Experimental Analysis of TCP Spurious Retransmission Timeout in 3G3.5G networks

1
Experimental Analysis of TCP Spurious
Retransmission Time-out in 3G/3.5G networks
Gennaro Boggia, Antonio Barbuzzi Assistant
Professor, PhD student Telematics Lab -
Politecnico di Bari
Paolo Dini Research Associate IP Technologies
Area - CTTC
2
Outline

• Introduction to COST TMA (Data Traffic Monitoring
  and Analysis)
• Collaboration between Poliba and CTTC
• Presentation of Polibas Telematics lab
• Statement of the Spurious Retransmission Time-Out
  problem in TCP connections
• Testbed design and development
• Experimental result analysis
• Conclusions and future work

3
Outline

• Introduction to COST TMA (Data Traffic Monitoring
  and Analysis)
• Collaboration between Poliba and CTTC
• Presentation of Polibas Telematics lab
• Statement of the Spurious Retransmission Time-Out
  problem in TCP connections
• Testbed design and development
• Experimental result analysis
• Conclusions and future work

4
COST Action IC0703 Data Traffic Monitoring and
Analysis

• Understanding, developing and managing modern
  packet networks is difficult and expensive
• Traffic monitoring and analysis has always been
  seen as a key methodology to understand
  telecommunication technology and operation
• TMA COST Action aims at coordinating and
  promoting the development of common and novel
  monitoring tools and analysis platforms, so as to
  catalyze the emergence of a European de-facto
  standard for traffic monitoring

5
COST Action IC0703 Data Traffic Monitoring and
Analysis

• Collaboration between IPTech (CTTC) and
  Telematics Lab (Poliba)
• Experimental study of TCP over 3G/3.5G network
• First identified problem Spurious Retransmission
  Time-Out (SRTO)
• Short Term Scientific Mission (STSM)
• Antonio Barbuzzi
• - Title Measurements of TCP performances over
  3/3.5G wireless networks- Date 11/01/2009 -
  28/02/2009 (extended to 13/03/2009)- Home
  institution Politecnico di Bari, Italy- Host
  institution Centre Tecnològic de
  Telecomunicacions de Catalunya (CTTC), Spain

6
Telematics Lab - General Description

• Telematics Lab is a research laboratory at the
  Electrical Electronics Engineering Department
  (DEE) of Politecnico di Bari, the Technical
  University of Bari.
• Its mission is the research on the most relevant
  technologies in the area of telecommunication
  networks.

7
Main Research Areas

• Multimedia Systems
• Multimedia Streaming using quality adaptive
  encoding schemes
• Bandwidth estimation algorithms for Admission
  Control
• QoS in wireless LAN/PAN
• Feedback-based bandwidth allocation algorithms in
  wireless networks (IEEE 802.11, 802.15.3, etc.)
• Active and Passive measurements in 3G
• measurements to infer parameter settings in 3G
  network
• performance analysis
• Wireless Sensor Networks for detecting adverse
  events
• energy efficient architectures for event
  detection
• analytical modeling
• Networked Control Systems
• Architectures for real-time communications in
  factory environment
• Wireless network for system control

8
People
Full Professor
PhD Students
Pietro Camarda
Antonio Barbuzzi
Francesco Capozzi
Claudia Cormio
Assistant Professor
Gennaro Boggia
Rossella Fortuna
Luigi Alfredo Grieco
Roberta Laraspata
Maria Rita Palattella
Domenico Striccoli
Carla Passiatore
Post-Doc
Giuseppe Piro
Roberto DellAquila
Visiting Researcher
Giammarco Zacheo
Alessandro DAlconzo
9
International Cooperation

• FTW at Vienna (Austria)
• Nokia Siemens Network at Aalborg (Denmark)
• CTTC at Barcelona (Spain)
• VTT at Oulu (Finland)
• INRIA at Sophia Antipolis (France)

