L Subramanian*, I Stoica*, H Balakrishnan +, R Katz* *UC Berkeley, MIT + USENIX NSDI’04, 2004...

25
L Subramanian*, I Stoica*, H Balakrishnan + , R Katz* *UC Berkeley, MIT + USENIX NSDI’04, 2004 Presented by Alok Rakkhit, Ionut Trestian

description

 Today’s Internet still continues to provide only a best-effort service. The main reason is the requirement of these proposals that all network elements implement QoS mechanisms.  The authors propose OverQoS, an overlay based QoS architecture for enhancing Internet QoS. 3

Transcript of L Subramanian*, I Stoica*, H Balakrishnan +, R Katz* *UC Berkeley, MIT + USENIX NSDI’04, 2004...

Page 1: L Subramanian*, I Stoica*, H Balakrishnan +, R Katz* *UC Berkeley, MIT + USENIX NSDI’04, 2004 Presented by Alok Rakkhit, Ionut Trestian.

L Subramanian*, I Stoica*, H Balakrishnan+, R Katz**UC Berkeley, MIT+

USENIX NSDI’04, 2004Presented by Alok Rakkhit, Ionut Trestian

Page 2: L Subramanian*, I Stoica*, H Balakrishnan +, R Katz* *UC Berkeley, MIT + USENIX NSDI’04, 2004 Presented by Alok Rakkhit, Ionut Trestian.

IntroductionOverQos ArchitectureControlled-Loss Virtual Link (CLVL)OverQoS ImplementationTwo Sample ApplicationEvaluationConclusions

2

Page 3: L Subramanian*, I Stoica*, H Balakrishnan +, R Katz* *UC Berkeley, MIT + USENIX NSDI’04, 2004 Presented by Alok Rakkhit, Ionut Trestian.

Today’s Internet still continues to provide only a best-effort service. The main reason is the requirement of these proposals that all network elements implement QoS mechanisms.

The authors propose OverQoS, an overlay based QoS architecture for enhancing Internet QoS.

3

Page 4: L Subramanian*, I Stoica*, H Balakrishnan +, R Katz* *UC Berkeley, MIT + USENIX NSDI’04, 2004 Presented by Alok Rakkhit, Ionut Trestian.

Enhancements: Smoothing losses

▪ Reduce or even eliminate the loss bursts by smoothing packet losses across time

Packet prioritization▪ Protect important packets

Statistical Bandwidth and Loss Guarantees

4

Page 5: L Subramanian*, I Stoica*, H Balakrishnan +, R Katz* *UC Berkeley, MIT + USENIX NSDI’04, 2004 Presented by Alok Rakkhit, Ionut Trestian.

Assumptions The placement of overlay nodes is pre-specified The end-to-end path on top of an overlay network

is fixed Using existing approaches like RON to

determine the overlay path. Terms

Virtual link – The IP path between two overlay nodes

Bundle – A stream of application data packets carried across the virtual link

5

Page 6: L Subramanian*, I Stoica*, H Balakrishnan +, R Katz* *UC Berkeley, MIT + USENIX NSDI’04, 2004 Presented by Alok Rakkhit, Ionut Trestian.

Overlay-based QoS challenges Node Placement and Cross Traffic Fairness

▪ Should not hurt the cross traffic Stability

▪ Many virtual links overlapping on congested physical links should be able to co-exist

6

Page 7: L Subramanian*, I Stoica*, H Balakrishnan +, R Katz* *UC Berkeley, MIT + USENIX NSDI’04, 2004 Presented by Alok Rakkhit, Ionut Trestian.

A Solution builds on two principles Bundle loss control

▪ Using controlled-loss virtual link (CLVL) to bound the loss rate

Resource management within a bundle▪ Control the loss and bandwidth allocations

7

Page 8: L Subramanian*, I Stoica*, H Balakrishnan +, R Katz* *UC Berkeley, MIT + USENIX NSDI’04, 2004 Presented by Alok Rakkhit, Ionut Trestian.

The CLVL provides a loss rate bound, q. Using a combination of FEC and ARQ The bandwidth overhead should be minimized

The total traffic consists of: The traffic of the bundle The redundancy traffic

The available bandwidth for the flows in the bundle

8

b(t): Traffic bound at time tr(t): Fraction of redundancy traffic

Page 9: L Subramanian*, I Stoica*, H Balakrishnan +, R Katz* *UC Berkeley, MIT + USENIX NSDI’04, 2004 Presented by Alok Rakkhit, Ionut Trestian.

