Manual: TCP Agents

From nsnam
Revision as of 19:14, 23 July 2008 by Lachlan (Talk | contribs) (One-Way Trace TCP Trace Dynamics: Renamed "two way")

Jump to: navigation, search

Previous Chapter "UDP Agents" | Index | Next Chapter "SCTP Agents"

This section describes the operation of the TCP agents in ns. There are two major types of TCP agents: one-way agents and a two-way agent. One-way agents are further subdivided into a set of TCP senders (which obey different congestion and error control techniques) and receivers (“sinks”). The two-way agent is symmetric in the sense that it represents both a sender and receiver. It is still under development.

The files described in this section are too numerous to enumerate here. Basically it covers most files matching the regular expression ~ns/tcp*.{cc, h}.

The one-way TCP sending agents currently supported are:

  • Agent/TCP - a “tahoe” TCP sender
  • Agent/TCP/Reno - a “Reno” TCP sender
  • Agent/TCP/Newreno - Reno with a modification
  • Agent/TCP/Sack1 - TCP with selective repeat (follows RFC2018)
  • Agent/TCP/Vegas - TCP Vegas
  • Agent/TCP/Fack - Reno TCP with “forward acknowledgment”
  • Agent/TCP/Linux - a TCP sender with SACK support that runs TCP congestion control modules from Linux kernel

The one-way TCP receiving agents currently supported are:

  • Agent/TCPSink - TCP sink with one ACK per packet
  • Agent/TCPSink/DelAck - TCP sink with configurable delay per ACK
  • Agent/TCPSink/Sack1 - selective ACK sink (follows RFC2018)
  • Agent/TCPSink/Sack1/DelAck - Sack1 with DelAck

The two-way experimental sender currently supports only a Reno form of TCP:

  • Agent/TCP/FullTcp

The section comprises three parts: the first part is a simple overview and example of configuring the base TCP send/sink agents (the sink requires no configuration). The second part describes the internals of the base send agent, and last part is a description of the extensions for the other types of agents that have been included in the simulator.

One-Way TCP Senders

The simulator supports several versions of an abstracted TCP sender. These objects attempt to capture the essence of the TCP congestion and error control behaviors, but are not intended to be faithful replicas of real-world TCP implementations. They do not contain a dynamic window advertisement, they do segment number and ACK number computations entirely in packet units, there is no SYN/FIN connection establishment/teardown, and no data is ever transferred (e.g. no checksums or urgent data).

The Base TCP Sender (Tahoe TCP)

The “Tahoe” TCP agent Agent/TCP performs congestion control and round-trip-time estimation in a way similar to the version of TCP released with the 4.3BSD “Tahoe” UN’X system release from UC Berkeley. The congestion window is increased by one packet per new ACK received during slow-start (when cwnd_ < ssthresh_) and is increased by 1/cwnd_ for each new ACK received during congestion avoidance (when cwnd_ ≥ ssthresh_).

Responses to Congestion Tahoe TCP assumes a packet has been lost (due to congestion) when it observes NUMDUPACKS (defined in tcp.h, currently 3) duplicate ACKs, or when a retransmission timer expires. In either case, Tahoe TCP reacts by setting ssthresh_ to half of the current window size (the minimum of cwnd_ and window_) or 2, whichever is larger. It then initializes cwnd_ back to the value of windowInit_. This will typically cause the TCP to enter slow-start.

Round-Trip Time Estimation and RTO Timeout Selection Four variables are used to estimate the round-trip time and set the retransmission timer: rtt_, srtt_, rttvar_, tcpTick_, and backoff_. TCP initializes rttvar to 3/tcpT ick_ and backoff to 1. When any future retransmission timer is set, it’s timeout is set to the current time plus max(bt(a + 4v + 1), 64) seconds, where b is the current backoff value, t is the value of tcpTick, a is the value of srtt, and v is the value of rttvar.

Round-trip time samples arrive with new ACKs. The RTT sample is computed as the difference between the current time and a “time echo” field in the ACK packet. When the first sample is taken, its value is used as the initial value for srtt_. Half the first sample is used as the initial value for rttvar_. For subsequent samples, the values are updated as follows:

srtt = 7*srtt/8 + sample/8
rttvar = 3*rttvar/4 + |sample − srtt|/4

Configuration

Running an TCP simulation requires creating and configuring the agent, attaching an application-level data source (a traffic generator), and starting the agent and the traffic generator.

