Performance Evaluation and Comparison of Westwood+, New Reno and Vegas TCP Congestion Control

Saverio Mascolo [email protected]

http://www-ictserv.poliba.it/mascolo/ Dipartimento di Elettrotecnica ed Elettronica Politecnico di Bari Via Orabona 4, 70125 Bari, Italy Gotland August 25, 2004 1

   

Outline

Brief description of Westwood+ TCP, New Reno and Vegas TCP Performance evaluation of Westwood+, New Reno and Vegas TCP using ns-2 Internet measurements using an implementation of Westwood+ in Linux 2.4 (Westwood+ is now in the official kernel of Linux 2.6 at www.kernel.org) For more details see: ACM Computer Communication Review, April 2004, and ICC04 Saverio Mascolo – WIP/Beats 04, Visby, Gotland 2

WESTWOOD+ TCP



key idea of Westwood+:

 to use the stream of ack packets to get an e2e estimate of the available bandwidth to be used for setting

cwnd and ssthresh

(whereas standard TCP implements a “blind” by half window decrease) after congestion Saverio Mascolo – WIP/Beats 04, Visby, Gotland 3

TCP Westwood+ cwnd ssthresh Congestion Avoidance BWE*RTTmin Slow start Adaptive decrease cwnd=ssthr=BWE*RTTmin Timeout time Westwood Adaptive decrease vs (New) Reno blind by ½ window shrinking

Saverio Mascolo – WIP/Beats 04, Visby, Gotland 4

Rate Halving



Another feature of the Linux implementation of Westwood+ TCP is the rate-halving mechanism (taken from New Reno):

 after congestion the cwnd is reduced by one every two ack packets up to the value BWE*RTTmin Saverio Mascolo – WIP/Beats 04, Visby, Gotland 5

E2E bandwidth estimation packets packets SENDER RECEIVER Filter Network Bandwidth estimate ACKs ACKs



The rate of returning ACKS is exploited to estimate the “best-effort” available bandwidth

An anti-aliasing filter in packet networks

b j



d j



j

Antialiased samples 

j



Last RTT d j



all data

acknowledg ed in the last RTT

Saverio Mascolo – WIP/Beats 04, Visby, Gotland 7

ACK compression effects

 ACK pairs give information about the bandwidth of the last link traversed on the backward path  To smooth ACK compression we accumulate ACKs over an RTT and then compute a bandwidth sample Saverio Mascolo – WIP/Beats 04, Visby, Gotland 8

 We are currently using the standard exponential filter

b

ˆ

k

  

b

ˆ

k

 1  ( 1   ) 

b k b k



d k RTT k

Saverio Mascolo – WIP/Beats 04, Visby, Gotland 9

Summary on bandwidth estimate

 Westwood TCP: one bandwidth sample computed for each ACK (Mobicom 01)=>> Bandwdith overestiamte (when ACK compression)  Westwood+ TCP: one bandwidth sample for each RTT (see ACM CCR, April 04) Saverio Mascolo – WIP/Beats 04, Visby, Gotland 10

Advantages of Westwood+ TCP



higher throughput over wireless links because losses due to unreliable links do not provoke overshrinking of the congestion window



Improved fairness wrt to Reno proportional to 1/sqrt(RTT) ) (Reno throughput is proportional to 1/RTT whreas Westwood throughput is

New Reno TCP

 New Reno is an improved version of Reno that avoids multiple reductions of the cwnd when several segments from the same window of data get lost  RFC 3782, S. Floyd, T. Henderson, A. Gurtov, “The NewReno Modification to TCP's Fast Recovery Algorithm”  New Reno is the leading Internet congestion control protocol Saverio Mascolo – WIP/Beats 04, Visby, Gotland 12

Vegas TCP

 Vegas TCP was the first attempt to depart from the loss driven paradigm of the TCP by introducing a mechanism of congestion detection before packet losses:  Vegas TCP computes the difference  between the actual input rate (cwnd/RTT) and the expected rate (cwnd/RTT

min

), to infer network congestion. If  is smaller than a threshold α then the cwnd is additively increased, whereas if the difference is greater than another threshold β then the cwnd is Additively Decreased; finally, if α < then the cwnd is kept constant. It has been shown that Vegas TCP ensures network stability but it is not able to grab its own bandwidth share when interacting with algorithms that systematically hits network queue capacity such as Reno.

