Carnegie Mellon University Computer Science Department 1 OPEN VERSUS CLOSED: A CAUTIONARY TALE Bianca Schroeder Adam Wierman Mor Harchol-Balter Computer.

Slides:



Advertisements
Similar presentations
Introduction to Queuing Theory
Advertisements

Walter Binder University of Lugano, Switzerland Niranjan Suri IHMC, Florida, USA Green Computing: Energy Consumption Optimized Service Hosting.
CPU Scheduling Questions answered in this lecture: What is scheduling vs. allocation? What is preemptive vs. non-preemptive scheduling? What are FCFS,
Anshul Gandhi (Carnegie Mellon University) Varun Gupta (CMU), Mor Harchol-Balter (CMU) Michael Kozuch (Intel, Pittsburgh)
1 Size-Based Scheduling Policies with Inaccurate Scheduling Information Dong Lu *, Huanyuan Sheng +, Peter A. Dinda * * Prescience Lab, Dept. of Computer.
Silberschatz, Galvin and Gagne  2002 Modified for CSCI 399, Royden, Operating System Concepts Operating Systems Lecture 19 Scheduling IV.
ANALYZING STORAGE SYSTEM WORKLOADS Paul G. Sikalinda, Pieter S. Kritzinger {psikalin, DNA Research Group Computer Science Department.
1 Web Server Performance in a WAN Environment Vincent W. Freeh Computer Science North Carolina State Vsevolod V. Panteleenko Computer Science & Engineering.
Simulation Evaluation of Hybrid SRPT Policies
Maryam Elahi Fairness in Speed Scaling Design Joint work with: Carey Williamson and Philipp Woelfel.
DotSlash – A Web Hotspot Rescue System Weibin Zhao Henning Schulzrinne Department of Computer Science Columbia University June 11, 2004.
A Hierarchical Characterization of a Live Streaming Media Workload E. Veloso, V. Almeida W. Meira, A. Bestavros, S. Jin Proceedings of Internet Measurement.
OS Fall ’ 02 Performance Evaluation Operating Systems Fall 2002.
1 Performance Evaluation of Computer Networks Objectives  Introduction to Queuing Theory  Little’s Theorem  Standard Notation of Queuing Systems  Poisson.
Effects and Implications of File Size/Service Time Correlation on Web Server Scheduling Policies Dong Lu* + Peter Dinda* Yi Qiao* Huanyuan Sheng* *Northwestern.
Looking at the Server-side of P2P Systems Yi Qiao, Dong Lu, Fabian E. Bustamante and Peter A. Dinda Department of Computer Science Northwestern University.
1 Connection Scheduling in Web Servers Mor Harchol-Balter School of Computer Science Carnegie Mellon
Performance Evaluation
Queueing Network Model. Single Class Model Open - Infinite stream of arriving customers Closed - Finite population eg Intranet users Indistinguishable.
Carnegie Mellon University Computer Science Department 1 CLASSIFYING SCHEDULING POLICIES WITH RESPECT TO HIGHER MOMENTS OF CONDITIONAL RESPONSE TIME Adam.
OS Fall ’ 02 Performance Evaluation Operating Systems Fall 2002.
Introduction to Queuing Theory. 2 Queuing theory definitions  (Kleinrock) “We study the phenomena of standing, waiting, and serving, and we call this.
1 Mor Harchol-Balter Carnegie Mellon University Joint work with Bianca Schroeder.
Understanding Factors That Influence Performance of a Web Server Presentation CS535 Project By Thiru.
Cmpt-225 Simulation. Application: Simulation Simulation  A technique for modeling the behavior of both natural and human-made systems  Goal Generate.
Can Internet Video-on-Demand Be Profitable? SIGCOMM 2007 Cheng Huang (Microsoft Research), Jin Li (Microsoft Research), Keith W. Ross (Polytechnic University)
Efficient Scheduling of Heterogeneous Continuous Queries Mohamed A. Sharaf Panos K. Chrysanthis Alexandros Labrinidis Kirk Pruhs Advanced Data Management.
Chapter 1: Introduction to Web
1 SCHEDULING FOR TODAY’S COMPUTER SYSTEMS: SCHEDULING FOR TODAY’S COMPUTER SYSTEMS: BRIDGING THEORY AND PRACTICE Adam Wierman Mor Harchol-Balter John.
Design and Implement an Efficient Web Application Server Presented by Tai-Lin Han Date: 11/28/2000.
(C) 2009 J. M. Garrido1 Object Oriented Simulation with Java.
Simulation Examples ~ By Hand ~ Using Excel
U NIVERSITY OF M ASSACHUSETTS, A MHERST – Department of Computer Science An Analytical Model for Multi-tier Internet Services and its Applications Bhuvan.