Simple Configuration

Creating the Agent

set ns [new Simulator] ;# preamble initialization
set node1 [$ns node] ;# agent to reside on this node
set node2 [$ns node] ;# agent to reside on this node

set tcp1 [$ns create-connection TCP $node1 TCPSink $node2 42]
$tcp set window_ 50 ;# configure the TCP agent

set ftp1 [new Application/FTP]
$ftp1 attach-agent $tcp1

$ns at 0.0 "$ftp start"

This example illustrates the use of the simulator built-in function create-connection. The arguments to this function are: the source agent to create, the source node, the target agent to create, the target node, and the flow ID to be used on the connection. The function operates by creating the two agents, setting the flow ID fields in the agents, attaching the source and target agents to their respective nodes, and finally connecting the agents (i.e. setting appropriate source and destination addresses and ports). The return value of the function is the name of the source agent created.

TCP Data Source

The TCP agent does not generate any application data on its own; instead, the simulation user can connect any traffic generation module to the TCP agent to generate data. Two applications are commonly used for TCP: FTP and Telnet. FTP represents a bulk data transfer of large size, and telnet chooses its transfer sizes randomly from tcplib (see the file tcplib-telnet.cc). Details on configuring these application source objects are in the chapter on applications.

Other Configuration Parameters

In addition to the window_ parameter listed above, the TCP agent supports additional configuration variables. Each of the variables described in this subsection is both a class variable and an instance variable. Changing the class variable changes the default value for all agents that are created subsequently. Changing the instance variable of a particular agent only affects the values used by that agent. For example,

Agent/TCP set window_ 100 ;# Changes the class variable
$tcp set window_ 2.0 ;# Changes window_ for the $tcp object only

The default parameters for each TCP agent are:

Agent/TCP set window_ 20 ;# max bound on window size
Agent/TCP set windowInit_ 1 ;# initial/reset value of cwnd
Agent/TCP set windowOption_ 1 ;# cong avoid algorithm (1: standard)
Agent/TCP set windowConstant_ 4 ;# used only when windowOption != 1
Agent/TCP set windowThresh_ 0.002 ;# used in computing averaged window
Agent/TCP set overhead_ 0 ;# !=0 adds random time between sends
Agent/TCP set ecn_ 0 ;# TCP should react to ecn bit
Agent/TCP set packetSize_ 1000 ;# packet size used by sender (bytes)
Agent/TCP set bugFix_ true ;# see explanation
Agent/TCP set slow_start_restart_ true ;# see explanation
Agent/TCP set tcpTick_ 0.1 ;# timer granulatiry in sec (.1 is NONSTANDARD)
Agent/TCP set maxrto_ 64 ;# bound on RTO (seconds)
Agent/TCP set dupacks_ 0 ;# duplicate ACK counter
Agent/TCP set ack_ 0 ;# highest ACK received
Agent/TCP set cwnd_ 0 ;# congestion window (packets)
Agent/TCP set awnd_ 0 ;# averaged cwnd (experimental)
Agent/TCP set ssthresh_ 0 ;# slow-stat threshold (packets)
Agent/TCP set rtt_ 0 ;# rtt sample
Agent/TCP set srtt_ 0 ;# smoothed (averaged) rtt
Agent/TCP set rttvar_ 0 ;# mean deviation of rtt samples
Agent/TCP set backoff_ 0 ;# current RTO backoff factor
Agent/TCP set maxseq_ 0 ;# max (packet) seq number sent

For many simulations, few of the configuration parameters are likely to require modification. The more commonly modified parameters include: window_ and packetSize_. The first of these bounds the window TCP uses, and is considered to play the role of the receiver’s advertised window in real-world TCP (although it remains constant). The packet size essentially functions like the MSS size in real-world TCP. Changes to these parameters can have a profound effect on the behavior of TCP. Generally, those TCPs with larger packet sizes, bigger windows, and smaller round trip times (a result of the topology and congestion) are more agressive in acquiring network bandwidth.

Other One-Way TCP Senders

Reno TCP The Reno TCP agent is very similar to the Tahoe TCP agent, except it also includes fast recovery, where the current congestion window is “inflated” by the number of duplicate ACKs the TCP sender has received before receiving a new ACK. A “new ACK” refers to any ACK with a value higher than the higest seen so far. In addition, the Reno TCP agent does not return to slow-start during a fast retransmit. Rather, it reduces sets the congestion window to half the current window and resets ssthresh_ to match this value.

Newreno TCP This agent is based on the Reno TCP agent, but which modifies the action taken when receiving new ACKS. In order to exit fast recovery, the sender must receive an ACK for the highest sequence number sent. Thus, new “partial ACKs” (those which represent new ACKs but do not represent an ACK for all outstanding data) do not deflate the window (and possibly lead to a stall, characteristic of Reno).

Vegas TCP This agent implements “Vegas” TCP ([4, 5]). It was contributed by Ted Kuo.

Sack TCP This agent implements selective repeat, based on selective ACKs provided by the receiver. It follows the ACK scheme described in [23], and was developed with Matt Mathis and Jamshid Mahdavi.

