RAQFM – a Resource Allocation Queueing Fairness Measure David Raz School of Computer Science, Tel Aviv University Jointly with Hanoch Levy, Tel Aviv University.

Slides:



Advertisements
Similar presentations
3/24//2004Fairness in Queues, H. Levy, CS, TAU1 How Fair is your Queue March, 2004 Hanoch Levy School of Computer Science, Tel Aviv University Jointly.
Advertisements

EISCambridge 2009MoN8 Exploring Markov models for gate-limited service and their application to network-based services Glenford Mapp and Dhawal Thakker.
. Markov Chains. 2 Dependencies along the genome In previous classes we assumed every letter in a sequence is sampled randomly from some distribution.
1 Size-Based Scheduling Policies with Inaccurate Scheduling Information Dong Lu *, Huanyuan Sheng +, Peter A. Dinda * * Prescience Lab, Dept. of Computer.
Short-Term Fairness and Long- Term QoS Lei Ying ECE dept, Iowa State University, Joint work with Bo Tan, UIUC and R. Srikant, UIUC.
Discrete Time Markov Chains
Load Balancing of Elastic Traffic in Heterogeneous Wireless Networks Abdulfetah Khalid, Samuli Aalto and Pasi Lassila
TCOM 501: Networking Theory & Fundamentals
Maryam Elahi Fairness in Speed Scaling Design Joint work with: Carey Williamson and Philipp Woelfel.
Queuing Theory For Dummies
Computer Science Deadline Fair Scheduling: Bridging the Theory and Practice of Proportionate-Fair Scheduling in Multiprocessor Servers Abhishek Chandra.
Algorithm Orals Algorithm Qualifying Examination Orals Achieving 100% Throughput in IQ/CIOQ Switches using Maximum Size and Maximal Matching Algorithms.
#11 QUEUEING THEORY Systems Fall 2000 Instructor: Peter M. Hahn
ECS 152A Acknowledgement: slides from S. Kalyanaraman & B.Sikdar
Performance analysis for high speed switches Lecture 6.
TCOM 501: Networking Theory & Fundamentals
Mean Delay in M/G/1 Queues with Head-of-Line Priority Service and Embedded Markov Chains Wade Trappe.
Fundamental Characteristics of Queues with Fluctuating Load VARUN GUPTA Joint with: Mor Harchol-Balter Carnegie Mellon Univ. Alan Scheller-Wolf Carnegie.
Carnegie Mellon University Computer Science Department 1 CLASSIFYING SCHEDULING POLICIES WITH RESPECT TO HIGHER MOMENTS OF CONDITIONAL RESPONSE TIME Adam.
A Multi-Agent Learning Approach to Online Distributed Resource Allocation Chongjie Zhang Victor Lesser Prashant Shenoy Computer Science Department University.
Queuing Networks: Burke’s Theorem, Kleinrock’s Approximation, and Jackson’s Theorem Wade Trappe.
1 TCOM 501: Networking Theory & Fundamentals Lecture 8 March 19, 2003 Prof. Yannis A. Korilis.
Fundamental Characteristics of Queues with Fluctuating Load (appeared in SIGMETRICS 2006) VARUN GUPTA Joint with: Mor Harchol-Balter Carnegie Mellon Univ.
Load Balancing and Switch Scheduling Xiangheng Liu EE 384Y Final Presentation June 4, 2003.
Computer Networks Performance Evaluation. Chapter 12 Single Class MVA Performance by Design: Computer Capacity Planning by Example Daniel A. Menascé,
yahoo.com SUT-System Level Performance Models yahoo.com SUT-System Level Performance Models8-1 chapter12 Single Class MVA.
Introduction to Queuing Theory
Flows and Networks Plan for today (lecture 5): Last time / Questions? Waiting time simple queue Little Sojourn time tandem network Jackson network: mean.
Efficient Scheduling of Heterogeneous Continuous Queries Mohamed A. Sharaf Panos K. Chrysanthis Alexandros Labrinidis Kirk Pruhs Advanced Data Management.
Decentralised load balancing in closed and open systems A. J. Ganesh University of Bristol Joint work with S. Lilienthal, D. Manjunath, A. Proutiere and.
Verification & Validation
1 Mor Harchol-Balter Carnegie Mellon University Computer Science Heavy Tails: Performance Models & Scheduling Disciplines.