10
Some projects

• Some recent projects (most of them funded by
  Apulia Region)
• COST Action IC0703, Data Traffic Monitoring and
  Analysis theory, techniques, tools and
  applications for the future networks,. Chair F.
  Ricciato, University of Salento, FTW (Vienna),
  2007-2012.
• Apulia Regional Strategic Proj., PS 121 -
  Telecommunication Facilities and Wireless Sensor
  Networks in Emergency Management, Chair Prof.
  B. Maione, Politecnico di Bari, 2006-2009.
• Apulia Reg. Strategic Proj., PS 092 -
  Distributed Production to Innovative System
  DIPIS, Chiar Prof. G. Visaggio, University of
  Bari, 2006-2009.
• Apulian Reg. Operative Prog. 2000-2006,
  Monitoring and Adaptive Control Mobility of
  dangerous material, Chair Prof. G. Visaggio,
  University di Bari, 2007-2008.
• Apulian Reg. Operative Prog. 2000-2006,
  Terrestrial Digital Platform for Television
  Services with high Social Impact, Actuator
  CO.S.TE, 2007-2008.
• Apulia Reg. Explorative Proj., ICT Technologies
  for tracking food farming with RFID tags, Chair
  Prof. P. Camarda, Politecnico di Bari, 2007.
• Apulia Reg. Explorative Proj., ICT Technologies
  for tourist assistance based on an interactive
  virtual guide, Chair Prof. G. Piscitelli,
  Politecnico di Bari, 2007.
• Apulian Reg. Operative Prog. 2000-2006, Robotic
  Systems for Micro Assembly, in cooperation with
  Masmec S.r.l Italy. Chair. Prof. L. Salvatore,
  Politecnico di Bari, 2006-2007.
• Apulian Reg. Operative Prog. 2000-2006, Wireless
  Communication Systems for Industrial Automation,
  in cooperation with Masmec S.r.l Italy. Chair.
  Prof. P. Camarda, Politecnico di Bari, 2006-2007.

11
Outline

• Introduction to COST TMA (Data Traffic Monitoring
  and Analysis)
• Collaboration between Poliba and CTTC
• Presentation of Polibas Telematics lab
• Statement of the Spurious Retransmission Time-Out
  problem in TCP connections
• Testbed design and development
• Experimental result analysis
• Conclusions and future work

12
Recalling TCP behavior

• As well known, TCP
• uses a retransmission timer when expecting an
  acknowledgment (ACK) for a given segment
• the ACK should arrive before the RTO
  (Retransmission TimeOut)
• the RTO is evaluated as RTO SRTT 4 DEV
• where SRTT estimated Round Trip Time DEV RTT
  standard deviation

13
SRTO definition

• A spurious timeout happens in case of a sudden
  delay spike on the link, where the round-trip
  time exceeds the expected value calculated for
  the retransmission timeout.
• ACK is received after the retransmission of the
  segment

Receiver
Ack. Numb. 3000
Seq. Numb. 1500
Seq. Numb. 1
Seq. Numb. 1500
Ack. Numb. 1500
Sender
RTO
14
Effects of SRTO

• TCP retransmits the oldest outstanding segment
  (not lost, but delayed)
• The retransmission is unnecessary (the segment
  has been received)
• The sender interprets the received ACK as related
  to the retransmission
• TCP retransmits all outstanding segments (slow
  start algorithm)

• Retransmitted segments generate 3DUPACKs and a
  spurious fast retransmit
• Transmission of new segments are delayed

- A. Gurtov, R. Ludwig, Responding to Spurious
Timeouts in TCP, In Proc. of IEEE INFOCOM, Mar.
2003
15
Some remarks on SRTO

• Pronounced RTT variations can fire RTO even in
  absence of packet loss (i.e., a SRTO)
• Standard TCP flavors are not able to distinguish
  between SRTOs and normal RTOs due to packet
  losses
• An SRTO implies
• unnecessary retransmission of segments which
  already arrived successfully at the receiver
  beforehand
• unnecessary reduction of the congestion window
• several 3 DUPACKs with consequent congestion
  window reduction
• Frequent SRTOs can significantly degrades network
  performance and TCP throughput.

