QoS II - Adaptive Virtual Queue - Fair Queueing for Multiple Link

QoS II- Adaptive Virtual Queue - Fair Queueing for Multiple Link

12th Mar., 2002

Eun-Chan Park

CSL, SoEECS, SNU

S. Kunniyur, R.Srikant

“Analysis and Design of an Adaptive Virtual

Queue (AVQ) Algorithm

for Active Queue Management,”

SIGCOMM 2001

Contents Background

Congestion control Active Queue Management

Related works RED, PI, REM

Adaptive Virtual Queue (AVQ) AVQ algorithm Stability analysis Simulation results

Conclusion

Congestion Control Congestion control schemes

End-to-end control

- TCP-Tahoe, TCP-Reno, TCP-Vegas Router supported control

- AQM: RED, REM, PI control, AVQ

TCP congestion control algorithm Window-based transmission control

sender limits its transmission rate by controlling window size

slow-start, congestion avoidance, fast-recovery Feedback

Implicit: Timeout, Duplicate ACKs Explicit Congestion Notification (ECN)

Active Queue Management Drop-tail queue has several problems

Reduce rates only after overflow loss Results in significant packet loss Packet drop could result in a global sync.

Active Queue Management Resolves the problem of drop-tail Drops or marks packets at the buffer of router

Random Early Detection (RED) S. Floyd and V. Jacobson, “random early detection gateways for

congestion avoidance,” IEEE/ACM trans. On networking, vol. 1, pp. 397-413, 1993.

Detect congestion using average queue size Intelligently drop/mark packet before buffer overflow

RED (cont.) Advantages

Prevent global synchronization Reduce packet loss Minimize biases against bursty traffic Simple, low-overhead

Disadvantages Difficulty of appropriate parameter setting Sensitive queueing delay and throughput to the traffic load

and to parameters Argument in the case of small buffer size

Random Exponential Marking (REM) S. Athuraliya et al., “REM: Active Queue

Management,” IEEE Network Magazine May/June. 2001

“Match Rate, Clear Buffer” Match aggregated input rate to network capacity Stabilize queue around a small target

Sum prices Prices: Differentiated from the calculation of dropping or

marking prob. End-to-End marking prob. exponentially increases with the

sum of prices

REM (Cont.)

Link price

Updated periodically Depends on mismatches of rate and queue length

Marking Prob.

])}[][()][({][]1[ * nCnxbnbnpnp lllllll

ll np ][

1

Proportional-Integral (PI) Controller C.Hollot et al., “On designing improved controllers for AQM

routers supporting TCP flows,” IEEE INFOCOM, 2001 Uses instantaneous queue length, while RED uses EWMA. Proportional to queue length mismatch and to its accumulation

(time integral)

Equivalent to the price of REM

is

refref

iebTkebakp

qkqbqkqakpkp

][][)(][

)][()]1[(][]1[

AVQ Algorithm C : link capacity, : virtual capacity VQ : virtual queue VQ=B (buffer size) On a packet arrival, it enqueues VQ, if no rooms

available in VQ, marks it updates on each packet arrival

If input rate is below than desired rate,

VC increases and less aggressive marking Otherwise, VC decreases and more aggressive marking

C~

CC ~

C~

)(~ CC

AVQ Algorithm (cont.) At a packet arrival Update VQ size: If VQ+b > B, then mark packet Else, VQVQ+b Update VC

)0,~

max( TCVQVQ

)0,),~

max(min(~

bCTCCC

Properties of AVQ Rate-based marking

Provides early feedback Achieves input rate to the desired utilization

Regulates utilization instead of queue length as RED,PI,REM

Robust to short flows Two design parameters (alpha, gamma)

determine robustness and stability

TCP/AVQ model for analysis

Similar to stochastic fluid-based TCP dynamics [14] Ignores slow-start and time-out of TCP Characterize AIMD

Linearize TCP/AQM model

))(~

),(()()(1

2dtCdtxpdtxtx

dx

jjiii

Stability Analysis of AVQ (1/3) Main ideas

Take Laplace Transform to the linearized TCP/AVQ model

Obtain the characteristic equation Find the condition where all roots of char. Eq. is on

the LPF.

Characteristic Eq.

Stability Analysis (2/3)

What value of K yields? Can it guarantee the unique d?

Make the condition less strict

Find necessary condition

Stability Analysis (3/3)

Simulation (1/3) Compare performance (loss, utilization, avg. queue

length) of various AQM schemes when traffic load varies

Simulation (2/3) Compare responsiveness of AQM schemes when

flows are dropped and then established again

t= 0 s, N=140

t=100 s, N=35

t=150 s, N=140

Simulation (3/3) Investigate the effect of short-flow

Conclusion AVQ algorithm is proposed

Maintains small queue length with consistent utilization and small loss

Robust to short-flows Stability is analyzed relating to control parameters

and feedback delay A design guideline is provided However,

Simulation result showing the validity of analysis is missed Simulation results are unfair (number of packet drop) Queueing delay is not effectively regulated

J. M. Blanquer, B. Ozden

“Fair Queueing for Aggregated Multiple Links,”

SIGCOMM 2001

Contents Introduction & Background

GPS, PGPS (WFQ) MSFQ

Preliminary properties of MSFQ Bound on packet delay of MSFQ Bound on per-flow service of MSFQ

Fairness and MSF2Q Applications Conclusion

Introduction Fair queueing/scheduling is required due to

Increased variety of traffic diverse requirement for QoS limited network resources

Fair queueing disciplines based on GPS have been studied considerably in case of single server, however, not in case of multiple server system

