Passive Monitoring of RTT spikes Jorma Kilpi VTT Information - - PowerPoint PPT Presentation

ISMA 2004 Workshop on Internet Signal Processing 1

Passive Monitoring of RTT spikes

Jorma Kilpi VTT Information Technology P.O.Box 1202, 02044 VTT, FINLAND Email: Jorma.Kilpi@vtt.fi (Joint work with Pasi Lassila form HUT) 5th November 2004

VTT Information Technology

ISMA 2004 Workshop on Internet Signal Processing 2

Problem formulation:

• Sudden increase in the RTT of a TCP/IP connection is called an RTT spike.

– Such a phenomenom may exist specially in mobile/wireless environments. – Spikes are not congestion related but, instead, are unpredictable for the TCP. – In mobile connections mobility management and mobile routing are possible sources of such spikes. – In wireless connections the distance to the base station is one possible source of spikes.

• In the worst case the RTT estimation algorithm of TCP gets confused, spurious

timeouts and unnecessary retransmissions cause loss of goodput for TCP. (A Strict RTT spike.)

• Even without a retransmission the goodput of TCP gets worse. (A [non-strict] RTT

spike.)

• What is the statistical significance of this phenomenom and how to monitor it

passively? Statistical inference and modeling?

VTT Information Technology

ISMA 2004 Workshop on Internet Signal Processing 3

Available measurements: (In GSM/GPRS, EDGE/GPRS)

• All down- and upstream traffic of

a TCP flow goes through the same GGSN.

• Time difference of a data segment

and the corresponding ACK, meas- ured at GGSN is called ’half’ RTT.

• There are four different ’half’ RTT

cases: End host Mobile Internet Client

• ✁

• ✁

Server

• ✁

• ✁

• We concentrate for
• ✁

.

VTT Information Technology

ISMA 2004 Workshop on Internet Signal Processing 4

Approach/methodology used: (For statistical inference and modeling purposes)

• We use spectroscopy (Radon transform) to find out link characteristics like the number
• f uplink Packet Data Channels (PDCHs) and channel codings used in the uplink.

Transmission Plane MS BSS BTS BSC SGSN GGSN Um Abis Gb Gn Gi IP IP TCP RF RF CCU RLC/MAC RLC/MAC

• Given packet sizes
• ✁

and packet interarrival times

, measured at the GGSN, we analyze empirical probabilities

• ✆

• The Radon transform in our case is defined as

• ✆

• ✆✸✷

• ✝✼

VTT Information Technology

ISMA 2004 Workshop on Internet Signal Processing 5

Preliminary results: (How the Radon transform method works)

40 200 576 1024 1500 Packet Size (B) 480 960 1440 1920 Interarrival Delay (ms) Example 2 500 1000 1500 2000 v (kb/s) 4.4 4.6 4.8 5 5.2 H(v) Four Cases CS-2 and 2 PDCHs

300 350 400 450 500 v (kb/s) 3.9 4 4.1 4.2 4.3 H(v) Example 2

• This analysis is based only on time

stamps and packet sizes.

• Minimum
• ✁

?

• According to link characteristics we

try to group TCP flows into as homo- geneous classes as possible.

VTT Information Technology

ISMA 2004 Workshop on Internet Signal Processing 6

Validation: (That RTT spikes can be observed passively)

25 50 75 100 125 150 Number of Retransmissions 0.001 0.01 0.1 1 10 100 Frequency (%) Observed Unnecessary Retransmissions TCP-flows with SACK Option Required

50 100 150 200 250 300 RTT count 2.5 5 7.5 10 12.5 15 17.5 20 Normalized Range of RTTMC RTTMC 1000 2000 3000 4000 Min RTTMC (ms) 1000 2000 3000 4000 5000 6000 7000 8000 Mean RTTMC (ms) RTTMC

• It seems that at least

• f TCP

flows suffer from strict RTT spikes.

• There is some evidence that non-

strict RTT spikes are much more common.

• An RTT spike for an EDGE user is

not a spike for a GSM user!

VTT Information Technology

ISMA 2004 Workshop on Internet Signal Processing 7

Next steps: (Work in progress and still in early phases)

• Practical issues:

– How to get high-quality mobile TCP data. (Subscribers’ privacy, operator’s business secrets, organization, technical issues et cetera).

• Theoretical issues:

– Statistically useful definition of an RTT spike? (Unnecessary retransmission may identify a strict RTT spike but non-strict spikes are also of importance.) – Homogeneous classes according to the link characteristics would be required for more fruitful statistical inference! – Robustness and minimum requirements of the Radon transform method?

VTT Information Technology