Probability Review Thinh Nguyen. Probability Theory Review Sample space Bayes’ Rule Independence Expectation Distributions.
Yossi Azar Tel Aviv University Joint work with Ilan Cohen Serving in the Dark 1.
1 Networks of queues Networks of queues reversibility, output theorem, tandem networks, partial balance, product-form distribution, blocking, insensitivity,
CSE 538 Instructor: Roch Guerin – Office hours: Mon & Wed 4:00-5:00pm in Bryan 509 TA: Junjie Liu
Introduction to Queueing Theory
Dominant Resource Fairness: Fair Allocation of Multiple Resource Types Ali Ghodsi, Matei Zaharia, Benjamin Hindman, Andy Konwinski, Scott Shenker, Ion.
Network Design and Analysis-----Wang Wenjie Queueing System IV: 1 © Graduate University, Chinese academy of Sciences. Network Design and Analysis Wang.
Networks of Queues Plan for today (lecture 6): Last time / Questions? Product form preserving blocking Interpretation traffic equations Kelly / Whittle.
Queuing Theory Basic properties, Markovian models, Networks of queues, General service time distributions, Finite source models, Multiserver queues Chapter.
Flows and Networks Plan for today (lecture 4): Last time / Questions? Output simple queue Tandem network Jackson network: definition Jackson network: equilibrium.
Modeling and Analysis of Computer Networks
Queuing Networks Jean-Yves Le Boudec 1. Contents 1.The Class of Multi-Class Product Form Networks 2.The Elements of a Product-Form Network 3.The Product-Form.
1 Our focus  scheduling a single CPU among all the processes in the system  Key Criteria: Maximize CPU utilization Maximize throughput Minimize waiting.
1 Markov Decision Processes Infinite Horizon Problems Alan Fern * * Based in part on slides by Craig Boutilier and Daniel Weld.
EE 685 presentation Optimization Flow Control, I: Basic Algorithm and Convergence By Steven Low and David Lapsley.
Packet Scheduling: SCFQ, STFQ, WF2Q Yongho Seok Contents Review: GPS, PGPS SCFQ( Self-clocked fair queuing ) STFQ( Start time fair queuing ) WF2Q( Worst-case.
Computer Networking Queueing (A Summary from Appendix A) Dr Sandra I. Woolley.
Analysis of SRPT Scheduling: Investigating Unfairness Nikhil Bansal (Joint work with Mor Harchol-Balter)
Queuing Theory.  Queuing Theory deals with systems of the following type:  Typically we are interested in how much queuing occurs or in the delays at.
CprE 458/558: Real-Time Systems (G. Manimaran)1 CprE 458/558: Real-Time Systems Energy-aware QoS packet scheduling.
Flows and Networks Plan for today (lecture 3): Last time / Questions? Output simple queue Tandem network Jackson network: definition Jackson network: equilibrium.
Random Variables r Random variables define a real valued function over a sample space. r The value of a random variable is determined by the outcome of.
QUEUING. CONTINUOUS TIME MARKOV CHAINS {X(t), t >= 0} is a continuous time process with > sojourn times S 0, S 1, S 2,... > embedded process X n = X(S.
Mohammad Khalily Islamic Azad University.  Usually buffer size is finite  Interarrival time and service times are independent  State of the system.
CPU Scheduling CS Introduction to Operating Systems.
Process Scheduling. Scheduling Strategies Scheduling strategies can broadly fall into two categories  Co-operative scheduling is where the currently.
CPU Scheduling Andy Wang Operating Systems COP 4610 / CGS 5765.
Discrete Time Markov Chains (A Brief Overview)
Flows and Networks Plan for today (lecture 4):
Scheduling Non-Preemptive Policies
Top Percentile Pricing and the Economics of Multi-Homing
Lecture on Markov Chain
Job Fairness in Queue Scheduling:
Autoscaling Effects in Speed Scaling Systems
How Fair is your Queue Hanoch Levy
Mean Delay Analysis of Multi Level Processor Sharing Disciplines
Presentation transcript:

RAQFM – a Resource Allocation Queueing Fairness Measure David Raz School of Computer Science, Tel Aviv University Jointly with Hanoch Levy, Tel Aviv University Benjamin Avi-Itzhak, RUTGERS University Sigmetrics – Performance, June 2004

June, 2004D. Raz, RAQFM, Sigmetrics04 2 Outline Motivation – The importance of fairness in queues – The physical properties of the problem – Related work The RAQFM approach Properties Analysis

June, 2004D. Raz, RAQFM, Sigmetrics04 3 Why is Fairness in Queues Important? Fundamental reason: – Isn’t fairness why we have queues in the first place? Scientific evidence, recent studies, Rafaeli et. al. [2003] (experimental psychology): – Experiments on humans in multi-queue and single queue – Fairness in queue is very important to people – Perhaps even more than the delay itself

June, 2004D. Raz, RAQFM, Sigmetrics04 4 Queueing Theory and Fairness Decades of research on: – Queueing structures / policies – Focused on efficiency – delay distribution/moments, utilization, etc. The issue of fairness is discussed, but not quantified – Larson (1988), Palm (1953), Mann (1969), Whitt (1984) etc. Existing measures for streams (WFQ) Little analysis on job fairness – Morris & Wang (1985) – Avi-Itzhak & Levy (2004) – Wierman & Harchol-Balter (Sigmetrics 2003)  We know very little about job fairness

June, 2004D. Raz, RAQFM, Sigmetrics04 5 Queues: The Physical Factors Size (Service Requirement) Resources Seniority (Arrival Time)

June, 2004D. Raz, RAQFM, Sigmetrics04 6 Fairness: Size and Seniority Size Seniority

June, 2004D. Raz, RAQFM, Sigmetrics04 7 The Size vs. Seniority Dilemma Short vs. Long Is it more fair to serve Short ahead of Long? What is the intuition? Long Short

June, 2004D. Raz, RAQFM, Sigmetrics04 8 Recent Related Work(1) Avi-Itzhak & Levy (2004): – Axiomatic approach to fairness – Departure point & emphasis: Seniority (Order of service) Some results (for equal service times) – FCFS is the most fair (among non preemptive) – LCFS is the most unfair (among non preemptive) Takes Seniority into account

June, 2004D. Raz, RAQFM, Sigmetrics04 9 Recent Related Work(2) Wierman & Harchol-Balter (2003): – Propose a Fairness Criterion – Slowdown: for job of size x compute E[T(x)/x]. If it is bounded the system is FAIR. Some results: – FCFS is “Always UNFAIR” – LCFS (preemptive) is FAIR Takes Size into account

June, 2004D. Raz, RAQFM, Sigmetrics04 10 Requirements From a Job Fairness Measure Reacts well to both Seniority and Size Aim for standard  have a consistent view and agree with Intuition  build confidence Yields to Analysis RAQFM – A Resource Allocation Queueing Fairness Measure

June, 2004D. Raz, RAQFM, Sigmetrics04 11 RAQFM Philosophy Equal Share of Resources  Fairness

June, 2004D. Raz, RAQFM, Sigmetrics04 12 RAQFM - How to Apply the Philosophy: Individual Discrimination At every epoch t with N(t) customers in the system, each customer should get 1/N(t) Warranted service: Granted service: Compare the warranted service with the granted service : discrimination:

June, 2004D. Raz, RAQFM, Sigmetrics04 13 Basic Properties of the Discrimination E[D]=0 (every sample path, every policy) Proof sketch: The momentary rate of discrimination is  ”zero sum” For PS individual discrimination is zero

June, 2004D. Raz, RAQFM, Sigmetrics04 14 RAQFM – System Fairness System Unfairness: differences in customer treatment (customer discrimination)  We measure the unfairness in a system using Var[D] ( D is r.v.) Var[D] ≥ 0  Property 1: PS is the most fair policy

June, 2004D. Raz, RAQFM, Sigmetrics04 15 Short vs. Long Revisited Seniority (Arrival Time) Difference Size (Service Requirement) Difference Is it more fair to serve Short ahead of Long?

June, 2004D. Raz, RAQFM, Sigmetrics04 16 The difference in unfairness resulting from serving short ahead of long is If size difference is small – serve by order of arrival If seniority difference is small – serve by order of service requirement RAQFM Agrees With Intuition Short vs. Long Revisited

June, 2004D. Raz, RAQFM, Sigmetrics04 17 Property 2: Bounds Individual discrimination is bound by – How good – How bad

June, 2004D. Raz, RAQFM, Sigmetrics04 18 Property 3: Reaction to Seniority Theorem: If customers have equal service requirements – For each pair of customers, it is more fair to serve the senior first ⇒FCFS is the most fair ⇒LCFS is the least fair (Proof sketch: compare scenarios) RAQFM Reacts Well to Seniority

June, 2004D. Raz, RAQFM, Sigmetrics04 19 Property 3: Reaction to Size Theorem: If customers arrive together – For each pair of customers, it is more fair to serve the shorter first ⇒SJF is the most fair ⇒LJF is the least fair (Proof sketch – prove ) RAQFM Reacts Well to Size