 <β Saverio Mascolo – WIP/Beats 04, Visby, Gotland 13

   TCP Vegas has been considered because:  Analogously to Westwood+, it is based on mechanism for throttling the congestion window based on RTT measurements.

Vegas TCP is behind the new Fast TCP (by researchers at Caltech ). In authors’ words, “Fast TCP is a sort of high speed version of Vegas”.

Fast TCP is still in a trial phase and no kernel code or ns-2 implementation available at time of this work  Being based on RTT measurements to infer congestion, Fast TCP could inherit all drawbacks of Vegas that will be illustrated ( incapacity to grab bandwidth when coexisting with Reno traffic or in the presence of reverse traffic) Saverio Mascolo – WIP/Beats 04, Visby, Gotland 14

 Following our suggestions, Les Cotrell at Stanford Linear Accelerator Center found that Fast TCP is, in his words “very handicapped in the presence of reverse traffic” (fall 2003) Saverio Mascolo – WIP/Beats 04, Visby, Gotland 15

Performamce evaluation

 Bandwidth estimate Saverio Mascolo – WIP/Beats 04, Visby, Gotland 16

Topology with ACK compression effects (10 Mbps) 10 TCP 10 TCP Sinks Forward traffic: West TCP connection R R Reverse traffic: 20 TCP connections

Saverio Mascolo – WIP/Beats 04, Visby, Gotland

Sink 10 TCP Connections

17

The 20 Westwood+ connections estimate a best-effort available bandwidth that reasonably approaches the fair share of 0.5Mbps

1.0E+09 1.0E+08 1.0E+07 1.0E+06 1.0E+05 1.0E+04 1.0E+03 1.0E+02 1.0E+01 0 Fair share 20 40 60 80 100 s Saverio Mascolo – WIP/Beats 04, Visby, Gotland 18

Westwood overestimates up to 100 times the fair share due to ACK compression

1.0E+09 1.0E+08 1.0E+07 1.0E+06 1.0E+05 1.0E+04 1.0E+03 1.0E+02 1.0E+01 0 20 40 s 60 Saverio Mascolo – WIP/Beats 04, Visby, Gotland 80 Fair share 100 19

Remark

  Westwood estimate is different from measuring the low frequency components of the sending rate cwnd/RTT (cwnd/RTT is the measure of the instantaneous throughput employed by Vegas TCP) In fact, the Vegas actual rate cwnd/RTT is a measure of the available bandwidth that is based on the number of sent packets (cwnd) and not on the number of acknowledged packets d

k

. As a consequence, Vegas samples do not take into account that a fraction of sent packets could be lost thus leading to available bandwidth overestimate. Saverio Mascolo – WIP/Beats 04, Visby, Gotland 20

1 Mbps

1.2E+06 1.1E+06 1.0E+06 9.0E+05 8.0E+05 7.0E+05 0 20 40 s 60 Saverio Mascolo – WIP/Beats 04, Visby, Gotland BWE Input Rate Bottleneck Capacity 80 100 21

Formula of Steady State Throughput

T

Reno  1

RTT

2

(

1 

p ) p T

West  1

RTT



T q

1 

p p

Saverio Mascolo – WIP/Beats 04, Visby, Gotland 22

Single connection + 10 TCP connections on the backward path following an OFF-ON-OFF-ON pattern to investigate the effect of reverse traffic

