All Rights Reserved, Copyright(C) 2007, Hitachi, Ltd. 1 Transport-layer optimization for thin-client systems Yukio OGAWA Systems Development Laboratory,

All Rights Reserved, Copyright(C) 2007, Hitachi, Ltd.－ 1－

Transport-layer optimizationfor thin-client systems

Yukio OGAWASystems Development Laboratory, Hitachi, Ltd.

E-mail: [email protected]

Go HASEGAWA, Masayuki MURATAOsaka University

2007 International CQR WorkshopMay 15-17, 2007

All Rights Reserved, Copyright(C) 2007, Hitachi, Ltd.－ 2－

Overview of thin-client systems

office

Internet

data center

thin client

intranet

VPN gateway

server

satellite office, home, ‥

desktop service

without data, apps

thin client

user event

screen updates

Isolating computer resources from users

resource management, user mobility

System performancedepends on network performance

VPN: Virtual Private Network

TCP proxy

All Rights Reserved, Copyright(C) 2007, Hitachi, Ltd.－ 3－

Research objective and our approach

System performance (usability) depends on network performance. - intranet performance – designed in advance, controllable - Internet performance – uncontrollable

Improve performance of thin-client traffic - especially of flows traversing Internet - thin-client traffic = long-lived interactive TCP data flows - affected by TCP's Nagle algorithm and delayed ACK

- affected by buffering of TCP segments and SSR

Transport layer optimization on basis of actual traffic observations - observation of Hitachi SDL's prototype system - Dec. 20, 2006 to Jan. 25, 2007 - 168 pairs of a server and a thin-client - number of co-existing sessions during office hours: several dozen

research objective

drawback

our approach

All Rights Reserved, Copyright(C) 2007, Hitachi, Ltd.－ 4－

Characteristics of thin-client traffic- traffic patterns

client server

time time

request

response

large interval

interactive data flow(character information)

client server

time time

bulk data flow(screen update information)

request

response

•distinguished by interarrival time of response packets

size of data segment

time

MSSm ( n MSS + a )

MSS: Maximum Segment Size

•~102

packets•short interval

All Rights Reserved, Copyright(C) 2007, Hitachi, Ltd.－ 5－

Characteristics of thin-client traffic- interarrival time distribution of request packets

data

segm

ent

size

(lo

g10 b

yte

s)

access from Internet

interarrival time of request packets (log10 sec)-6 -5 -4 -3 -2 -1 0 1 2 3

0.5

1.0

1.5

2.0

2.5

3.0

3.5

4.0

0.0

All Rights Reserved, Copyright(C) 2007, Hitachi, Ltd.－ 6－

Characteristics of thin-client traffic- interarrival time distribution of response packets

data

segm

ent

size

(lo

g10 b

yte

s)

access from Internet

interarrival time of response packets (log10 sec)-6 -5 -4 -3 -2 -1 0 1 2 3

0.5

1.0

1.5

2.0

2.5

3.0

3.5

4.0

0.010-2.2（ 6.3 m） sec

interactive

bulk (head)

bulk (inside of 'nMSS+a')

bulk(head of 'nMSS+a')

data segment size

time

MSShead

inside

head of 'nMSS+a'inside of 'nMSS+a'

bulk

interactive

All Rights Reserved, Copyright(C) 2007, Hitachi, Ltd.－ 7－

Proposed methods for improving performance- interactive data flow

client server

time time

request

response

gateway (TCP proxy)

×

Ti

ti

sending copy of data packet

sending interval ：

　 ti = 　 　 min( RTT – RTTmin , Ti / 2 )

1h

All Rights Reserved, Copyright(C) 2007, Hitachi, Ltd.－ 8－

Proposed methods for improving performance- bulk data flow

client server

time time

gateway (TCP proxy)

requestresponse

no SSR

data segment size

time

MSSm ( n MSS + a )

data segment size

time

MSSn MSS + a

MSS: Maximum Segment SizeSSR: Slow-Start Restartpaused for buffering

resegmenting TCP data segments

All Rights Reserved, Copyright(C) 2007, Hitachi, Ltd.－ 9－

Simulation model- system model

R

R

R router

20 Mbps, 5 msec

sender host(server)

gateway(TCP proxy)

receiver host(client)

20 Mbps, 0.1 msec

1 Mbps, 30 – 300 msec

100 Mbps, 0.1 msec

intranet

Internet

receiver hosts

sender hosts

thin-client traffic

background traffic(UDP: 64 bytes, 128 Kbps) x n

packet drop ratio: 0, 3%

tail-drop router(buffer size ： 50, 1024 packets)

bottleneck link

All Rights Reserved, Copyright(C) 2007, Hitachi, Ltd.－ 10－

Simulation model- thin-client traffic for evaluation

interactive

bulk(-0.6, -1.3) : -1.3 = mean - 2 std

(-0.6, -0.6)

access from Internet

aver

age

inte

rarr

ival

tim

e of

resp

onse

pac

kets

(log

10 s

ec )

evaluation traffic•number ： 30•duration ： 60 sec

-6 -5average interarrival time of response data flows (contiguous packets) (log10 sec)

-4 -3 -2 -1 0 1 2 3

-5

-4

-3

-2

-1

0

1

2

-6

3

All Rights Reserved, Copyright(C) 2007, Hitachi, Ltd.－ 11－

Simulation results- interactive data flow – packet drop

avera

ge n

um

ber

of

pack

et

dro

ps

(log

10

)

transmission delay of bottleneck link (log10 msec)

(UDP 1024 Kbps)

(UDP 1152 Kbps)

(UDP 1280 Kbps)

bottleneck link - 1 Mbps - 3% drop ratiorouter buffer - 50 packets

send a copy with pause

send a copy without pause

send no copies

bg: background

102

100

10-1

101

101 102 103 101 102 103 101 102 103

102

100

10-1

101

102

100

10-1

101

drop from tail-drop router

random drop from bottleneck link

All Rights Reserved, Copyright(C) 2007, Hitachi, Ltd.－ 12－

Simulation results- bulk data flow – transfer time

med

ian

tra

nsf

er

tim

e (

log

10

sec)

number of packets in bulk data flow (log10)

buffer size = 1024 packets

bottleneck link - 1 Mbps - 80 msec - 0% drop ratiobackground - 3 UDP flows (= 384 Kbps)

no-SSR, resegmentationno-SSR

101

100

10-1

10-2

101

100

10-1

10-2

100 101 100 101

buffer size = 50 packetsSSR, resegmentation

SSR

All Rights Reserved, Copyright(C) 2007, Hitachi, Ltd.－ 13－transmission delay of bottleneck link (log10 msec)

Simulation results- bulk data flow – drop from tail-drop router

(UDP 384 Kbps)

no-SSR, resegmentationno-SSR

SSR, resegmentationSSR

bottleneck link - 1 Mbps - 0% drop ratiorouter buffer - 50 packets

(UDP 768 Kbps)

bg: background

avera

ge n

um

ber

of

pack

et

dro

ps

(log

10

)102

101

10-1

10-2

100

102

101

100

101 102 103101 102 103

All Rights Reserved, Copyright(C) 2007, Hitachi, Ltd.－ 14－

Conclusion

•for interactive data flows (transferring character information) - send a packet copy with pause

⇒ increases tolerance for packets drops

•for bulk data flows (transferring screen update information) - disable TCP slow-start restart

⇒ increases packet sending rate⇒ increases burstiness of traffic

- resegment TCP data segments⇒ reducees burstiness of traffic

TCP optimization for improving performance of thin-client traffic