Freshness-Aware Scheduling of Continuous Queries in the Dynamic Web Mohamed A. Sharaf Alexandros Labrinidis Panos K. Chrysanthis Kirk Pruhs Advanced Data.
Management of Waiting Lines McGraw-Hill/Irwin Copyright © 2012 by The McGraw-Hill Companies, Inc. All rights reserved.
Modeling and Performance Evaluation of Network and Computer Systems Introduction (Chapters 1 and 2) 10/4/2015H.Malekinezhad1.
Budget-based Control for Interactive Services with Partial Execution 1 Yuxiong He, Zihao Ye, Qiang Fu, Sameh Elnikety Microsoft Research.
Classification of scheduling policies Preemptive methods (typical representative: RR) Non-preemptive methods (typical representative: FCFS) Preemption.
NETE4631:Capacity Planning (2)- Lecture 10 Suronapee Phoomvuthisarn, Ph.D. /
Entities and Objects The major components in a model are entities, entity types are implemented as Java classes The active entities have a life of their.
CPU Scheduling CSCI 444/544 Operating Systems Fall 2008.
1 Queuing Systems (2). Queueing Models (Henry C. Co)2 Queuing Analysis Cost of service capacity Cost of customers waiting Cost Service capacity Total.
1 The Effect of Heavy-Tailed Job Size Distributions on System Design Mor Harchol-Balter MIT Laboratory for Computer Science.
1 Challenges in Scaling E-Business Sites  Menascé and Almeida. All Rights Reserved. Daniel A. Menascé Department of Computer Science George Mason.
ICOM 6115: Computer Systems Performance Measurement and Evaluation August 11, 2006.
An Optimal Service Ordering for a World Wide Web Server A Presentation for the Fifth INFORMS Telecommunications Conference March 6, 2000 Amy Csizmar Dalal.
Performance Analysis of Real Traffic Carried with Encrypted Cover Flows Nabil Schear David M. Nicol University of Illinois at Urbana-Champaign Department.
Simulation Examples and General Principles
1 Admission Control and Request Scheduling in E-Commerce Web Sites Sameh Elnikety, EPFL Erich Nahum, IBM Watson John Tracey, IBM Watson Willy Zwaenepoel,
OPERATING SYSTEMS CS 3530 Summer 2014 Systems with Multi-programming Chapter 4.
Chapter 10 Verification and Validation of Simulation Models
Improving Disk Throughput in Data-Intensive Servers Enrique V. Carrera and Ricardo Bianchini Department of Computer Science Rutgers University.
Measuring the Capacity of a Web Server USENIX Sympo. on Internet Tech. and Sys. ‘ Koo-Min Ahn.
OPERATING SYSTEMS CS 3530 Summer 2014 Systems and Models Chapter 03.
Chapter 10 Verification and Validation of Simulation Models Banks, Carson, Nelson & Nicol Discrete-Event System Simulation.
Scheduling MPI Workflow Applications on Computing Grids Juemin Zhang, Waleed Meleis, and David Kaeli Electrical and Computer Engineering Department, Northeastern.
(C) J. M. Garrido1 Objects in a Simulation Model There are several objects in a simulation model The activate objects are instances of the classes that.
1 Mor Harchol-Balter Carnegie Mellon with Nikhil Bansal with Bianca Schroeder with Mukesh Agrawal.
Management of Waiting Lines Copyright © 2015 McGraw-Hill Education. All rights reserved. No reproduction or distribution without the prior written consent.
Queuing Theory Simulation & Modeling.
Building Valid, Credible & Appropriately Detailed Simulation Models
Test Loads Andy Wang CIS Computer Systems Performance Analysis.
System Simulation (CAP 4800) May 30, of xx Notes on Barford SURGE paper Ken Christensen Department of Computer Science and Engineering College of.
Dynamic Resource Allocation for Shared Data Centers Using Online Measurements By- Abhishek Chandra, Weibo Gong and Prashant Shenoy.
Chapter-04 Building an Ecommerce Website. Building an E-commerce Site: A Systematic Approach The two most important management challenges in building.
OPERATING SYSTEMS CS 3502 Fall 2017
OPERATING SYSTEMS CS 3502 Fall 2017
Chapter 10 Verification and Validation of Simulation Models
Capacity Analysis, cont. Realistic Server Performance
Approximate Mean Value Analysis of a Database Grid Application
Presentation transcript:

Carnegie Mellon University Computer Science Department 1 OPEN VERSUS CLOSED: A CAUTIONARY TALE Bianca Schroeder Adam Wierman Mor Harchol-Balter Computer Science Department Carnegie Mellon University OPEN VERSUS CLOSED: A CAUTIONARY TALE Bianca Schroeder Adam Wierman Mor Harchol-Balter Computer Science Department Carnegie Mellon University To appear at NSDI 2006 presenter: 吳泰廷

Carnegie Mellon University Computer Science Department 2 standard system new system old new new system has smaller response time! This comparison requires testing the two systems on realistic workloads THE RESEARCH PROCESS

Carnegie Mellon University Computer Science Department 3 INTRODUTION Need system models that “accurately represent" the real system. Representing a system accurately involves many things: bottleneck resource behavior, the scheduling of requests at that bottleneck, workload parameters such as the distribution of service request demands……. One factor that researchers typically pay little attention to is whether the job arrivals obey a closed or an open system model.

Carnegie Mellon University Computer Science Department 4 We show that closed and open system models yield significantly different results, even when both models are run with the same load and service demands. Conclude with guidelines for choosing a system model.

Carnegie Mellon University Computer Science Department 5 think receive send server CLOSED SYSTEM MODEL User requests web page, receives page, reads page, clicks on new link Closed System N=MPL (multiprogramming level)

Carnegie Mellon University Computer Science Department 6 1:01.12 ip1 GET a.gif HTTP/1.0 1:01.20 ip2 GET b.htm HTTP/1.0 1:01.25 ip1 GET c.jpg HTTP/1.0 1:01.27 ip1 GET d.txt HTTP/1.0 1:01.28 ip3 GET a.htm HTTP/1.0 1:01.35 ip4 GET d.gif HTTP/1.0 1:01.45 ip2 GET e.htm HTTP/1.0 : Trace driven OPEN SYSTEM MODEL service demands x x x server new arrivals arrival times file sizes from trace next arrival time from trace Open System

Carnegie Mellon University Computer Science Department 7 Distribution driven Use distributions of interarrival times and service demands (typically using trace info) x x x server new arrivals OPEN SYSTEM MODEL interarrival time dist. service demand dist. sample dist. sample dist. Open System

Carnegie Mellon University Computer Science Department 8 OPEN MODEL CLOSED MODEL Arrivals are independent of completions Arrivals are completely dependent on completions There is no max number of simultaneous users There is a fixed population of users, called the Multi-Programming-Level (MPL)

Carnegie Mellon University Computer Science Department 9

Carnegie Mellon University Computer Science Department 10 WEB WORKLOAD GENERATORS CLOSED MODEL CLOSED MODEL OPEN MODEL Surge SPECWeb TPC-W Sclient RUBiS WebBench Webjamma 1.Workload generators for the same purpose use different system models! 2.It’s often not clear which model workload generators use! Do you use an open or closed model?

Carnegie Mellon University Computer Science Department 11 NEITHER THE OPEN OR CLOSED MODEL IS COMPLETELY REALISTIC

Carnegie Mellon University Computer Science Department 12 x x x new arrivals server think send receive leave system with probability q return to the system PARTLY-OPEN MODEL PARTLY-OPEN SYSTEM

Carnegie Mellon University Computer Science Department 13 What is the impact of the choice of an open or closed model? OUR GOAL

Carnegie Mellon University Computer Science Department 14 HOW DO WE COMPARE OPEN AND CLOSED SYSTEMS? CLOSED CLOSED OPEN 1.Fix the service distribution across the systems 2.Fix the load across the systems load depends only on mean arrival rate and mean service demands load depends on MPL, think times, mean of service demands, variability of service demands … adjust load using the think time adjust load using the arrival rate

Carnegie Mellon University Computer Science Department 15 How do open and closed response times compare? FCFS scheduling open  Poisson arrival process closed  Exponential think times

Carnegie Mellon University Computer Science Department 16 load mean response time FCFS scheduling open  Poisson arrival process closed  Exponential think times Open Closed (MPL=10) CLOSED << OPEN

Carnegie Mellon University Computer Science Department 17 load mean response time Open Closed (MPL=10) Closed (MPL=100) Closed (MPL=1000) CLOSED  OPEN FCFS scheduling open  Poisson arrival process closed  Exponential think times

Carnegie Mellon University Computer Science Department 18 OPEN MODEL CLOSED MODEL VS CLOSED  OPEN AS MPL GROWS Schatte [36, 37] proves formally that as N grows to infinity, a closed FCFS queue converges to an open queue.

