spain 03
DESCRIPTION
Spain 03TRANSCRIPT
Protocol Tracing in GPRS
Andrei Gurtov
6.03.2003Workshop on Internet Access over 2.5G3G Networks
Barcelona, Spain
2/21
Outline
Delay spikes in GPRS and their effect on TCPMulti-layer protocol tracing in GPRSTCP and TFRC during intersystem (vertical) handovers
3/21
Delay Spikes in GPRS during Handovers
A. Gurtov, Effect of Delays on TCP Performance, In Proceedings of IFIP Personal Wireless Communications (PWC'01), August 2001.
A. Gurtov, Making TCP Robust Against Delay Spikes, University of Helsinki, Department of Computer Science, No C-2001-53, November 2001.
4/21
Delay Spikes in GPRS during Handovers
02
468
10
1214
0 50 100 150 200 250 300 350
Ping sequence number
RTT
, sec
onds
5/21
Effect of Delay Spikes on TCP -R.Ludwig
• Delay spikes cause spurious
timeouts in TCP
• Unnecessary go-back-N
retransmissions
• Too aggressive congestion
control in short run,
underutilization in long run
• The Eifel algorithm detects a
spurious timeout and responds by
preventing go-back-N and CC 35000
40000
45000
50000
55000
41 46 51 56 61 66Time of Day (s)
Sequ
ence
Num
ber
Snd_DataSnd_AckRcv_DataRcv_Ackhiccup
(3)
(1)
(2)
TCP
TCP+Eifel
6/21
Papers on the Eifel Algorithm
• R. Ludwig and R. H. Katz. The Eifel algorithm: Making TCP robust against spurious retransmissions. ACM CCR 30(1), January 2000
• A. Gurtov, R. Ludwig, Responding to Spurious Timeouts in TCP, IEEE INFOCOM'03.
• A. Gurtov, R. Ludwig, Evaluating the Eifel Algorithm for TCP in a GPRS network, In Proceedings of European Wireless, February 2002.
7/21
IETF Documents on TCP Enhancements
H. Inamura, G. Montenegro, R. Ludwig, A. Gurtov, F. Khafizov, TCP over Second (2.5G) and Third (3G) Generation Wireless Networks, RFC3481.
R. Ludwig, A. Gurtov, The Eifel Response Algorithm for TCP, draft-ietf-tsvwg-tcp-eifel-response-03.txt , work in progress.
A. Gurtov, On Treating DUPACKs in TCP, draft-gurtov-tsvwg-tcp-delay-spikes-01.txt, work in progress.
8/21
Multi-Layer Tracing in GPRS
A. Gurtov, M. Passoja, O. Aalto, M. Raitola, Multi-Layer Protocol Tracing in a GPRS Network, In Proceedings of IEEE Vehicular Technology Conference (VTC'02), September 2002.
J. Korhonen, O. Aalto, A. Gurtov, H. Laamanen, Measured Performance of GSM HSCSD and GPRS, In Proceedings of the IEEE International Conference on Communications (ICC'01), June 2001.
9/21
Multi-Layer Tracing in GPRS
Stationary/mobile measurements in test/live networks=>Compare to earlier results presented in ICC´01=>Find out any remaining inefficienciesEnd-to-end TCP tracing (tcpdump)=>Throughput, RTT, goodput, network buffers, delaysLLC tracing at Gb interface (NetHawk)=>Fragmentation, mobility signallingRLC tracing at Abis interface (NetHawk)=>Retransmissions, radio channel allocation
10/21
Testbed
Abis
Abis
Mobile client tcpdump
BSC
Firewall
Fixed servertcpdump
BTS SGSNGGSN
BTS
BSC
Gb
Nethawk
11/21
Stationary TCP Throughput is Optimal
0
10
20
30
40
50
60
T 3 9 m R 5 2 0 m N 8 3 10 T 2 8 0
Thro
ughp
ut, k
pbs
M IN M A X A V G 9 0 % T C P M A X Line ra t e
0
10
20
30
40
50
60
T39m R520m N8310 T280
Thro
ughp
ut, k
pbs
MIN MAX AVG 90% TCP MAX Line rate
Downlink
Uplink
Very little space for TCPimprovements (perhaps header compression?)
