Action Breakout Session

Slides:



Advertisements
Similar presentations
Scheduling in Web Server Clusters CS 260 LECTURE 3 From: IBM Technical Report.
Advertisements

Distributed Systems Major Design Issues Presented by: Christopher Hector CS8320 – Advanced Operating Systems Spring 2007 – Section 2.6 Presentation Dr.
CROWN “Thales” project Optimal ContRol of self-Organized Wireless Networks WP1 Understanding and influencing uncoordinated interactions of autonomic wireless.
High speed links, distributed services, can’t modify routers  Lack of visibility But, need for more visibility and control  Increased number and complexity.
Web Caching Schemes1 A Survey of Web Caching Schemes for the Internet Jia Wang.
Internet Traffic Patterns Learning outcomes –Be aware of how information is transmitted on the Internet –Understand the concept of Internet traffic –Identify.
PSMC Proxy Server-based Multipath Connection CS 526 Advanced Networking - Richard White.
OCT1 Principles From Chapter One of “Distributed Systems Concepts and Design”
1 Action Breakout Session Anil, AP, Nina Bhatti, Charles Berdnall, Joe Hellerstein, Wei Hu, Anthony Joseph, Randy Katz, Li, Machi Mukund Kimmo Raatikanen,
Using Prices to Allocate Resources at Access Points Jimmy Shih, Randy Katz, Anthony Joseph One Administrative Domain Access Point A Access Point B Network.
ISCSI Performance Experiments Li Yin EECS Department U.C.Berkeley.
Using Prices to Allocate Resources at Access Points Jimmy Shih, Randy Katz, Anthony Joseph One Administrative Domain Access Point A Access Point B Network.
1 Web Content Delivery Reading: Section and COS 461: Computer Networks Spring 2007 (MW 1:30-2:50 in Friend 004) Ioannis Avramopoulos Instructor:
Computer Science Cataclysm: Policing Extreme Overloads in Internet Applications Bhuvan Urgaonkar and Prashant Shenoy University of Massachusetts.
FIREWALL TECHNOLOGIES Tahani al jehani. Firewall benefits  A firewall functions as a choke point – all traffic in and out must pass through this single.
1 Issues in Benchmarking Intrusion Detection Systems Marcus J. Ranum.
Computer Security: Principles and Practice First Edition by William Stallings and Lawrie Brown Lecture slides by Lawrie Brown Chapter 8 – Denial of Service.
Performance of Web Applications Introduction One of the success-critical quality characteristics of Web applications is system performance. What.
SEDA: An Architecture for Well-Conditioned, Scalable Internet Services
This document is for informational purposes only, and Tekelec reserves the right to change any aspect of the products, features or functionality described.
Web Application Firewall (WAF) RSA ® Conference 2013.
1 06/00 Questions 10/6/2015 QoS in DOS ECOOP 2000John Zinky BBN Technologies ECOOP 2000 Workshop on Quality of Service in Distributed Object Systems
What are the main differences and commonalities between the IS and DA systems? How information is transferred between tasks: (i) IS it may be often achieved.
Challenges towards Elastic Power Management in Internet Data Center.
10/20/2015 ©2006 Scott Miller, University of Victoria 1 User Authentication Content Generation The Use of Cookies Content Pooling Rev 1.5.
Denial of Service Attack 발표자 : 전지훈. What is Denial of Service Attack?  Denial of Service Attack = DoS Attack  Service attacks on a Web server floods.
Jennifer Rexford Fall 2014 (TTh 3:00-4:20 in CS 105) COS 561: Advanced Computer Networks TCP.
1 BBN Technologies Quality Objects (QuO): Adaptive Management and Control Middleware for End-to-End QoS Craig Rodrigues, Joseph P. Loyall, Richard E. Schantz.
6.894: Distributed Operating System Engineering Lecturers: Frans Kaashoek Robert Morris
KEYNOTE OF THE FUTURE 3: DAVID BECKETT CSIT PhD Student QUEEN’S UNIVERSITY BELFAST.
Spark on Entropy : A Reliable & Efficient Scheduler for Low-latency Parallel Jobs in Heterogeneous Cloud Huankai Chen PhD Student at University of Kent.
INTRODUCTION TO WEB HOSTING
Fail-stutter Behavior Characterization of NFS
Introduction to Operating Systems
VPN Extension Requirements for Private Clouds
University of Maryland College Park
SEDA: An Architecture for Scalable, Well-Conditioned Internet Services
Threads vs. Events SEDA – An Event Model 5204 – Operating Systems.
Software Architecture in Practice
Panagiotis Demestichas
Securing the Network Perimeter with ISA 2004
PROTEAN: A Scalable Architecture for Active Networks
IBTA Developers’ Conference Link Working Group - QoS
Regulating Data Flow in J2EE Application Server
TCP Congestion Control
Introduction to Networks
Measuring Service in Multi-Class Networks
Oracle Solaris Zones Study Purpose Only
Software Architecture in Practice
Storage Virtualization
Auburn University COMP7500 Advanced Operating Systems I/O-Aware Load Balancing Techniques (2) Dr. Xiao Qin Auburn University.
Transparent Adaptive Resource Management for Middleware Systems
Introduction to Operating Systems
Network Security: IP Spoofing and Firewall
Load Shedding in Stream Databases – A Control-Based Approach
Storage: Optimize, Monitor, Automate
Reference-Driven Performance Anomaly Identification
Provision of Multimedia Services in based Networks
COS 561: Advanced Computer Networks
Speaker: Jin-Wei Lin Advisor: Dr. Ho-Ting Wu
NeST: Network Storage Technologies
Specialized Cloud Architectures
Chapter 2: Operating-System Structures
Control Theory in Log Processing Systems
EE 122: Lecture 22 (Overlay Networks)
Building Topic/Trend Detection System based on Slow Intelligence
Chapter 2: Operating-System Structures
Performance-Robust Parallel I/O
The Dark Side of the Internet
Modeling and Evaluating Variable Bit rate Video Steaming for ax
Presentation transcript:

Action Breakout Session Anil, AP, Nina Bhatti, Charles Berdnall, Joe Hellerstein, Wei Hu, Anthony Joseph, Randy Katz, Li, Machi Mukund Kimmo Raatikanen, Siva

Breakout Goal Identify research questions and issues related to adaptive action invocation to enhance the dependability and security of distributed systems Customer is the “system administrator,” not the end user

Breakout Process Define actions by example Discuss cross-layer interaction and coordination Distill underlying principles

Key Observations Distinguish between control actions (e.g., “slow down”) and data actions (e.g., “drop packets”) Distinguish between internal/locally performed actions and actions that affect global behavior Control loops operating in multiple levels, regionally and globally Performance-related actions are the basic building block Control system itself can be target of an adversarial attack

Working Examples Network Storage Service; Media Streaming Service Multiple instances of service various places in network Direct requests to best available service instance Balance requests among service instances Fall back to alternative service instance in the face of failure or DOS attack Coordinate measurements on client-side and server-side to reduce load through admission control and content adaptation Distinguish between server overload and network overload For clients “not in the loop” (heterogeneous clients, adversarial clients), proxy the necessary behavior inside the network Network Denial of Service Overload data traffic and starve control traffic Secondary performance effects: session resets, router CPUs driven to high utilization, etc.

Control Theoretic Viewpoint Black boxes that are managed by a control system Actuation points that can acted upon to control the system E.g., Apply backpressure to clients to slow down request rate (control); degrade content quality (data) E.g., Prioritize/reserve bandwidth for control traffic; Policy settings are control actions, enforcement of policy are data actions Single vs. independent control loops: which is better? Theory provides tools for managing “disturbances” Note that the control system can itself be the target of attack Hellerstein: Action is a change to a configuration E.g., buffer pool size, weights in load balancer E.g., uninstall/reinstall software

General Observations Causality and Visibility Actions can lead to cascaded actions Can interactions/side effects be modeled/made explicit? Action graph model: probability that a following action will be invoked as the result of a given current action In general, difficult to determine in advance Could it be learned via observe/analyze? Feasible to place action points at every potential bottleneck site? Note that routers are badly designed black boxes, difficult and time consuming to extract their internal state Tradeoff between centralized collection of state that may be “complete” but out-of-date vs. decentralized collection that may be more timely but globally incomplete Principle of containment: first do no harm, local actions potential less disastrous than global actions

General Observations Managing Disturbances Instabilities arise where delays in taking action are introduced Latencies in response Imperfect knowledge of the state Tradeoff in making decisions based on longer intervals spanning more state vs. shorter intervals spanning less state Time intervals adapt … short time to ensure useful work always being done E.g., Disk scheduling in Storage Server You can only do work you are aware of Keep the queues short to achieve best performance

General Observations Predictive actions Waiting too long to detect problem limits ability to respond Characterize workload/response changes as signature of impending system performance failure Response to workload changes: “gradual” vs. cliff degradation E.g., as I/O workload grows, predict increases in response latency E.g., IBM detects changes to slope of activity to trigger resource allocation to manage flash crowds in web server farms

General Observations Don’t ignore the human decision maker Human operators in the loop Research challenge: visualizing the configuration and state of the system to a human decision maker Higher order configuration and administration tools and frameworks