550 likes | 683 Views
PhD Thesis Defense. QoS Support in Edge Routers. Idris A. Rai Institut Eurecom/Telecom Paris France 15 th September 2004. Overview. Motivation Proposed solution LAS scheduling Analysis of LAS scheduling for jobs LAS scheduling in packet networks Differentiated LAS-based policies
E N D
PhD Thesis Defense QoS Support in Edge Routers Idris A. Rai Institut Eurecom/Telecom Paris France 15th September 2004
Overview • Motivation • Proposed solution • LAS scheduling • Analysis of LAS scheduling for jobs • LAS scheduling in packet networks • Differentiated LAS-based policies • Summary
Motivation: Internet QoS • Original Internet: • Best-effort service • TCP: guarantees only packet delivery • Emerged applications • QoS architectures: • IntServ, DiffServ, MPLS, TE, etc • Scalability issues in core • Complex signaling protocols • Parameters hard to tune • Backward compatibility issue Current Internet still offersonly a best-effortservice with TCP
Motivation: Internet traffic • Emerged applications include • Web (short flows) and peer-to-peer (large flows) • Internet traffic measurements • Most of flows are short and a few largest ones (1%) contribute a large fraction of the total load • High variability property (HVP) of flow size distribution • We use coefficient of variation (C = /m) to indicate variability level of a distribution • Distributions that model HVP: Pareto, Bounded Pareto (BP), Weibull, etc
Fraction of the total mass Percentile Motivation: Internet traffic High variability property
Proposed solution • Given the HVP of Internet traffic • Give priority to the many short flows • Don’t penalize large flows too much • We propose using LAS scheduling • At the network edges routers • Keeping best-effort service at the network core
Analytical approaches • Mathematical analysis • Analyze LAS for CUP/Server scheduling • Job:is a workload that arrives in the system all at once • Simulation • Evaluate LAS for packet networks • Flow:is a sequence of packets arriving in bursts • Modeling • We model policies for packet networks
S(x)PS x S(x) Performance Metrics:X = flow/job size • Response time • Conditional mean: • Mean: • Slowdown:S(x) := T(x)/x • Conditional mean: • Mean: • Penalty metric Penalty region
What is LAS scheduling? LAS: Least attained service • Gives service to the job with the least attained service of all • Favors short jobs and reduces the mean T(x) • In queuing theory: (a.k.a. FB and SET) • Never proposed for packet networks before • Known to penalize largest jobs a lot when analyzed under M/M/1 queue [Kleinrock Vol. 2]
LAS scheduling:Response time • T(x)LAS :
SRPT requires the knowledge of job sizes • LAS schedulingdoes not require to know the size • of the jobs Related previous work • Shortest Remaining Processing Time (SRPT) First • A policy thatfavors short jobs • An optimal policy • Unfairness analysis of SRPT [Mor Balter et. al] • Implementation in Web servers
Outline • Analysis of LAS for jobs • Studying LAS under M/G/1 queue • Comparing LAS to other policies: • PS; penalty analysis, servers • FIFO; routers • SRPT; optimal • A study of LAS in packet networks • Proposing new differentiated LAS-based policies
Exponential BP PS Expected Slowdown S(x) Job Size Penalty analysis S(x) under LAS • Penalty level decreases with increasing variability of a distribution
SLAS/SPS Load Upper Bound of SLAS/SPS SLAS/ SPS< S(x) LAS<, x for < 1, and all distributions
Lemma: Proof: mn:= nth moment
Small C and -> 1 TLAS> TFIFO TLAS< TFIFO Load C Upper Bound of TLAS/TFIFO From Lemma TLAS= TFIFO TLAS /TFIFO
Exp BP: HVP T(x)LAS/T(x)SRPT Percentile of job size distribution Comparison of LAS and SRPT T(x)LAS T(x)SRPT for distributions with HVP at all load values E[T(x)] LAS /E[T(x)]SRPT
Outline • Analyzing of LAS for jobs • A study of LAS in packet networks • Impact of TCP and FIFO with droptail to short flows • Simulation results for LAS vs. FIFO • Analytical model of LAS in packet networks • What about long-lived flows under LAS? • Proposing new differentiated LAS-based policies
Client Server Initiate TCP connection Request object Firstwindow = S/C RTT RTT RTT Secondwindow = 2S/C Thirdwindow = 4S/C time at client time at server TCP during Slow Start (SS) • The transfer time during SS is dominated by RTT • Short flows are transferred during Slow Start phase of TCP
Impact of FIFO to RTT FIFO queue • Queuing delay can be high andprolongs the RTT • FIFO with droptail can lose packets from short flows p Arrivingpacket o RTT = 2 prop. delay + o + p
Wo/2 Wo Window Size (packets) Timeout RTO~3sec Time Impact losses to short TCP flows Packet losses at Slow Start • Packet losses at SS prolong transfer times of short flows Packet loss at SS Fast retransmit 1
Priorityqueue Highest Priority Pkt Lowest Priority Pkt LAS in packet networks First packet from a new flow coming to a full queue First packet from a new flow • LAS reduces queuing delay for the first packets of a flow • Reduces RTT (for short TCP flows) RTT 2 prop. delay + p LAS Reduces loss rate for short flows Avoids Timeouts and the use of RTO to recover losses Reduces transfer time for short flows
10Mb/s, 10ms Clients S3 Servers 3Mb/s, 30ms LAS in packet networks R1 R0 Bottleneck link • LAS is implemented in ns2 • Feldman et. al. [Sigcomm’99] • Flow sizes: Pareto
FIFO LAS FIFO LAS LAS in packet networks Transfer time Mean transfer time (sec) Percentile Flow size (packets) • LAS reduces the mean transfer time of short flows • It does not penalize the largest flows a lot
