300 likes | 551 Views
Information and Control in Gray-Box Systems. Arpaci-Dusseau and Arpaci-Dusseau SOSP 18, 2001 John Otto Wi06 CS 395/495 Autonomic Computing Systems. Overview. OS and Gray-Box Advantages Techniques Previous Approaches Case-Studies Gray Toolbox Autonomic Perspective. What is Gray-Box?.
E N D
Information and Control in Gray-Box Systems Arpaci-Dusseau and Arpaci-Dusseau SOSP 18, 2001 John Otto Wi06 CS 395/495 Autonomic Computing Systems
Overview • OS and Gray-Box Advantages • Techniques • Previous Approaches • Case-Studies • Gray Toolbox • Autonomic Perspective
What is Gray-Box? • Premise • Operating systems cannot be easily modified without performance risks • Goal • Incorporate new, “special application” OS ideas into systems without modifying the OS itself • Method • Using knowledge of OS algorithms, observe the OS “state” and present an optimized interface for the user (the Information and Control Layer, ICL)
General Capabilities • Applications do not necessarily need to be designed to interface with the ICL • Easy to port—ICLs usually assume an algorithm and perform general tests to determine the OS state.
Overview • OS and Gray-Box Advantages • Techniques • Previous Approaches • Case-Studies • Gray Toolbox • Autonomic Perspective
Gaining Information • Obtain Algorithmic Knowledge • Trade-off between generality and optimization • Monitor Outputs • Information in “covert channels” implies state • Use Statistical Methods • Generate a “system profile” to distinguish normal and abnormal system performance • Use Microbenchmarks • Judiciously conduct performance tests on the system • Insert Probes • Probes help obtain, but also modify, the system state
Asserting Control • Exploit algorithmic knowledge to simply achieve a goal • e.g. prefetching a file • Move the system to a known state • Implement feedback systems • Repeated use should optimize the ICL • Design should keep OS in known state
Overview • OS and Gray-Box Advantages • Techniques • Previous Approaches • Case-Studies • Gray Toolbox • Autonomic Perspective
Existing Microbenchmarks • Typically run in a controlled environment • Collect static data • Time restrictions are not imposed • Hence, they do not offer insight into the unknown state of a system—only static parameters
Existing Gray-Box Systems • Capabilities • TCP: diagnose network congestion • Implicit Coscheduling: run communicating processes concurrently • MS Manners: optimize resource (CPU) availability for important processes
Overview • OS and Gray-Box Advantages • Techniques • Previous Approaches • Case-Studies • Gray Toolbox • Autonomic Perspective
File-Cache Content Detector • Goal • Order data accesses to maximize cache hits, minimize disk accesses • Methods • Internal Simulation vs. Inference by Observation • Simulation expensive, requires all processes to cooperate • Exploit spatial locality (page loading algorithms) • Probing one region of a file can indicate whether that region of the file is in cache • Limitations • Probing small files significantly alters the cache state of that file
FCCD: Implementation and Interface • Resilient Interface • Library: built-in application ICL functionality • Command line: orders a list of files passed to command line tool • Implementation • Differentiation between cache hit and miss • Sort files/regions of a file by shortest probe access time • Choice of Access Unit size—minimize disk seek time • Choice of Prediction Unit size—minimize probe use • Perform a few probes per access unit (prediction unit smaller than access unit) • Select random byte in prediction unit
File Layout Detector and Controller • Goal • To ascertain the layout on disk of a set of files • “Gray-Box” Knowledge • Most file systems localize contents of a directory on the same set of disk cylinders • Methods • Refresh directory structure • Use knowledge of i-node assignment to order file accesses • Implementation • Call stat() on each file • Refresh the directory • Return list of files sorted by i-node • Limitations • UNIX-oriented optimization (i-nodes!) • Dependence of other running applications on i-node numbers
Memory-based Admission Control • Goal • Prevent overuse of memory resources • Methods • Measure amount of memory that can be referenced without causing a page replacement • Applications are notified when there is not enough free memory for an allocation request • Limitations • Accuracy limited by page-replacement algorithm • Just because the MAC application is “nice” doesn’t mean that other applications can’t cause thrashing.
Overview • OS and Gray-Box Advantages • Techniques • Previous Approaches • Case-Studies • Gray Toolbox • Autonomic Perspective
Gray Toolbox • Microbenchmark results stored in common repository for use by ICLs at system level • Overhead-sensitive operations use system-optimized “plug-in” functionality • e.g. timers • Provide tools for simple statistical calculations
Overview • OS and Gray-Box Advantages • Techniques • Previous Approaches • Case-Studies • Gray Toolbox • Autonomic Perspective?
Autonomic Perspective—Observations • Knowledge: In order for an autonomic tool to function well, the state of the system must be well-known. • Hence, keeping the system in a known state is an important objective for autonomic tools. • Trust: If a system can provide evidenceand reasonsfor its actions, a user is more likely to trust the system. • A user interface detailing decisions and the benchmarks leading to an action would be beneficial. • Simplicity: Autonomic systems should operate based on known algorithms; actions would be predictable and explainable.
Information and Control in Gray-Box Systems Arpaci-Dusseau and Arpaci-Dusseau SOSP 18, 2001 John Otto Wi06 CS 395/495 Autonomic Computing Systems