Browsing by Author "Iyer, Sitaram"

Now showing 1 - 4 of 4
  • Thumbnail Image
    Item
    A Resource Management Framework for Predictable Quality of Service in Web Servers
    (2003-07-07) Aron, Mohit; Druschel, Peter; Iyer, Sitaram
    This paper presents a resource management framework for providing predictable quality of service (QoS) in Web servers. The framework allows Web server and proxy operators to ensure a probabilistic minimal QoS level, expressed as an average request rate, for a certain class of requests (called a service), irrespective of the load imposed by other requests. A measurement-based admission control framework determines whether a service can be hosted on a given server or proxy, based on the measured statistics of the resource consumptions and the desired QoS levels of all the co-located services. In addition, we present a feedback-based resource scheduling framework that ensures that QoS levels are maintained among admitted, co-located services. Experimental results obtained with a prototype implementation of our framework on trace-based workloads show its effectiveness in providing desired QoS levels with high confidence, while achieving high average utilization of the hardware.
  • Thumbnail Image
    Item
    Advanced memory management and disk scheduling techniques for general-purpose operating systems
    (2006) Iyer, Sitaram; Druschel, Peter
    Operating systems have evolved into sophisticated, high-performance virtualizing platforms, to support and be fair towards concurrently running applications. However, since applications usually run oblivious of each other and prefer narrow system interfaces, they inadvertently contend for resources, resulting in inappropriate allocations and significant performance degradations. This dissertation identifies and eliminates two such problems: one we call rigidity in physical memory management which we solve using adaptive memory management, and a second we call deceptive idleness in disk schedulers that we solve through anticipatory disk scheduling . Many applications, their libraries, and runtimes can trade memory consumption for performance by maintaining caches, triggering garbage collection, etc. However, due to ignorance of memory pressure in the system, they are forced to be conservative about memory usage. Adaptive memory management is a technique that informs applications of the severity of memory pressure via a metric that quantifies the cost of using memory. This enables applications to allocate memory liberally when available (with performance benefits of 20% to 300%); and to release it under contention. The system thus reaches an equilibrium that balances the impact of memory pressure on each application; adapts to avoid paging during load bursts and improves stability and responsiveness; and reduces the need for manual configuration of memory footprints. It also provides finer control on memory usage by adapting proportional to application priorities. Disk schedulers generally schedule a request as soon as the previous request has finished. Unfortunately, many applications perform synchronous I/O by issuing a request after the previous request has been served. This causes the scheduler to suffer from deceptive idleness, a condition where it incorrectly assumes that the process has no further requests, and seeks to a request from another process. Anticipatory dish scheduling transparently solves this problem by sometimes injecting a small, controlled delay into the disk scheduler before it makes a scheduling decision, whenever it expects the current request to be quickly followed by another nearby request. This improves performance by up to 70% and enables proportional schedulers to achieve their contracts. Anticipatory scheduling has been ported to Linux, where it is now the default disk scheduler.
  • Thumbnail Image
    Item
    The Effect of Deceptive Idleness on Disk Schedulers
    (2001-06-19) Druschel, Peter; Iyer, Sitaram
    Disk schedulers in operating systems are generally work-conserving; they schedule a request immediately after the previous request has finished. Such schedulers need multiple outstanding requests to make good decisions. Unfortunately, many applications issue synchronous, almost-continuous streams of read requests. This forces the scheduler into making decisions too early, falsely assuming that the process has become momentarily idle. This phenomenon of deceptive idleness causes significant degradation in performance and quality of service objectives on current systems. We solve deceptive idleness by designing and implementing a transparent, non-work-conserving scheduling framework for various scheduling policies. We evaluate this solution on micro benchmarks and real workloads, and observe large benefits. The Apache web server delivers 56% and 16% more throughput for two configurations. The Andrew Benchmark runs faster by 8% (54% for the read-intensive phase). Variants of the TPC-B database benchmark exhibit improvements between 4% and 60%. Proportional-share schedulers become empowered to efficiently deliver application-desired proportions.
  • Thumbnail Image
    Item
    The effect of deceptive idleness on disk schedulers
    (2001) Iyer, Sitaram; Druschel, Peter
    Disk schedulers in operating systems are generally work-conserving; they schedule a request immediately after the previous request has finished. Such schedulers need multiple outstanding requests to make good decisions. Unfortunately, many applications issue synchronous, almost-continuous streams of read requests. This forces the scheduler into making decisions too early, falsely assuming that the process has become momentarily idle. This phenomenon of deceptive idleness causes significant degradation in performance and quality of service objectives on current systems. We solve deceptive idleness by designing and implementing a transparent, non-work-conserving scheduling framework for various scheduling policies. We evaluate this solution on microbenchmarks and real workloads, and observe large benefits. The Apache webserver delivers 56% and 16% more throughput for two configurations. The Andrew Benchmark runs faster by 8% (54% for the read-intensive phase). Variants of the TPC-B database benchmark exhibit improvements between 4% and 60%. Proportional-share schedulers become empowered to efficiently deliver application-desired proportions.