June, 2004D. Raz, RAQFM, Sigmetrics04 20 Property 4: RAQFM is Analyzable Example: analysis of M/M/1 FCFS – Markov chain state is ( a,b )=(# ahead, # behind) – c(a,b) =momentary discrimination rate at state ( a,b ) – d(a,b) =expected discrimination of a walk starting at ( a,b ) – – Similarly for the second moment d (2) (a,b) –

June, 2004D. Raz, RAQFM, Sigmetrics04 21 Fairness of Service Policies Not Discriminating For Size as function of load Empty system: everyon e is alone LCFS: Severe seniority violation FCFS: no seniority violation PS: Absolut e Fairness

June, 2004D. Raz, RAQFM, Sigmetrics04 22 Is FCFS Always More Fair Than LCFS? FCFS = 0.9 Preemptive LCFS = 0.15

June, 2004D. Raz, RAQFM, Sigmetrics04 23 Summary Fairness is extremely important, yet there are little quantification methods for jobs RAQFM – Philosophy: fair share of service – Agrees with intuition – Reacts well to seniority and size – Yields to analysis

June, 2004D. Raz, RAQFM, Sigmetrics04 24 Future Work More service policies – SRPT, SJF, FB etc. Generalized service requirement Different architectures – multi-class, multi- server, multi-queue, polling etc.

RAQFM – a Resource Allocation Queueing Fairness Measure David Raz School of Computer Science, Tel Aviv University Jointly with Hanoch Levy, Tel Aviv University Benjamin Avi-Itzhak, RUTGERS University Sigmetrics – Performance, June 2004 Thank You

June, 2004D. Raz, RAQFM, Sigmetrics04 26 Property 4: RAQFM is Analyzable RAQFM yields to analysis via standard queueing theory techniques We compute – E[D|k] - discrimination given number seen on arrival – Var[D] - system unfairness Conducted analysis for M/M/1 type: Variety of service orders (FCFS, LCFS, ROS, more…)

June, 2004D. Raz, RAQFM, Sigmetrics04 27 Sensitivity to Size and Seniority: Mr. Short vs. Mrs. Long Revisited Observe a simple scenario: – C 0 in the system, being served – C 1 and C 2 arrive – Compare the fairness depending on the order of service. The difference in fairness is:

June, 2004D. Raz, RAQFM, Sigmetrics04 28 Analysis of RAQFM RAQFM yields to analysis via standard queueing theory techniques We compute for M/M/1 – E[D|k] – Discrimination given number seen on arrival – Var[D] - system unfairness

June, 2004D. Raz, RAQFM, Sigmetrics04 29 FCFS Analysis Time can be slotted by arrivals and departures, slot size is independent State description: a customers ahead of C, b customers behind C. If we assume C is in state (a,b) in slot i – customer arrives – probability λ̃, state (a,b+1) – customer departs - probability μ̃, state (a-1,b)

June, 2004D. Raz, RAQFM, Sigmetrics04 30 FCFS Analysis d(a,b) is the discrimination of a walk starting in (a,b), with moment d(a,b) c(a,b) is the momentary discrimination at (a,b) t the slot size

June, 2004D. Raz, RAQFM, Sigmetrics04 31 FCFS Analysis Solving the recursive equation: – Choose a maximum value of b, say N – Run through the matrix iteratively, column by column – O(N 2 ) time, O(N) space

June, 2004D. Raz, RAQFM, Sigmetrics04 32 Individual discrimination for FCFS k – number seen on arrival Bingo – clear road at rush hour Traffic Jam at 4AM

June, 2004D. Raz, RAQFM, Sigmetrics04 33 Short vs. Long Revisited C1C1 C2C2 C0C0 Seniority (Arrival Time) Difference Size (Service Requirement) Difference

June, 2004D. Raz, RAQFM, Sigmetrics04 34 Sensitivity to Size and Seniority: Mr. Short vs. Mrs. Long Revisited Size (Service Requirement) Difference

June, 2004D. Raz, RAQFM, Sigmetrics04 35 Sensitivity to Size and Seniority: Mr. Short vs. Mrs. Long Revisited C1C1 C2C2 C1C1 C2C2 C2C2 C1C1 A B

June, 2004D. Raz, RAQFM, Sigmetrics04 36 RAQFM Agrees With Intuition II Bounds Individual discrimination is bound by – How good – How bad (Conveniently, System unfairness is bound by Bounds provide a scale for comparison RAQFM is Sensitive to Size and Seniority

June, 2004D. Raz, RAQFM, Sigmetrics04 37 Property 1: PS is the most fair Theorem: Processor Sharing (PS) is the most fair policy PS - each customer gets 1/N(t) ⇒ PS is the most fair policy