If the traffic arrival rate is larger than available bandwidth c, the extra traffic is dropped at the entry overlay node With priority

Statistical bandwidth guarantees , where u represents the

probability of not meeting the bandwidth guarantee

As long as the total allocated bandwidth is less than cmin

9

Page 10: L Subramanian*, I Stoica*, H Balakrishnan +, R Katz* *UC Berkeley, MIT + USENIX NSDI’04, 2004 Presented by Alok Rakkhit, Ionut Trestian.

Application-OverQoS Interface

10

It needs to tunnel its packets through the overlay network using an OverQoS proxy

The proxy is responsible for signaling the application specific requirements to OverQoS

OverQoS proxy is application specific

Page 11: L Subramanian*, I Stoica*, H Balakrishnan +, R Katz* *UC Berkeley, MIT + USENIX NSDI’04, 2004 Presented by Alok Rakkhit, Ionut Trestian.

End-to-end Recovery vs. Overlay CLVL Using FEC to apply end-to-end loss control is

far more expensive than on an aggregate level With a better distribution of overlay nodes,

they expect the overlay links to have much smaller RTTs than end-to-end RTTs▪ ARQ recovery is better in overlay-level

Delay guarantees Overlay has no control in queuing delays

Over-provisioning Overlay are the right platform for translating

intra domain QoS to end-to-end QoS guarantees

11

Page 12: L Subramanian*, I Stoica*, H Balakrishnan +, R Katz* *UC Berkeley, MIT + USENIX NSDI’04, 2004 Presented by Alok Rakkhit, Ionut Trestian.

Estimating b Based on an N-TCP pipe abstraction which

provides a bandwidth which is N times the throughput of a single TCP connection.▪ Use MulTCP to emulate the behavior▪ N is equal to the number of flows in the bundle

Node Architecture

12

q: target loss-ratec: available bandwidth

p: loss rateb: maximum sending rate

Page 13: L Subramanian*, I Stoica*, H Balakrishnan +, R Katz* *UC Berkeley, MIT + USENIX NSDI’04, 2004 Presented by Alok Rakkhit, Ionut Trestian.

Achieving target loss rate q FEC vs. ARQ trade-off

▪ Bandwidth overhead and packet recovery time FEC+ARQ based CLVL

▪ Restrict # of retransmissions to at most one▪ The expected packet loss rate

▪ The expected bandwidth overhead

▪ The optimal solution is when r1 = 013

After two rounds Goal

Minimizes

r is the redundancy factor

Page 14: L Subramanian*, I Stoica*, H Balakrishnan +, R Katz* *UC Berkeley, MIT + USENIX NSDI’04, 2004 Presented by Alok Rakkhit, Ionut Trestian.

Application-dependent proxy Choosing parameters

N as the average number of flows observed over a larger period of time q = 0.1%

Startup phase Using a slow-start phase to estimate the initial value of b

FEC implementation Operating on small window sizes (n < 1000) coding is not a

bottleneck

Page 15: L Subramanian*, I Stoica*, H Balakrishnan +, R Katz* *UC Berkeley, MIT + USENIX NSDI’04, 2004 Presented by Alok Rakkhit, Ionut Trestian.

Two enhancements The quality can be enhanced by converting

bursty losses into smooth losses for streaming audio

Recovering packets preferentially can improve the quality for MPEG streaming

Not consume any additional bandwidth Retransmits an important lost packet and drops

a later lesser important packet

15

Page 16: L Subramanian*, I Stoica*, H Balakrishnan +, R Katz* *UC Berkeley, MIT + USENIX NSDI’04, 2004 Presented by Alok Rakkhit, Ionut Trestian.

Streaming Audio

MPEG streaming

16

Perceptual Evaluation of Speech Quality (PESQ)(5 is ideal)

Increase0.15 – 0.2

Average loss rate Mazu-Korea – 2% Intel-Lulea – 3%

Not only improves the quality in the average casebut also the minimum quality of a stream

Page 17: L Subramanian*, I Stoica*, H Balakrishnan +, R Katz* *UC Berkeley, MIT + USENIX NSDI’04, 2004 Presented by Alok Rakkhit, Ionut Trestian.