12/21
Buffering of User Data in GPRS
Networks use huge drop-tail buffers of 50-150 KB=>’the more the better’ buffering approach in telecom?Terminals use buffers of 3KB-32KBProblems with too large buffer=>Large RTT (> 10 sec)=>Significant losses during routing area updates=>Delivering unnecessary data after aborting transfersProblems with too small buffers=>High loss rate reduces throughputOptimal buffer size is about 10 KB (2*bandwidth*delay)
13/21
Fast Reset for TCP
When the user clicks from a web page to another, TCP connections
from the old page are aborted
Buffered packets in the network are delivered unnecessary
TCP receiver generates RST for packets on aborted conn.
The idea of ’Fast Reset’
• Look for RST packets in the bottleneck buffer
• Delete packets that belong to aborted connections in other direction
Seems to provide substantial efficiency savings (e.g. 100%) without
affecting robustness
Seems to works for all applications using TCP, not just Web
14/21
Major Findings from Tracing
Packets have to wait for channel (TBF) allocation 80-200 ms• Extended TBF release [Wiemann et al, Ericsson]
Unnecessary retransmissions at the RLC layer• Control RLC SACK frequency and introduce a guard timer [Walke]
Inefficient fragmentation of TCP segments into LLC frames• Set larger frame size
Tracing of cell reselections: duration, frequency & loss• Duration 3-6s; most packets get lost in downlink, few in uplink
GPRS performance got better in two years of time• More stability, higher throughput, lower RTT
15/21
RLC Trace of Bulk TCP Transfer
0
20
40
60
80
100
120
5.0 5.5 6.0 6.5 7.0 7.5 8.0 8.5 9.0 9.5 10.0 10.5 11.0 11.5 12.0Time, s
RLC
sequ
ence
num
ber
DL RLC data block
DL RLC ACK
UL RLC data block
Uplink Channel Assignment
UL Channel Request
TCP ACKs
TBF alloc. delay 200ms
TBF upkeeping
Unnecessary RLCretransmissions
TCP segments
16/21
LLC/TCP Trace of Cell Reselection
0
50
100
150
200
250
300
350
80 85 90 95 100 105 110 115 120 125 130
Time, s
LLC
fram
e nu
mbe
r
0
10000
20000
30000
40000
50000
60000
70000
80000
90000
100000
TCP
sent
byt
es
LLC snd_data
TCP snd_data
TCP snd_ack
Cell reselections
TCP recovers with timeout and slow start
TCP recovers with fast retransmit (and spurious timeout)
17/21
Summary of Tracing Results
Max throughput in downlink 43 kbps, in uplink 21 kbps Unloaded RTT 0.6-0.8 s, under load 5-10 sIn general, smaller network buffers are neededExtended channel (TBF) release reduces RTTRLC fixes to avoid unnecessary data retransmissionsLarge frame size in UNACK LLC mode to avoid
inefficient fragmentation of TCP/IP packetsCell reselections are the main problem=>take 3-15 s, occur every minute while driving in a city=>cause packet losses in downlink, but rarely in uplink
18/21
Ongoing Work: Vertical Handovers
Handovers between GPRS, UMTS, WLAN, LANVertical handovers cause up to two orders of magnitude
change in link characteristics Trigger error losses, delay spikes, reordering and
duplicationSlowly-responsive congestion control (TFRC) provides
smooth rates for streamingHow TCP and TFRC behave and share bandwidth in
presence of vertical handovers?
19/21
Ongoing Work: Vertical Handovers
Mobile IP testbed with GPRS, UMTS, WLAN, LANTraces of TCP and TFRC over intersystem handoversFindings:
• TFRC is not always fair towards TCP
• ACK compression affects TFRC smoothness
• TFRC gets over 200 packets in network when 3 is enough
• As expected, TFRC overshoots and underutilizes bandwidth
during vertical handovers
20/21
Ongoing Work: Vertical Handovers
Continue experiments using an NS2 model of ideal handover
Concentrate on fundamental changes in link characteristics, not on avoidable disruptions such as delays
Two enhancement mechanisms are evaluated: overbuffering and an explicit handover notification
A Fast Reset algorithm is proposed to make overbuffering more feasible
21/21
Thank you!
More information on http://www.cs.helsinki.fi/~gurtov