Maximizing End-to-End Network Performance Thomas Hacker University of Michigan October 5, 2001.

MaximizingEnd-to-End Network Performance

Thomas Hacker

University of Michigan

October 5, 2001

Introduction

• Applications experience network performance from a end customer perspective

• Providing end-to-end performance has two aspects– Bandwidth Reservation– Performance Tuning

• We have been working to improve actual end-to-end throughput using Performance Tuning

• This work allows applications to fully exploit reserved bandwidth

Improve Network Performance• Poor network performance arises from a

subtle interaction between many different components at each layer of the OSI network stack

• Physical• Data Link• Network• Transport• Application

TCP Bandwidth Limits – Mathis Equation• Based on characteristics from physical

layer up to transport layer.

• Hard Limits

• TCP Bandwidth, Max Packet Loss

p

C

RTT

MSSBW

2

*

RTTBW

MSSp

Packet Loss and MSS

• If the minimum link bandwidth between two hosts is OC-12 (622 Mbps), and the average round trip time is 20 msec, the maximum packet loss rate necessary to achieve 66% of the link speed (411 Mbps) is approximately 0.00018%, which represents only 2 packets lost out of every 100,000 packets.

• If MSS is increased from 1500 bytes to 9000 bytes (Jumbo frames), limit on TCP BW will rise by a factor of 6.

The Results

Web100 Collaboration

Effects of Host Tuning on EdgeWarp Data Transmission Performance

0

10

20

30

40

50

60

70

Untuned Bandw idth Tuned Bandw idth

Dat

a Tr

ansm

issi

on

Rat

e (M

b/s

ec)

Performance of PSockets on Abilene Network

0

20

40

60

80

0 50 100 150

Number of PSockets

Da

ta R

ate

in M

b/s

TCP WindowSize = 64K

SOURCE:Harimath Sivakumar, Stuart Bailey, Robert L. Grossman. “PSockets: The Case for Application-level Network Striping for Data Intensive Applications using High Speed Wide Area Networks,” SC2000: High-Performance Network and Computing Conference, Dallas, TX, 11/00

Parallel TCP Connections…a clue

Why Does This Work?• Assumption is that network gives best effort

throughput for each connection

• But end-to-end performance is still poor, even after tuning the host, network, and application

• Parallel Sockets are being used in GridFTP, Netscape, Gnutella, Atlas, Storage Resource Broker, etc.

Packet Loss

• Bolot* found that Random losses are not always due to congestion– local system configuration (txqueuelen in

Linux)– Bad cables (noisy)

• Packet losses occur in bursts• TCP throttles transmission rate on ALL

packet losses, regardless of the root cause• Selective Acknowledgement (SACK) helps,

but only so much* Jean-Chrysostome Bolot. “Characterizing End-to-End packet delay and loss in the Internet.”, Journal of High Speed Networks, 2(3):305--323, 1993.

Expression for Parallel Socket Bandwidth

nn

nagg

pRTT

MSS

pRTT

MSS

pRTT

MSSBW

22

2

11

1

n

aggp

MSS

p

MSS

p

MSS

RTTBW

21

1

Example

MSS = 4418, RTT = 70 msec, p = 1/10000 for all connections

Number of Connections

Aggregate Bandwidth

1 100 50 Mb/sec

2 100+100 100 Mb/sec

3 100+100+100 150 Mb/sec

4 4 (100) 200 Mb/sec

5 5 (100) 250 Mb/sec

n ip

1

Measurements

• To validate theoretical model, 220 4 minute transmissions performed from U-M to NASA AMES in San Jose, CA

• Bottleneck was OC-12, MTU=4418• 7 runs MSS=4366, 1 to 20 sockets• 2 runs MSS=2948, 1 to 20 sockets• 2 runs MSS=1448, 1 to 20 sockets• Iperf used for transfer, Web100 used to

collect TCP observations on sender side

Actual: MSS 1448 Bytes

Measured TCP Bandwidth vs. Number of Sockets (MSS 1448)

0

50

100

150

200

250

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20

Number of Parallel TCP Connections

Mea

sure

d A

ggre

gate

TC

P B

andw

idth

(Mb/

sec)

Actual: MSS 2948 Bytes

Measured TCP Bandwidth vs. Number of Sockets (MSS 2948)

0

50

100

150

200

250

300

350

400

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20

Number of Parallel TCP Connections

Mea

sure

d A

ggre

gate

TC

P B

andw

idth

(Mb/

sec)

Actual: MSS 4366 BytesMeasured TCP Bandwidth vs. Number of Sockets (MSS 4366)

0

50

100

150

200

250

300

350

400

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 19 18 20

Number of Parallel TCP Connections

Mea

sure

d A

gg

reg

ate

TC

P

Ban

dw

idth

(M

b/s

ec)

Sunnyvale – Denver Abilene Link

Initial Tests

Yearly Statistics

Abilene Weather Map

Conclusion

• High Performance Network Throughput is possible with a combination of host, network and application tuning along with using parallel TCP connections

• Parallel TCP Sockets mitigate negative effects of packet loss in random congestion regime

• Effects of Parallel TCP Sockets similar to using larger MSS

• Using Parallel Sockets is aggressive, but as fair as using large MSS