Carnegie Mellon University Computer Science Department 19 low variabilityhigh variability mean response time Open Closed (MPL=10) Closed (MPL=100) Closed (MPL=1000) Web Workloads How quickly does Closed  Open?

Carnegie Mellon University Computer Science Department 20 There principles 1.For a given load, mean response times are significantly lower in closed systems than in open systems. 2. As the MPL grows, closed systems become open, but convergence is slow for practical purposes. 3.While variability has a large effect in open systems, the effect is much smaller in closed systems.

Carnegie Mellon University Computer Science Department 21 What is the impact of the choice of an open or closed model? OUR GOAL It matters a lot! 1. What is the impact on the effectiveness of scheduling? 2.What is the impact in practice?

Carnegie Mellon University Computer Science Department 22 FCFS (First-Come-First-Served) PS (Processor-Sharing) PESJF (Preemptive-Expected-Shortest-Job-First) SRPT (Shortest-Remaining-Processing-Time-First) PELJF (Preemptive-Expected-Longest-Job-First)

Carnegie Mellon University Computer Science Department 23 SCHEDULING IS A KEY COMPONENT OF SYSTEM DESIGN Improved design Shortest Remaining Processing Time (SRPT) Standard design Processor Sharing (PS) WEB SERVERS Does the effectiveness of scheduling depend on the system model (open vs. closed)? Compare using a workload generator

Carnegie Mellon University Computer Science Department 24 SCHEDULING IN OPEN SYSTEMS OPEN mean response time load PLJF FCFS PS SRPT How do the closed results compare?

Carnegie Mellon University Computer Science Department 25 CONTRASTING THE IMPACT OF SCHEDULING OPEN CLOSED mean response time load load PLJF FCFS PS SRPT PLJF FCFS PS SRPT

Carnegie Mellon University Computer Science Department 26

Carnegie Mellon University Computer Science Department 27 Three priciples 1.While open systems benefit significantly from scheduling with respect to response time, closed systems improve much less. 2. Scheduling only significantly improves response time in closed systems under very specific parameter settings: moderate load (think times). 3. Scheduling can limit the effect of variability in both open and closed systems.

Carnegie Mellon University Computer Science Department 28 What is the impact of the choice of an open or closed model? OUR GOAL It matters a lot! Especially when evaluating scheduling policies What is the impact in practice?

Carnegie Mellon University Computer Science Department 29 OPEN VS CLOSED IN PRACTICE 3 CASE STUDIES 1.Serving static web content 2.Database backend of an e-commerce site 3. Auctioning web site testbed implementation trace-based simulation

Carnegie Mellon University Computer Science Department 30 Case study Open generator Closed generator Scheduling policies Static web (LAN) Sclient on World Cup trace Modified Sclient on World Cup trace PS, SRPT E-commerceModified TPC-WTPC-WPS, PESJF Auctioning Trace-based simulation (top 10 auction site trace) PS, SRPT

Carnegie Mellon University Computer Science Department 31 OPEN VS CLOSED IN PRACTICE OPENCLOSED mean response time PS SRPT PS SRPT load load STATIC WEB SERVER Different models give different conclusion about benefits of SRPT MPL=50

Carnegie Mellon University Computer Science Department 32 OPENCLOSED PS mean response time load PS SRPT PS SRPT load load PESJF E-COMMERCE SITE AUCTION SITE MPL=50

Carnegie Mellon University Computer Science Department 33 What is the impact of the choice of an open or closed model? OUR GOAL TODAY It matters a lot in practice! Especially when evaluating scheduling policies How can we identify whether to use an open or closed model?

Carnegie Mellon University Computer Science Department 34 A MORE REALISTIC ALTERNATIVE x x x new arrivals server think send receive leave system with probability q return to the system PARTLY-OPEN MODEL What parameters affect the load? Does think time affect the load? How do think times affect response times?

Carnegie Mellon University Computer Science Department 35 FITTING A PARTLY-OPEN MODEL 12 ip1 GET a.gif HTTP/ ip2 GET b.htm HTTP/ ip1 GET c.jpg HTTP/ ip1 GET d.txt HTTP/ ip3 GET a.htm HTTP/ ip4 GET d.gif HTTP/ ip2 GET e.htm HTTP/1.0 : Trace service demands file sizes from trace PARTLY-OPEN PARTLY-OPEN

Carnegie Mellon University Computer Science Department 36 FITTING A PARTLY-OPEN MODEL 12 ip1 GET a.gif HTTP/ ip2 GET b.htm HTTP/ ip1 GET c.jpg HTTP/ ip1 GET d.txt HTTP/ ip3 GET a.htm HTTP/ ip4 GET d.gif HTTP/ ip2 GET e.htm HTTP/1.0 : Trace PARTLY-OPEN PARTLY-OPEN Fitting the interarrival times Distinguish users e.g. use ip address in a web trace Identify user session boundaries  Use periods of inactivity of length > timeout