Fack TCP This agent implements “forward ACK” TCP, a modification of Sack TCP described in [22].

Linux TCP This agent runs TCP congestion control modules imported from Linux kernel. The agent generates simulation results that are consistent, in congestion window trajectory level, with the behavior of Linux hosts.

Simulation users can update or import new congestion control modules from Linux kernel source code for this agent. The Linux congestion control modules are compiled into the NS-2 binary. Users can select different congestion control algorithms, different congestion controlmodule parameters, and different Linux TCP parameters for different instances of this agent. This agent supports SACK. A receiver that supports SACK is recommended to work with this agent. There is a tutorial for using this agent in [32].

The implementation of this agent loosely follows the Linux TCP packet processing routine and calls the congestion control source codes from Linux kernel to change congestion control related parameters (e.g. congestion window, slow start threshold and etc). The design and implementation details are described in [33].

To achieve simulation results close to Linux performance, this agent changes the default values of the following parameters according to the Linux parameters:

Agent/TCP/Linux set maxrto_ 120
Agent/TCP/Linux set ts_resetRTO_ true
Agent/TCP/Linux set delay_growth_ false

TCP Receivers (sinks)

The TCP senders described above represent one-way data senders. They must peer with a “TCP sink” object.

The Base TCP Sink

The base TCP sink object (Agent/TCPSink) is responsible for returning ACKs to a peer TCP source object. It generates one ACK per packet received. The size of the ACKs may be configured. The creation and configuration of the TCP sink object is generally performed automatically by a library call (see create-connection above).

configuration parameters

Agent/TCPSink set packetSize_ 40

Delayed-ACK TCP Sink

A delayed-ACK sink object (Agent/Agent/TCPSink/DelAck) is available for simulating a TCP receiver that ACKs less than once per packet received. This object contains a bound variable interval_ which gives the number of seconds to wait between ACKs. The delayed ACK sink implements an agressive ACK policy whereby only ACKs for in-order packets are delayed. Out-of-order packets cause immediate ACK generation.

configuration parameters

Agent/TCPSink/DelAck set interval_ 100ms

Sack TCP Sink

The selective-acknowledgment TCP sink (Agent/TCPSink/Sack1) implements SACK generation modeled after the description of SACK in RFC 2018. This object includes a bound variable maxSackBlocks_ which gives the maximum number of blocks of information in an ACK available for holding SACK information. The default value for this variable is 3, in accordance with the expected use of SACK with RTTM (see RFC 2018, section 3). Delayed and selective ACKs together are implemented by an object of type Agent/TCPSink/Sack1/DelAck.

configuration parameters

Agent/TCPSink set maxSackBlocks_ 3

Two-Way TCP Agents (FullTcp)

The Agent/TCP/FullTcp object is a new addition to the suite of TCP agents supported in the simulator and is still under development. It is different from (and incompatible with) the other agents, but does use some of the same architecture. It differs from these agents in the following ways: following ways:

  • connections may be establised and town down (SYN/FIN packets are exchanged)
  • bidirectional data transfer is supported
  • sequence numbers are in bytes rather than packets

The generation of SYN packets (and their ACKs) can be of critical importance in trying to model real-world behavior when using many very short data transfers. This version of TCP currently defaults to sending data on the 3rd segment of an initial 3-way handshake, a behavior somewhat different than common real-world TCP implementations. A “typical” TCP connection proceedswith an active opener sending a SYN, the passive opener respondingwith a SYN+ACK, the active opener responding with an ACK, and then some time later sending the first segment with data (corresponding to the first application write). Thus, this version of TCP sends data at a time somewhat earlier than typical implementations. This TCP can also be configured to send data on the initial SYN segment. Future changes to FullTCP may include a modification to send the first data segment later, and possibly to implement T/TCP functionality.

Currently FullTCP is only implemented with Reno congestion control, but ultimately it should be available with the full range of congestion control algorithms (e.g., Tahoe, SACK, Vegas, etc.).

Simple Configuration

Running an Full TCP simulation requires creating and configuring the agent, attaching an application-level data source (a traffic generator), and starting the agent and the traffic generator.

Creating the Agent

# set up connection (do not use "create-connection" method because
# we need a handle on the sink object)
set src [new Agent/TCP/FullTcp] ;   # create agent
set sink [new Agent/TCP/FullTcp] ;  # create agent
$ns_ attach-agent $node_(s1) $src ; # bind src to node
$ns_ attach-agent $node_(k1) $sink ;# bind sink to node
$src set fid_ 0 ;                   # set flow ID field
$sink set fid_ 0 ;                  # set flow ID field
$ns_ connect $src $sink ;           # active connection src to sink

# set up TCP-level connections
$sink listen ;                      # will figure out who its peer is
$src set window_ 100;

The creation of the FullTcp agent is similar to the other agents, but the sink is placed in a listening state by the listen method. Because a handle to the receiving side is required in order to make this call, the create-connection call used above cannot be used.

