Micro- and macroscopic analysis of RTT variability in GPRS ...

Jorma Kilpi @ PAN-NET Research Seminar 16.3.2006. 1

Micro- and macroscopic analysis of RTT variability in

GPRS and UMTS networks

Jorma Kilpi Pasi Lassila

Systems Research Networking Laboratory

VTT Technical Research Centre of Finland Helsinki University of Technology

P.O.Box 12022, FIN 02044 VTT, Finland P.O.Box 3000, FIN 02015 TKK, Finland

Email: [email protected] Email: [email protected]

Jorma Kilpi @ PAN-NET Research Seminar 16.3.2006. 2

Semi-RTT

GPRS network

GGSN FirewallFirewall

Service network

PublicInternetPublicInternet

mobile hostmobile host external hostexternal host

Gn Gi

datadata

ACKsACKs

semi-RTT

• RTT process:

(ti, RTT (ti)) i = 1, . . . , n

wheren = number ofvalid semi-RTT samples observed from the flow andti is the

time stamp of the ACK packet at the Gi interface.

Jorma Kilpi @ PAN-NET Research Seminar 16.3.2006. 3

Why micro- and macroscopic analysis?

• How individual flows see the RTT process? (Microscopic level)

• What information does theaggregate RTT process tell us? (Macroscopic level)

• Please note:We use the word ’aggregate’ in two cases when we consider

1. aggregate of all TCP flows from the same mobile or

2. aggregate of all TCP flows from all of the mobiles.

Jorma Kilpi @ PAN-NET Research Seminar 16.3.2006. 4

Reconstruction of TCP flows

• Mobile TCP connections: GSM/GPRS and UMTS access.

• Mobile host <-> Internet host.

• All down- and upstream traffic of a TCP connection goes through the same GGSN.

• We used a program called Tstat (http://tstat.tlc.polito.it/) which

reconstructs TCP connections from TCP/IP packet level data.

• Moreover, Tstat was modified slightly in order to obtain RTT processes(ti, RTT (ti)),

i = 1, . . . , n of a large number of flows.

• Biased view of mobile traffic in the sense that we only present analysis of

non-anomalious successful TCP connections.

Jorma Kilpi @ PAN-NET Research Seminar 16.3.2006. 5

Examples of observed RTT processes

15 16 17 18 19Time (hour)

2

4

6

8

10

RT

T(s

)

10:30-18:00, longest

17 18Time (hour)

2

4

6

8

10

RT

T(s

)

10:30-18:00, 2:nd longest

13 14 15Time (hour)

2

4

6

8

10

RT

T(s

)

10:30-18:00, 3:rd longest

13 14 15 16 17Time (hour)

2

4

6

8

10

RT

T(s

)

10:30-18:00, 4:th longest

Jorma Kilpi @ PAN-NET Research Seminar 16.3.2006. 6

11Time (hour)

2

4

6

8

10

RT

T(s

)

10:30-18:00, 5:th longest

Time (hour)

2

4

6

8

10

RT

T(s

)

10:30-18:00, 6:th longest

12Time (hour)

2

4

6

8

10

RT

T(s

)

10:30-18:00, 7:th longest

12Time (hour)

2

4

6

8

10

RT

T(s

)

10:30-18:00, 8:th longest

Jorma Kilpi @ PAN-NET Research Seminar 16.3.2006. 7

17Time (hour)

2

4

6

8

10

RT

T(s

)

10:30-18:00, 9 :th longest

18Time (hour)

2

4

6

8

10

RT

T(s

)

10:30-18:00, 10:th longest

11Time (hour)

2

4

6

8

10

RT

T(s

)

10:30-18:00, 11:th longest

11 12 13Time (hour)

2

4

6

8

10

RT

T(s

)

10:30-18:00, 12:th longest

Jorma Kilpi @ PAN-NET Research Seminar 16.3.2006. 8

Microscopic level: Flow 1 in TCP port 80 (HTTP) but not a web page downloading!

