Franco Davoli, Giuseppe Span

1

Performance Measurements and Comparison of
Modified TCP Control Algorithms over Rain-Faded
Satellite Channels

• Franco Davoli, Giuseppe Spanò,
• Stefano Vignola, Sandro Zappatore
• CNIT - Italian National Consortium for
  Telecommunications
• National Multimedia Communications Laboratory,
  Via Diocleziano 328, Napoli, Italy
• and
• Department of Communications, Computer and
  Systems Science (DIST)
• University of Genoa, Via Opera Pia 13, Genova,
  Italy
• (franco.davoli, giuseppe.spano, stefano.vignola,
  sandro.zappatore)_at_cnit.it

2
Introduction and Objectives

• Quite a few TCP modifications have been proposed,
  analyzed and implemented, to cope with large
  bandwidth-delay product and error-prone networks,
  where the standard mechanisms of congestion
  control present problems.
• For many of such implementations, patches to the
  Linux kernel are available, which can be used to
  effectively configure the modified TCP and to use
  it in real world tests.
• However, in most cases the performance has been
  evaluated by simulation, and there are rather few
  works on measurements effected on real satellite
  networks or on emulated laboratory environments.
• The main goal of the paper is to compare some
  modified TCP congestion control algorithms in a
  laboratory setting, by performing repeatable
  experiments, involving traffic generators,
  routers, and an accurate satellite channel
  emulation with recorded fading data.

3
Laboratory setting
4
Laboratory setting (contd)

• We have emulated the satellite link by using the
  facilities offered by the CNIT National
  Laboratory for Multimedia Communications in
  Naples, Italy.
• The LAN consists of a set of hosts devoted to
  generate TCP traffic. According to the specific
  test, up to ten connections can be running at the
  same time.
• All packets produced feed the router 1 (CISCO
  3640), which conveys them to a serial V.35
  interface, adopting a HDLC protocol at the data
  link layer. The synchronous data flow is input to
  a Fairchild SM290 modem (modem 1), operating in
  quadrature phase-shift keying (QPSK), at an
  information rate of 2048 kbit/s, with 3/4
  sequential FEC.
• The outgoing modulated signal is applied to a
  Channel Simulator (here used as a simple
  programmable attenuator) and then added to a
  Gaussian noise, produced by a noise generator.
  Both the attenuation and noise variance are
  suitably changed, to reproduce the effects of
  real world fading as concerns the bit error
  probability observed at the receiver end.

5
Laboratory setting (contd)

• The link budget was derived according to the
  Italsat satellite characteristics (G/T23 dB at
  the receiver, under the adopted parameter
  values).
• After a delay of 250 ms, the corrupted signal
  feeds the receiver portion of a second modem,
  which interfaces with the router 2.
• The return channel (i.e., the link connecting the
  outgoing signal from modem 2 to the input of
  modem 1) introduces again a delay of 250 ms,
  but it is not affected by errors.
• Fading attenuation values are taken from the
  results of the propagation experiment carried out
  in Ka band on the Olympus satellite by the CSTS
  (Centro Studi sulle Telecomunicazioni Spaziali)
  Institute, on behalf of the Italian Space Agency
  (ASI). The attenuation samples considered were
  1-second averages, expressed in dB, of the signal
  power attenuation with respect to clear sky
  conditions. The adopted patterns are taken from
  data recorded at the Spino dAdda (Northern
  Italy) station on September 21 and 23, 1992,
  respectively.

6
Fading Attenuation Patterns
The pattern indicated as B, which corresponds
to a quite severe fading, causes a temporary sync
loss at the demodulator
7
TCP modifications

• We have investigated the behaviour of 3 TCP
  modifications, and compared it with that of New
  Reno
• TCP Westwood
• TCP Peach
• Complete Knowledge Satellite Transport Protocol
  (CK-STP)
• The first two introduce modifications to the
  congestion control mechanism the third one only
  acts on the TCP New Reno parametrization, by
  assuming that the characteristics of the
  satellite bandwidth pipe are known at the
  transmitting station and losses are only due to
  channel errors.

8
TCP New Reno

• New Reno is an improved version of TCP Reno that
  avoids waiting for a retransmit timer when
  multiple segments are lost from a window.
• The change concerns the sender's behaviour during
  Fast Recovery, when a partial ACK is received
  that acknowledges some, but not all of the
  segments that were outstanding. In New Reno,
  partial ACKs received during Fast Recovery are
  treated as an indication that the segment
  immediately following the acknowledged one in the
  sequence space has been lost, and should be
  retransmitted.
• Thus, when multiple segments are lost from a
  single window of data, New-Reno can recover
  without a retransmission timeout, retransmitting
  one lost segment per round-trip time until all of
  the lost segments from that window have been
  retransmitted.

