in hardware, OS, or application. We present a comprehensive evaluation of the relative merits of these and other techniques to estimate MRCs. Our evaluation study enables a data center administrator to select the technique most suitable to his (her) specific data center to provision caches for consolidated applications.