TCP 1 source TCP 1 Sink Forward Traffic 10 TCP Sinks Reverse Traffic Saverio Mascolo – WIP/Beats 04, Visby, Gotland 10 TCP sources 23

cwnd and ssthresh dynamics

100 90 80 30 20 10 0 70 60 50 40 0 NewReno cwnd ssthresh 100 90 80 70 60 50 40 30 20 10 0 0 Westwood+ cwnd ssthresh 100 200 300 400 500 s 600 700 800 900 1000 100 200 300 400 500 s 600 700 800 900 1000 100 90 80 70 60 50 40 30 20 10 0 0 Vegas cwnd ssthresh 120 100 80 60 40 20 cwnd ssthresh Reno 100 200 300 400 500 s 600 700 800 900 1000 0 0 100 200 300 400 500 s 600 700 800 900 1000 Saverio Mascolo – WIP/Beats 04, Visby, Gotland 24

70000 60000 50000 40000 30000 20000 10000 0 0

Visual look at fairness 20 connections

NewReno 200 Vegas 400 600 s 70000 60000 50000 40000 30000 20000 10000 0 0 800 1000 70000 60000 50000 40000 30000 20000 10000 0 0 Westwood + 200 200 400 s 600 800 1000 Saverio Mascolo – WIP/Beats 04, Visby, Gotland 400 s 600 800 1000 25

Wireless terrestrial scenario

one way delay of TCP1 =125ms; 20ms delay on the wireless link (2Mbps)

5 TCP sources Cross Traffic 5 TCP Sinks Wireless link TCP1 source 10 TCP Sinks Reverse Traffic 10 TCP sources TCP1 sink Saverio Mascolo – WIP/Beats 04, Visby, Gotland 26

   RTTs 5 cross traffic connections and 10 New Reno backward traffic connections are uniformly spread in the intervals [66ms,250ms] and [46ms,250ms], respectively.

wireless link affected by bursty segment losses in both directions A Gilbert two state Markov chain models the loss process      loss probability p equal to 0, when channel in the Good state, p =0.1 when the channel is in the Bad state.

permanence time in the Good state deterministic and equal to 1s permanence time in the Bad state also deterministic ranging from 0.1ms to 100 ms.

When the permanence time in a state elapses, the state can transit to a Good or Bad state with a probability p=0.5.

Saverio Mascolo – WIP/Beats 04, Visby, Gotland 27

Single connection

  For each considered case, we run 10 simulations by varying the seed of the random loss process. For each value of the BAD state duration we report the maximum, minimum and average goodputs.

In order to analyze only the impact of bursty losses on the TCP behavior, we have first turned off both the cross and reverse traffic sources. This simple scenario is particularly useful to investigate the effectiveness of the adaptive decrease paradigm when losses not due to congestion are experienced by the TCP. Saverio Mascolo – WIP/Beats 04, Visby, Gotland 28

Goodput of TCP1 connection without reverse traffic: DACK enabled

2 .0E+06 1 .8E+06 1 .6E+06 1 .4E+06 1 .2E+06 1 .0E+06 8 .0E+05 6 .0E+05 4 .0E+05 2 .0E+05 0 .0E+00 0 .00 01 W es twood+, DA CK e nabled New R eno, D AC K ena ble d 0.0 01 0. 01 D uration of the B A D sta te (s) Saverio Mascolo – WIP/Beats 04, Visby, Gotland 0 .1

29

Goodput of TCP1, no reverse traffic: DACK disabled

2 .0E+06 1 .8E+06 1 .6E+06 1 .4E+06 1 .2E+06 1 .0E+06 8 .0E+05 6 .0E+05 4 .0E+05 W est wood+, DA CK disabled New R eno, D AC K disabled 2 .0E+05 0 .0E+00 0 .00 01 SA C K 0.0 01 0. 01 D uration of the B A D sta te (s) Saverio Mascolo – WIP/Beats 04, Visby, Gotland 0 .1

30

Cwnd and ssthresh + ( duration of the BAD 0.01s)

Westwood+ 100 90 80 70 60 50 40 30 20 10 0 0 cwnd ssthresh 100 200 300 400 500 s 600 700 800 900 1000 New Reno 100 90 80 70 60 50 40 30 20 10 0 0 cwnd ssthresh 100 200 300 400 500 s 600 700 800 900 1000 Saverio Mascolo – WIP/Beats 04, Visby, Gotland 31

reverse + cross traffic

  One further point valuable of investigation is when Westwood+ shares the wired portion of the network with several TCP flows on the forward and backward paths.