16
SRTO in 3G/3.5G network

• In a cellular network (3G/3.5G) a SRTO can be due
  to
• mobility the mobile terminal may experience
  handovers with consequent signaling and packet
  storage
• propagation a sudden change in radio conditions
  usually causing a spike in the RTT of stored
  packets (retransmissions at link layer)
• priority a sudden increase of high priority
  traffic (voice) reduces resources available for
  data traffic
• configuration some buffers in the core network
  can be overdimensioned
• cell state the mobile terminal switches to
  another channel

17
How to reveal a SRTO

• There are no algorithms (sender side) that
  reveal all possible SRTOs. To this aim, we need
  to look at both sender and receiver sides.
• D-SACK option
• duplicated segments are signaled
• extension of the well-known TCP SACK option
• the SRTO is revealed only when at reception of
  the first ACK of the retransmitted segment
• Eifel Algorithm
• it reveals SRTOs and can react to them
• TCP timestamp option is used
• it works if some ACKs are lost
• F-RTO
• no TCP options are required
• sometimes SRTOs are not revealed
• the segment transmission can go on without
  congestion window reduction

18
Outline

• Introduction to COST TMA (Data Traffic Monitoring
  and Analysis)
• Collaboration between Poliba and CTTC
• Presentation of Polibas Telematics lab
• Statement of the Spurious Retransmission Time-Out
  problem in TCP connections
• Testbed design and development
• Experimental result analysis
• Conclusions and future work

19
Testbed

• The mobile station and the wired host are on the
  same physical machine
• A single clock allows us to calculate the end to
  end delay (NTP would be an alternative)
• To simplify the realization of the testbed in
  different environments (we dont have always two
  machines)
• Elimination of synchronization issues

M
H
W
Internet
Mobile Station
Wired Host
20
Loopback Avoidance (I)

• Locally originated traffic with local IP address
  as destination would be delivered through the
  loopback interface
• The original addresses of each interface is
  changed to a fictitious one.
• Packets exiting from each interface need to have
  the real source address
• A rule in the NAT table in the POSTROUTING chain
  change source IP to the real IP
• Also a fictitious router is used for the Ethernet
  path
• Packets coming to the interfaces need to have the
  fictitious IP as destination
• A rule in the NAT table in PREROUTING chain
  changes the source IP to the fictitious one

21
Loopback Avoidance (II)

• The host needs to know the MAC address of the
  Ethernet router
• Static entry in the MAC table
• The Ethernet router needs to know the MAC address
  of the laptop
• We cant control this router ?
• Incoming ARP requests are modified to have
  fictitious IP addresses
• Outgoing ARP replies are modified to have real IP
  addresses
• The packets needs to flow through the internet
• Routing tables rule sent packets with real UMTS
  IP address destination through the Ethernet card
  and vice versa.
• Alternative solutions
• Raw sockets (but would require a userspace TCP
  implementation)
• The use of two different hosts

22
Loopback Avoidance (III)
ppp0
eth0
Fake Eth0 IP.src ? Real ppp0 IP.dst
Real Eth0 IP.src ? Fake ppp0 IP.dst
Fake ppp0 IP.src ? Real Eth0 IP.dst
Real ppp0 IP.src ? Fake Eth0 IP.dst
POSTROUTING
Kernel Space Outside
PREROUTING
POSTROUTING
PREROUTING
Real ppp0 IP.src ? Real Eth0 IP.dst
Real ppp0 IP.src ? Real Eth0 IP.dst
Real Eth0 IP.src ? Real ppp0 IP.dst
Real Eth0 IP.src ? Real ppp0 IP.dst
23
Access to TCP Internals States Web100 patch