Packet loss rates Loss rates Flow size (packets) FIFO LAS LAS in packet networks • Flow of sizes < 40 pkts are unaffected • Loss rates of large flows under LAS remain moderate
LAS model in packet networks • Claim: Given that • the lead packets of flows arrive at a random point in time • no or low packet loss rate 1% then the model of LAS for packet networks model for jobs if: Packets service instances *approximates* jobs service instances
RTT 2 3 LAS model in packet networks A job of 3 service units 1 Arrival instances 2 3 Service instances 1 2 3 Time A TCP flow of 3 service units 1 1 Arrival instances Service instances 1 2 3 Time
Analysis Simulation Mean transfer time (sec) Flow size (packets) LAS model validation • Excellent agreement
C1 S1 C2 S2 10Mb/s,10ms Clients 3Mb/s, 30ms Servers C3 3Mb/s, 30 ms S3 R1 R0 C4 S4 10Mb/s, 10ms C5 S5 Performance of long-lived flows Sink Source R1 R0 Note: We consider extremely long flows
Long-lived FTP flows • Mean throughput (packets/sec) The performance of long-lived flows under LAS deteriorates under high load values
Outline • Analyzing LAS for jobs • A studying LAS in packet networks • Designing new differentiated LAS-based policies • Derive models • Validate the models
Differentiated LAS-based models • LAS-based models • Preserving nice properties of LAS • 2 types of flows in LAS: • Ordinary (r) and Priority (p) flows • Ordinary flows:Use the same priority change as under plain LAS • Priority flows:Use a priority function P(x)
LAS-linear Priority value LAS-log Packet number Differentiated LAS-based models • Models and Priority Functions: • Plain LAS : P(x) = x • LAS-fixed(k) : P(x) = k, k>1 • LAS-linear(k) : P(x) = x/k, k>1 • LAS-log(k) : P(x) = log2(x)1/kkR+ • Motivation of priority functions • Reducing the number of ordinary flows (resp. their packets) that a priority flow has to compete against
Differentiated LAS-based models • The expression ofT(x)for type i{r,p} flow is: • The truncated nth moment for flow sizex q = ratio of priority flows
2-class LAS based models • Moments in differentiated LAS based models
Analysis Simulation Mean transfer time (sec) Flow size (packets) Models validationLAS-linear(k = 5) Ordinary flows Priority flows Similar accuracy for other k values and for LAS-log(k) and LAS-linear(k)
LAS-log(k=1) LAS-log(k=0.5) LAS Throughput (packets/sec) Simulation time (sec) Performance of FTP long-lived flow Load = 0.92, Link speed = 120packets/sec
We analyzed the performance of LAS for different job size distributions Contributions • We studied the interaction of LAS and TCP • We proposed new scheduling policies • We derived analytical models of the policies in packet networks • We validated the policies using simulation
Couldn’t be covered… • LAS in heterogeneous networks • Heterogeneous propagation delays • Heterogeneous transport protocols • Multiple congested routers • LAS-FCFS scheduling • LAS-FCFS differentiated architectures
Outlook Discrepancy due to losses of packets from long flows • Modeling the impact of packet losses to analytical models LAS Inside LAS at loss rate = 3.6%
Publications • “Performance modeling of LAS based scheduling policies in packet switched networks “Proceedings of ACM Sigmetrics 2003 • “Analysis of LAS scheduling for job size distributions with high variance” Proceedings of ACM Sigmetrics 2003 • “LAS scheduling approach to avoid bandwidth hogging in heterogeneous TCP networks“7th IEEE International Conference on High Speed Networks and Multimedia communications HSNMC'04 • “Size-based scheduling with differentiated services to improve response time of highly varying flows” 5th ITC Specialist Seminar, Internet Traffic Engineering and Traffic Management Wurzburg, Germany, July 2002. • “Analyzing the performance of TCP flows in packet networks with LAS schedulers” Submitted for a journal publication 2004
Exponential job distribution • Theorems: For an exponential job size distribution SLAS≤ SPS TLAS = TPS
Exponential BP PS Expected Slowdown E[S(x)]) Percentile of job size distribution Penalized Jobs vs. C • C≥ 6, less than 1% of largest jobs see a small penalty
LAS at Overload ( ≥ 1) • For FIFO and PS, E[T(x)] = , for ≥ 1 x • Theorem: • E[T(x)] < for ≥ 1 if x < xLAS() • For SRPT [Bansal and Harchol-Balter, Sigmetrics 2001] • E[T(x)] < for ≥ 1 if x < xSRPT ()
FIFO LAS FIFO LAS LAS in packet networks Transfer time Mean transfer time (sec) Percentile Flow size (packets) • LAS reduces the mean transfer time of short flows • It does not penalize the largest flows a lot
Foreground traffic: LAS Background traffic:FIFO LAS-FCFS architecture Incoming traffic Outgoing traffic • Compared to LAS, LAS-FCFS helps the largest jobs for job size distribution with high variability property • Moderate size jobs under LAS-FCFS suffer
LAS-FCFS based differentiated architectures • Differentiated models • Classify traffic into high priority (HP) and low priority (LP) • Fixed priority LAS-FCFS • Extended LAS-FCFS • Differential LAS-FCFS
HP Incoming traffic Outgoing traffic Classifier LAS-FCFS LAS-FCFS LP Fixed priority LAS-FCFS • Can eliminate penalty for high priority jobs • However: Short low priority jobs receive a heavy penalty • But: Short jobs experience very low response time under LAS
Foreground traffic: LAS Incoming traffic Outgoing traffic HP Classifier LP Background traffic:FCFS Differential LAS-FCFS architecture • It guarantees low response time for all short flows • It can significantly improve the performance of high priority large jobs