For that purpose, we turn on the cross and reverse traffic and we measure the goodput of the TCP1 connections for various values of the BAD state duration.

Saverio Mascolo – WIP/Beats 04, Visby, Gotland 32

9 00000 8 00000 7 00000 6 00000 5 00000 4 00000 3 00000 2 00000 1 00000 0.0001

W es tw ood+, D A CK disabled New R eno, D AC K disa ble d SA C K W es tw ood+, D A CK ena ble d New R eno, DA CK e nabled 0.001

0.01

Du ration of the B A D state (s) Saverio Mascolo – WIP/Beats 04, Visby, Gotland 0 .1

33

Remarks

  The delayed ACK option plays a major role in this scenario.

In fact, protocols that do not employ the delayed ACK option provides goodputs that are roughly two times larger than those obtained when the delayed ACK option is enabled. The reason is that the delayed ACK option slows down the TCP probing phase.

In these scenarios Westwood+ TCP (DACK disabled) still improves the goodput with respect to New Reno (DACK disabled) and SACK TCP, but the improvement is now only up to roughly 20% since in this case the TCP1 connection loses bandwidth in favor of the cross traffic that, being wired, is not penalized by losses not due to congestion.

Saverio Mascolo – WIP/Beats 04, Visby, Gotland 34

Satellite scenario

20 TCP senders 20 TCP sinks 10 TCP sinks 10 TCP senders •20 TCP forward connections in the presence of reverse traffic contributed by 10 long-lived New Reno connections. •large leaky pipe: 10Mbps bottleneck link with one-way delay equal to 275ms •RTTs of the forward connections are equal to 590ms. Saverio Mascolo – WIP/Beats 04, Visby, Gotland 35

100 00000 90 00000 80 00000 70 00000 60 00000 50 00000 40 00000 30 00000 20 00000 10 00000 0 0.0001

W e stwood+ , D AC K disa ble d W es tw ood+, D A CK e nabled SA C K New R eno, D AC K disable d Ne w R eno , DA CK e nabled 0.001

0.01

0.1

Du ration of the B A D state (s) Saverio Mascolo – WIP/Beats 04, Visby, Gotland 1 36

Linux 2.4.19

implementation of Westwood+ TCP

 More than 4000 FTP froma Bari, South Italy to:  signserv.signal.uu.se (Uppsala)  panther.cs.ucla.edu (UCLA)  main.penguin.it (Parma)

Average goodput during ftp Angeles to Los

50 45 40 15 10 5 0 35 30 25 20 New Reno Westwood+ 21 Feb 2003 32MB 26 Feb 2003 3.2MB

28 Feb 2003 3.2 MB 14 Mar 2003 32MB 19 Mar 2003 32 MB 21 Mar 2003 32 MB Saverio Mascolo – WIP/Beats 04, Visby, Gotland 38