9
TCP Westwood
The main features of Westwood are the following

• End-to-end, sender-side estimate of the bandwidth
  B available to a TCP connection and seen at the
  receiver, obtained by measuring and low-pass
  filtering the rate of returning ACKs.
• When 3 DUPACKs are received
• ssthresh (B RTTmin) / seg_size
• cwnd ssthresh.
• When a coarse timeout expires
• ssthresh (B RTTmin) / seg_size
• cwnd 1.
• When ACKs are successfully received, TCPW
  increases cwnd according to Reno's congestion
  control algorithm.

10
TCP Peach

• TCP Peach implements Congestion Avoidance and
  Fast Retransmit (the same as Reno), and two new
  algorithms Rapid Recovery and Sudden Start,
  which are based on the use of dummy segments, low
  priority segments generated by the sender as a
  copy of the last transmitted data packet.
• If a router is congested, it discards dummy
  segments first, due to their low priority. On the
  contrary, the sender interprets the ACKs for
  dummy segments as the evidence that there are
  unused resources on the link and, accordingly, it
  can increase the transmission rate.

11
Complete Knowledge Satellite Transport Protocol
(CK-STP)
Complete Knowledge Satellite Transport Protocol
(CK-STP)

• It follows the TCP New Reno implementation, with
  the adoption of the SACK option. "Complete
  Knowledge" refers to the assumed knowledge about
  the physical characteristics of the satellite
  link, in terms of propagation delay and channel
  bandwidth.
• It assumes that all losses over the satellite
  link are due exclusively to link errors. A
  constant transmission window set to the
  bandwidth-delay product is employed, in order to
  achieve the maximum channel bandwidth.
• With respect to the TCP implementation, the slow
  start phase is no longer entered, and for each
  received acknowledgement, the congestion window
  keeps its constant value. As far as the loss
  detection is concerned, the arrival of only one
  duplicate acknowledgement is necessary to trigger
  the recovery phase, which consists of
  retransmitting the lost data segments. Once the
  recovery phase has terminated, the protocol
  continues transmitting with the constant
  transmission window as set previously.
• The sender buffer is adjusted according to the
  bandwidth delay-product.

12
Complete Knowledge Satellite Transport Protocol
(CK-STP)
Experimental measurements

• We have used the original fading patterns over a
  longer time window (900 s), by replicating each
  sample 3 times, in order to have a longer time
  scale for observing the TCP reaction. Fading
  pattern A is milder, whereas B is quite severe,
  and produces a loss of receiver synchronization
  for over 100 seconds, starting at 400 s.
• All graphs are obtained by averaging the results
  of 5 repeated experiments (under the same fading
  pattern), in order to increase their statistical
  significance.
• In all kernel implementations adopted for the
  various protocols (except CK-STP) the sender
  buffer dimension has been left untouched, with
  respect to the native New Reno implementation for
  the Linux Kernel 2.4.18 (64 kbytes).
• The sender buffer is set at 320 kbytes for STP
  (independently of the number of connections).

13
Throughput (goodput) vs. time of the different
TCP implementations with only one connection
active over a noisy satellite link (fading
pattern A)
14
Aggregate throughput vs. time of the different
TCP implementations with 2 connections active
over a noisy satellite link (fading pattern A).
15
Aggregate throughput vs. time of the different
TCP implementations with 5 connections active
over a noisy satellite link (fading pattern A).
16
Throughput vs. time of the different TCP
implementations with only one connection active
over a noisy satellite link(fading pattern B).
17
Aggregate throughput vs. time of the different
TCP implementations with 2 connections active
over a noisy satellite link (fading pattern B).
18
Aggregate throughput vs. time of the different
TCP implementations with 5 connections active
over a noisy satellite link (fading pattern B).
19
Throughput vs. time of three connections out of
5, measured during the realization of a single
experiment with STP,under fading pattern A.
20
Throughput vs. time of three connections out of
5, measured during the realization of a single
experiment with Westwood, under fading pattern A.
21
Conclusions and future work

• When a single connection is active, CK-STP
  achieves the best performance. In this sense, it
  would be fruitfully employed within PEPs for the
  creation of a bandwidth pipe.
• In terms of fairness and aggregate throughput in
  different situations, TCP Westwood (which, by the
  way, is becoming a standard feature of recent
  Linux kernels) appears to offer the best
  performance.
• In the presence of multiple connections, the
  effect of TCP modifications becomes less evident,
  and the advantage of using modified algorithms
  may be hindered by the necessity of updating the
  operating systems kernel of all the machines
  involved in the communication.
• The measurements are currently being repeated
  under the HB6 Ka payload link budget some
  measurements will also be performed on the Ka/Ku
  band CNIT satellite network. Software tools have
  been already developed to make the laboratory
  test-bed remotely accessible as a web service.