Carnegie Mellon University Computer Science Department 37 CHOOSING A TIMEOUT VALUE Number of sessions 2e5 1e min Timeout length financial world cup dept store

Carnegie Mellon University Computer Science Department 38 THE EFFECT OF THINK TIME STATIC WEB SERVER mean think time mean response time SRPT PS

Carnegie Mellon University Computer Science Department 39 CLOSED OPEN q1q1 q0q0 x x x new arrivals server think send receive leave system with probability q return to the system PARTLY-OPEN MODEL number of requests per visit ↑ number of requests per visit ↓ ?? A MORE REALISTIC ALTERNATIVE Workload generators are only Open/Closed!

Carnegie Mellon University Computer Science Department 40 THE TRANSITION FROM OPEN  CLOSED STATIC WEB SERVER PS open PS closed PS SRPT mean response time mean number of requests per visit OPEN CLOSED

Carnegie Mellon University Computer Science Department 41 THE PARTLY-OPEN SYSTEM IN PRACTICE mean number of requests per visit mean response time STATIC WEB PS SRPT E-COMMERCE SITE PS PESJF AUCTIONING PS SRPT

Carnegie Mellon University Computer Science Department 42 PS SRPT PS SRPT OPEN CLOSED PS SRPTPARTLY-OPENVS THESE DIFFERENCES ARE IMPORTANT IN PRACTICE

Carnegie Mellon University Computer Science Department 43 Two Principles 1.A partly-open system behaves similarly to an open system when the expected number of requests per session is small (≤ 5) and similarly to a closed system when the expected number of requests per session is large (≥ 10 as a rule-of-thumb). 2.In a partly-open system, think time has little effect on mean response time.

Carnegie Mellon University Computer Science Department 44 CHOOSING A SYSTEM MODEL Web workloads Open or closed? Use a partly-open model Large corporate web 2. CMU web server 3. Online department store 4. Science institute (USGS) 5. Online gaming site 6. Financial service provider 7. Supercomputing web site 8. Kasparov-DeepBlue match 9. Site seeing “slashdot effect” 10. Soccer world cup

Carnegie Mellon University Computer Science Department 45 CHOOSING A SYSTEM MODEL Web workloads Open or closed? Use a partly-open model......to decide which is more accurate 1. Large corporate web 2. CMU web server 3. Online department store 4. Science institute (USGS) 5. Online gaming site 6. Financial service provider 7. Supercomputing web site 8. Kasparov-DeepBlue match 9. Site seeing “slashdot effect” 10. Soccer world cup

Carnegie Mellon University Computer Science Department 46 HOW TO CHOOSE A SYSTEM MODEL Gather a trace How many simult. users are there? Fit a partly open model to the trace OPEN ≈ CLOSED >>1000 else What is the expected num. of visits? OPENCLOSED??? <55-10 >10 Mean num. of visits min Timeout length world cup dept store financial

Carnegie Mellon University Computer Science Department 47

Carnegie Mellon University Computer Science Department 48

Carnegie Mellon University Computer Science Department 49 CHOOSING A SYSTEM MODEL <5 expected visits >10 expected visits CLOSED 5-10 expected visits Web Workloads OPEN PARTLY OPEN 1. Large corporate web 2. CMU web server 3. Online department store 4. Science institute (USGS) 5. Online gaming site 6. Financial service provider 7. Supercomputing web site 8. Kasparov-DeepBlue match 9. Site seeing “slashdot effect” 10. Soccer world cup

Carnegie Mellon University Computer Science Department 50 CHOOSING A SYSTEM MODEL <5 expected visits 1. Large corporate web 2. CMU web server 4. Science institute (USGS) 6. Financial service provider 8. Kasparov-DeepBlue match 9. Site seeing “slashdot effect” >10 expected visits 5. Online gaming site 10. Soccer world cup CLOSED 5-10 expected visits 3. Online department store 7. Supercomputing web site Web Workloads OPEN PARTLY OPEN

Carnegie Mellon University Computer Science Department 51 CONCLUSION The differences in behavior of closed, open,and partly-open systems. These principles underscore the importance of choosing the appropriate system model. Our findings provide guidelines for choosing whether an open or closed model is the better approximation based on characteristics of the workload. Understanding the appropriate system model is essential to understanding the impact of scheduling.