Generalized Processor Sharing (GPS) L. Kleinrock “Queueing Systems Vol 2: Computer Applications,”

Wiley, 1976 GPS server serving N flows is characterized by

where, is the amount of traffic for flow i served in With Leaky Bucket algorithm, it guarantees bandwidth share

Also provides an end-to-end bounded delay service

Nii ,2,1,

j

i

j

i

tW

tW

),(

),(

),( tWi ],[ t

rr

jj

ii

GPS (Cont.) Idealized discipline that it can not be implemented

A server can transmit only one packet at a time, not several packets simultaneously

Traffic can not be divided infinitely As a solution to implementation, several realizable

schemes proposed Packet-by-packet GPS (PGPS), Weighted Fair Queueing

(WFQ) Virtual clock, Self-clocked fair queueing, Start-time fair

queueing

Packet-by-packet GPS (PGPS)

K. Parekh, “A Generalized Processor Sharing Approach to Flow Control in IntServ Network,” IEEE Tran. On Networking, 1993

Also known as Weighted Fair Queueing (WFQ) A. Demers, et al. “Design and Analysis of a Fair Queueing Algorithm,”

SIGCOMM 1989 Provides guarantees on throughput and worst-case packet delay

Packet delay compared to that of GPS is not grater than the transmission time of one maximum size packet

Bits served for each flow do not fall behind corresponding GPS

by more than one maximum size packet

r

Ldd GPSpPGPSp

max,,

max,, ),0(),0( LtWtW PGPSiGPSi

GPS & PGPS

Packet Arrivals of flow 1 and flow 2 < Comparison of GPS & PGPS >

MSFQMulti-server version of WFQ multi-server version of GPS

(MSFQ,N,r) (GPS,1,Nr) Compare how well (MSFQ,N,r) approximates

(GPS,1,Nr) in terms of worst-case packet delay amount of traffic served for each flow

Preliminary properties of MSFQ: Total service Let the total # of bits serviced in by GPS, MSFQ be

and , respectively, then

Left ineq. implies: When GPS is busy, MSFQ is busy, too. However, the converse is not true.

Right ineq. implies the need for a buffer size of

),0( GPSW),0( MSFQW

],0[

max)1(),0(),0(0 LNWW MSFQGPS

max)1( LN

Preliminary properties of MSFQ:Waiting time of packet

Upper bound of waiting time for packet k to be scheduled

Pf: Consider the worst case:

i) The previous packets have occupied all N server

just before the arrival of packet k,

ii) all servers finish at the same time

NrLabti

ikkkW )(,

Bound on packet delay of MSFQ

(1) Consider two extreme cases

For GPS, best case: (2)with assumption

For MSFQ, worst case: (3) (3)-(2) yields (1)

Compare it with the delay in single server

r

L

Nr

LNdd k

GPSkMSFQkmax

,,

)1(

Nr

Lad k

GPSkGPSk ,,

2,12 flowk

r

L

r

Lad p

MSFQkMSFQk max,,

r

Ldd GPSkPGPSk

max,,

Bound on per-flow service of MSFQ

Maximum difference occurs when flow i becomes idle in GPS a packet of flow i begins transmission in MSFQ

Proof done case by case (follow yourself ^^) For total service: For a single server:

max,, ),0(),0( NLWW MSFQiGPSi

kMSFQkMSFQiGPSkGPSi LLNbWdW max,,,, )1(),0(),0(

max)1(),0(),0(0 LNWW MSFQGPS

max,, ),0(),0( LtWtW PGPSiGPSi

Fairness of MSFQ The eq. is incompl

ete to guarantee fairness Why?

The eq. does not ensure that the amount of per-flow service does not exceed arbitrary the amount under GPS

i.e., there is no lower limit in the eq. To resolve this problem

Introduce MSF2 Q, which is an extended version of WF2 Q for multi-server system

max,, ),0(),0( NLWW MSFQiGPSi

MSF2 Q (1/3) Queued packets at t=0:

ten packets of flow 1 one packet of each flow 2~N

GPS Scheduling MSF2 Q Scheduling

Ni

ii

,2,05.0

,1,5.0

MSF2 Q (2/3) Scheduling discipline

Define # of outstanding packet of flow i at time t

where, outstanding packet is a packet being transmitted of picked for transmission

At time t, when a server is idle and there is a packet to serve, MSF2Q schedules among flows satisfying:

)(2,to

QMSFi

r

trtotWtW i

QMSFiGPSiQMSFi

)()(and),0(),0( 22 ,,,

MSF2 Q (3/3) Properties of MSF2 Q

MSF2 Q provides the lower bound of difference of per-flow service

Similar to WF2Q

Note that MSF2 Q is not work-conserving Future work: investigate implement of work-conser

ving scheduler

max,,max, ),0(),0( NLWWNL MSFQiGPSii

max,,max, ),0(),0(1 2 LWWLr

rQWFiGPSii

i

Applications Ethernet link aggregation

Cost-effective and fault tolerant solution for scaling the network capacity

IEEE 802.3ad: Standard for Ethernet link aggregation

Access of storage I/O RAID system with multiple SCIS channels to

improve I/O performance MSF2Q is expected to provide QoS guarantee and

fair sharing of multiple I/O channels

Conclusion Service guarantee and Fairness and for aggregated lin

ks have been studied Extended version of PGPS for multiple server has been analy

zed in terms of packet delay and per-flow service Proposed a new fair queueing in multiple servers, MSF2Q

Future works Implementation issues Quantitative comparison to the approach of partitioning flows Extension of hierarchal GPS and servers with different rates