15:00 16:00 17:00 18:00 19:00Time (hour)

1

2

3

4

5

6

RT

T(s

)

Flow 1

2.25 2.5 2.75 3 3.25 3.5 3.75 4Time Period (s)

0

50

100

150

200

Spe

ctra

lPow

er

Lomb Periodogram of Flow 1

Begin

Middle

2.25 2.5 2.75 3 3.25 3.5 3.75 4Time Period (s)

0

50

100

150

200

Spe

ctra

lPow

er

Lomb Periodogram of Flow 1

Middle

End

1 2 3 4 5 6 7 8Time Period (s)

20

40

60

80

100S

pect

ralP

ower

Lomb Periodogram of Flows 1 and 2

Flow 1

Flow 2

Jorma Kilpi @ PAN-NET Research Seminar 16.3.2006. 9

Microscopic level: Flow 2, also in TCP port 80 but lasts abouthalf an hour!

17:30 17:45Time (hour)

2.55

7.510

12.515

17.520

RT

T(s

)

Flow 2

0 5 10 15 20 25 30 35Time from the First Segment (min)

0

2

4

6

8

10

Vo

lum

e(M

B)

Flow 2

End: 26.8 kb/s

Begin: 40.2 kb/s

17:15 17:30 17:45 18:00 18:15Flow End Points

0

20

40

60

80

100

i:th

Flo

w

Simultaneous flows with Flow 2

Flow 2

• An example of the effect of simul-

taneous TCP-connections from the

same mobile.

Jorma Kilpi @ PAN-NET Research Seminar 16.3.2006. 10

Macroscopic level: Dominating RTT values

2 4 6 8 10 12j:th Scale

23456789

log 2

Ej

Energy Function Plot

0 1 2 3 4 5 6 7 8 9 10RTT (s)

0

0.002

0.004

0.006

0.008

PD

F

1.3 4.3Positions of Spikes in PDFs

Sample 1

Sample 2

0.7 1.9

• A paper by Huang, Feldmann and

Willinger used wavelets to detect

network performance problems.

• Energy Function Plots (EFPs) using

Haar wavelets showed local period-

icity in the range600ms − 5s.

• PDF of aggregate RTT processes

show that the probability mass essen-

tially lies between the same600ms−

5s!

• The positions of spikes are due to de-

terministic reasons.

• The (backbone) network as a whole

is not significantly congested during

the busy hours.

Jorma Kilpi @ PAN-NET Research Seminar 16.3.2006. 11

Macroscopic level: Self-congestion

0 5 10 15 20 25

Simultaneous TCP Connections (Max )

2.55

7.510

12.515

17.520

Max

RT

T(s

)

Conditional Expectation

Group 1

Group 0

0 5 10 15 20 25Simultaneous TCP Connections (Max)

2.55

7.510

12.515

17.520

Max

RT

T(s

)

Conditional Expectation (UMTS)

• Aggregate traffic from the same mo-

bile.

• Group 0: The set of those mobiles

that essentially did not send any data

in the uplink.

• Group 1: At least one of the flows

of the mobile in this group had non-

trivial simultaneous uploading.

• Robust estimates of conditional ex-

pectation.

• For individual flows, self-congestion

is due to simultaneous flows from

same mobiles are the main reason for

the observed RTT variability.

Jorma Kilpi @ PAN-NET Research Seminar 16.3.2006. 12

Macroscopic level: Uploading is more critical than simultaneousdownloadings

20 40 60 80 100 120

TCP Connections (Max )

50

100

150

200

250

Max

RT

T(s

)

Group 0

20 40 60 80 100 120

TCP Connections (Max )

50

100

150

200

250

Max

RT

T(s

)

Group 1

• Group 0: The set of those mobiles

that essentially did not send any data

in the uplink.

• Group 1: At least one of the flows of

the mobile in this group had a non-

trivial simultaneous uploading.

Jorma Kilpi @ PAN-NET Research Seminar 16.3.2006. 13

Conclusions and further research topics

• Dominating RTT values told that the (backbone) network was not significantly

congested during busy hours.

• Self-congestion and uploading are critical.

• Wavelets seem to be really a powerful tool.

– EFP gave the same information about the congestion level of the network from the

packet level data than the histogram of all RTTs,i.e., without reconstruction ofTCP flows.

• Usefulness of the Lomb periodogram?

– TCP port 80 does not indicate the true application.

– Distinguishing streaming applications from true file downloadings?

– Could the ACK packets alone be used? (Without reconstruction of TCP

connections)

Jorma Kilpi @ PAN-NET Research Seminar 16.3.2006. 14

• Poster (short paper) in Networking 2006 conference, May 15th-19th, Coimbra,

Portugal.

• Questions?