Retransmission ratio during ftp UCLA to

12 10 8 6 4 New Reno Westwood+ 2 0 21 Feb 2003 32MB 26 Feb 2003 3.2MB

28 Feb 2003 3.2MB

14 Mar 2003 32MB 19 Mar 2003 32MB 21 Mar 2003 32MB Saverio Mascolo – WIP/Beats 04, Visby, Gotland 39

Average goodput during ftp Uppsala to

250 200 New Reno Westwood+ 150 100 50 0 14 Dec 2002 32 MB 17 Dec 2002 3.2MB

10 Jan 2003 3.2MB

12 Jan 2003 3.2MB

13 Jan 2003 32MB 17 Jan 2003 3.2MB

3 Feb 2003 32MB 7 Feb 2003 32MB Saverio Mascolo – WIP/Beats 04, Visby, Gotland 40

Retransmission ratio during ftp Uppsala to

4.5

4 3.5

3 2.5

2 1.5

1 0.5

0 17 Dec 2002 3.2MB

10 Jan 2003 3.2MB

12 Jan 2003 3.2 MB 13 Jan 2003 32MB New Reno Westwood+ 17 Jan 2003 3.2MB

3 Feb 1003 32MB 7 Feb 2003 32MB Saverio Mascolo – WIP/Beats 04, Visby, Gotland 41

Average goodput during ftp to Parma

160 140 120 100 80 60 40 20 0 26 Mar 2003 32MB New Reno Westwood+ 27 Mar 2003 3.2MB

28 Mar 2003 3.2MB

4 Apr 2003 32MB 7 Apr 2003 3.2MB

9 Apr 2003 3.2MB

11 Apr 2003 3.2MB

Saverio Mascolo – WIP/Beats 04, Visby, Gotland 42

Retransmission ratio during ftp Parma to

14 12 10 8 6 4 2 0 26 Mar 2003 32MB New Reno Westwood+ 27 Mar 2003 3.2MB

28 Mar 2003 3.2MB

4 Apr 2003 32MB 7 Apr 2003 3.2MB

9 Apr 2003 3.2MB

11 Apr 2003 3.2MB

Saverio Mascolo – WIP/Beats 04, Visby, Gotland 43

Uploads to panther.cs.ucla.edu (1)

File size=3.2MB, From: rigel.poliba.it, To: panther.cs.ucla.edu, Total number of uploads = 197, Average New Reno Goodput = 16.86Kbyte/s, Average Westwood+ Goodput = 25.21Kbyte/s 60 50 40 Westwood+ New Reno 30 20 10 0 Wed Feb 26 13:35:16 2003 Wed Feb 26 19:38:03 2003 Wed Feb 26 22:53:27 2003 Thu Feb 27 Thu Feb 27 Thu Feb 27 Thu Feb 27 01:41:26 2003 04:10:01 2003 06:11:28 2003 08:23:06 2003 Date

Uploads to panther.cs.ucla.edu (2)

File size=32MB 80 70 60 Westwood+ New Reno 50 40 30 20 10 0 Fri Feb 21 18:15:56 2003 Sat Feb 22 06:04:31 2003 Sat Feb 22 20:26:44 2003 Sun Feb 23 Sun Feb 23 Mon Feb 24 07:24:58 19:06:51 03:52:51 2003 2003 2003 Date

Uploads to panther.cs.ucla.edu (3)

File size=32MB 60 50 40 30 Westwood+ New Reno 20 10 0 Fri Mar 14 18:59:15 2003 Sat Mar 15 06:09:27 2003 Sat Mar 15 13:29:48 2003 Sat Mar 15 22:20:43 Date 2003 Sun Mar 16 Mon Mar 17 09:57:01 2003 02:47:05 2003

Uploads to panther.cs.ucla.edu (4)

File size=32MB 70 60 50 40 30 20 10 Westwood+ New Reno 0 Wed Mar 19 14:15:25 2003 Wed Mar 19 22:55:38 2003 Thu Mar 20 02:47:10 2003 Thu Mar 20 06:08:36 2003 Thu Mar 20 10:45:34 2003 Date Thu Mar 20 19:52:36 2003 Fri Mar 21 01:07:10 2003 Fri Mar 21 05:29:19 2003

Further research

 Realistic Characterization of wireless links still needed in ns-2!

 Can we expect the same result with Westwood+ over the gigabit Internet? (tests are going at SLAC and CERN) Saverio Mascolo – WIP/Beats 04, Visby, Gotland 48

 Thanks for the attention and Questions?

Saverio Mascolo – WIP/Beats 04, Visby, Gotland 49