• The Web100 software implements instruments in the
  Linux TCP/IP stack.
• kernel patch adding the instruments
• suite of "userland" libraries and tools for
  accessing the kernel instrumentation.
• we have developed a python script to save all the
  tcp parameters exposed by web100 in a
  asynchronous way
• Instantaneous values of Cwnd, SampleRTT, Smoothed
  RTT, RTO, etc
• Access to RTO events needs to be synchronous
• Patch in the TCP retransmit timer handler to
  record sequence number, wcid, timestamps of each
  RTO

24
Execution of the tests iperf

• The TCP flows are generated using Iperf
• Iperf is a commonly used network testing tool
  that can create TCP and UDP data streams and
  measure the throughput of a network that is
  carrying them (wikipedia)
• Causes for rate limitation
• Application (low rate or burst traffic
  generation)
• Limited Receiver Buffer (erroneous settings)
• Full Receiver Buffer (the application empties the
  buffer too slowly, because it has to write
  receiver data on the disk, or its engaged in
  other jobs)
• Network Limitation (bandwidth, packet loss)
• Using iperf our tests should be limited only by
  the network.

25
Some details

• UMTS Rel. 5
• Linux kernel 2.6.27-web100 kernel
• TCP Reno with default parameters
• Sack
• Dsack
• Timestamp Option
• Windows Scaling Option,
• cat /proc/sys/net/ipv4/tcp
• 1h TCP/IP Flow
• The traffic of each interface is dumped. To
  reduce discarded packets
• Snaplen limited to Layer2 header (14-16B) IP
  Header (20B) TCP header (20B) TCP Option
  Header (40B)

26
An algorithm for distinguish SRTOs and NRTOs

• Given the sequence number and the timestamp TRTO
  of a RTO, how to distinguish between NRTO e SRTO?
• Find IP.id of the segment that caused the RTO on
  the Sender Side (A)
• Find the correspondent segment (having the same
  TCP.seq and IP.id) on the receiver (B). If lost
  ? NRTO
• Find IP.id of the correspondent ACK on B
• Find the timestamp T5 of the received ACK on A.
  If lost ? NRTO
• If T5 gt TRTO ? SRTO else ? NRTO

27
Outline

• Introduction to COST TMA (Data Traffic Monitoring
  and Analysis)
• Collaboration between Poliba and CTTC
• Presentation of Polibas Telematics lab
• Statement of the Spurious Retransmission Time-Out
  problem in TCP connections
• Testbed design and development
• Analysis of the experimental results
• Conclusions and future work

28
Commercial Network - Download Case (I)
29
Commercial Network - Download Case (II)
30
Commercial Network - Upload Case (I)
31
Commercial Network - Upload Case (II)
32
Packet Losses
33
A Comparison of Different Network Operators
34
Outline

• Introduction to COST TMA (Data Traffic Monitoring
  and Analysis)
• Collaboration between Poliba and CTTC
• Presentation of Polibas Telematics lab
• Statement of the Spurious Retransmission Time-Out
  problem in TCP connections
• Testbed design and development
• Experimental result analysis
• Conclusions and future work

35
Conclusions and Future Work

• SRTO is only a problem in a congested cell (Bari
  experiment)
• Identification of the location of the congestion
• Radio interface
• Wired interfaces
• Radio access network
• Core network
• Network configuration problem or technology
  limitation?
• Use of extreMe cellUlar System Architeture (MUSA)
• High number of RTOs due to the UMTS uplink
• RLC retransmissions do not succeed in hiding
  channel losses to TCP
• Scarce utilization of uplink radio resource by
  TCP
• Behavior of other implementations of TCP
• Cubic
• Westwood

36
Download in MUSA
37
Upload in MUSA
38
Thanks for your kind attention!

• Questions?

Gennaro Boggia, Antonio Barbuzzi Assistant
Professor, PhD Student Telematics Lab Politecnico
di Bari boggia_at_poliba.it a.barbuzzi_at_poliba.it
Paolo Dini Research Associate IP
Technologies CTTC paolo.dini_at_cttc.es