Problem Client unable to connect to the server Cause skips or get disconnected

Alleviate the problem of bursty losses by performing: Recover from bursty network losses by using

an FEC+ARQ based CLVL Smoothly drop data packets equivalent to the

size of the burst at the overlay node Identify control packets based on packet size

and not drop these packets

17

Page 18: L Subramanian*, I Stoica*, H Balakrishnan +, R Katz* *UC Berkeley, MIT + USENIX NSDI’04, 2004 Presented by Alok Rakkhit, Ionut Trestian.

Sequence number plot illustrating smoothing of packet losses using OverQoS

18

Smoothing losses works well only when the bursty loss-periods are relatively short by compensating

Unable to achieve the target loss-rate due to congestion periods with very high loss-rates

10% loss-rate

Page 19: L Subramanian*, I Stoica*, H Balakrishnan +, R Katz* *UC Berkeley, MIT + USENIX NSDI’04, 2004 Presented by Alok Rakkhit, Ionut Trestian.

Methodology Wide-Area Evaluation Testbed

▪ RON and PlanetLab – use 19 diverse nodes Simulation Environment

▪ Ns-2 – a single congested link of 10 Mbps where they vary the background traffic▪ Long lived TCP connections▪ Self similar traffic▪ Web traffic

19

Page 20: L Subramanian*, I Stoica*, H Balakrishnan +, R Katz* *UC Berkeley, MIT + USENIX NSDI’04, 2004 Presented by Alok Rakkhit, Ionut Trestian.

Simulations

Wide Area Evaluation Achieve target over 80 of the 83 virtual links The causes of the other 3 virtual links

▪ Short outages – a period of time all packets are lost (< 5s)

▪ Bi-modal loss distributions – bursty losses

20

q = 0.1%

Page 21: L Subramanian*, I Stoica*, H Balakrishnan +, R Katz* *UC Berkeley, MIT + USENIX NSDI’04, 2004 Presented by Alok Rakkhit, Ionut Trestian.

21

Monitor 83 unique virtual linksu = 0.01 and u = 0.005

The value of cmin is greater than 100Kbpsfor more than 80% of the links

N-TCP, N = 10

Stability of cmin

1) The value of cmin is very stable, which does not deviate more than 10% around its mean

2) Set P = 1%, the actual value is no more than 1.3%

Calculate cmin based on a history of 200 seconds

The average sending rate of N-TCP is between 120Kbps to 2Mbps

Page 22: L Subramanian*, I Stoica*, H Balakrishnan +, R Katz* *UC Berkeley, MIT + USENIX NSDI’04, 2004 Presented by Alok Rakkhit, Ionut Trestian.

Overhead Characteristics

22

The difference between avg. loss & FEC+ARQ is the amount of FEC used in the second round

The burstier the background traffic, the higher the amount of FEC required to recover from these losses

Page 23: L Subramanian*, I Stoica*, H Balakrishnan +, R Katz* *UC Berkeley, MIT + USENIX NSDI’04, 2004 Presented by Alok Rakkhit, Ionut Trestian.

Delay Characteristics Two reasons for increasing delay

▪ The recovery process▪ Support in-sequence delivery of packets

23

Three different models(a)No packet ordering(b)End-to-end ordering(c) Hop-by-hop ordering

1) E2E is better than Hop-by-hop2) Adding new OverQoS nodes

increasing limited delay

Page 24: L Subramanian*, I Stoica*, H Balakrishnan +, R Katz* *UC Berkeley, MIT + USENIX NSDI’04, 2004 Presented by Alok Rakkhit, Ionut Trestian.

Three OverQoS bundles (with N=2, N=4, N=8) compete on a shared bottleneck under two different scenarios

24

No cross-traffic Cross-traffic

consisting of five long lived TCPs

1) Three OverQoS bundles co-exist with each other and with the background traffic

2) The ratio of throughputs of the three bundles is preserved

Page 25: L Subramanian*, I Stoica*, H Balakrishnan +, R Katz* *UC Berkeley, MIT + USENIX NSDI’04, 2004 Presented by Alok Rakkhit, Ionut Trestian.

OverQoS can enhance Internet QoS without any support from the underlying IP network

OverQoS is able to achieve the three enhancements with little (i.e., 5%) or no extra bandwidth.

Future work Combine admission control and path selection Determine the “optimal” placement of the

OverQoS nodes in the network

25