Configuration Parameters The following configuration parameters are available through Tcl for the FullTcp agent:

Agent/TCP/FullTcp set segsperack_ 1 ;      # segs received before generating ACK
Agent/TCP/FullTcp set segsize_ 536 ;       # segment size (MSS size for bulk xfers)
Agent/TCP/FullTcp set tcprexmtthresh_ 3 ;  # dupACKs thresh to trigger fast rexmt
Agent/TCP/FullTcp set iss_ 0 ;             # initial send sequence number
Agent/TCP/FullTcp set nodelay_ false ;     # disable sender-side Nagle algorithm
Agent/TCP/FullTcp set data_on_syn_ false ; # send data on initial SYN?
Agent/TCP/FullTcp set dupseg_fix_ true ;   # avoid fast rxt due to dup segs+acks
Agent/TCP/FullTcp set dupack_reset_ false ;# reset dupACK ctr on !0 len data segs containing dup ACKs
Agent/TCP/FullTcp set interval_ 0.1 ;      # as in TCP above, (100ms is non-std)

BayFullTcp

A different implementation of two-way TCP has been ported into ns from Kathy Nicholes/Van Jacobson’s group. It is called BayFullTcp. The basic difference between BayFullTcp and FullTcp (the two-way tcp version already present in ns) are as follows:

  • BayTcp supports a client-server application model while FullTcp makes no assumption about its application layer.
  • The tcp-application interface is different for both;
  • FullTcp supports partial ack (BayTcp doesn’t).
  • FullTcp supports different flavors of tcp (tahoe, reno etc) which is not the case for baytcp.
  • Both implementations have different set of API’s .

There might be other finer differences between the two as well. One of our future plans is to redefine the APIs to allow fulltcp to use baytcp’s client-server model.

Architecture and Internals

The base TCP agent (class Agent/TCP) is constructed as a collection of routines for sending packets, processing ACKs, managing the send window, and handling timeouts. Generally, each of these routines may be over-ridden by a function with the same name in a derived class (this is how many of the TCP sender variants are implemented).

The TCP header The TCP header is defined by the hdr_tcp structure in the file ~ns/tcp.h. The base agent only makes use of the following subset of the fields:

ts_ /* current time packet was sent from source */
ts_echo_ /* for ACKs: timestamp field from packet associated with this ACK */
seqno_ /* sequence number for this data segment or ACK (Note: overloading!) */
reason_ /* set by sender when (re)transmitting to trace reason for send */

Functions for Sending Data Note that generally the sending TCP never actually sends data (it only sets the packet size).

send_much(force, reason, maxburst) - this function attempts to send as many packets as the current sent window allows. It also keeps track of how many packets it has sent, and limits to the total to maxburst. The function output(seqno, reason) sends one packet with the given sequence number and updates the maximum sent sequence number variable (maxseq_) to hold the given sequence number if it is the greatest sent so far. This function also assigns the various fields in the TCP header (sequence number, timestamp, reason for transmission). This function also sets a retransmission timer if one is not already pending.

Functions for Window Management The usable send window at any time is given by the function window(). It returns the minimum of the congestion window and the variable wnd_, which represents the receiver’s advertised window.

opencwnd() - this function opens the congestion window. It is invoked when a new ACK arrives. When in slow-start, the function merely increments cwnd_ by each ACK received. When in congestion avoidance, the standard configuration increments cwnd_ by its reciprocal. Other window growth options are supported during congestion avoidance, but they are experimental (and not documented; contact Sally Floyd for details).

closecwnd(int how) - this function reduces the congestion window. It may be invoked in several ways: when entering fast retransmit, due to a timer expiration, or due to a congestion notification (ECN bit set). Its argument how indicates how the congestion window should be reduced. The value 0 is used for retransmission timeouts and fast retransmit in Tahoe TCP. It typically causes the TCP to enter slow-start and reduce ssthresh_ to half the current window. The value 1 is used by Reno TCP for implementing fast recovery (which avoids returning to slow-start). The value 2 is used for reducing the window due to an ECN indication. It resets the congestion window to its initial value (usually causing slow-start), but does not alter ssthresh_.

Functions for Processing ACKs recv() - this function is the main reception path for ACKs. Note that because only one direction of data flow is in use, this function should only ever be invoked with a pure ACK packet (i.e. no data). The function stores the timestamp from the ACK in ts_peer_, and checks for the presence of the ECN bit (reducing the send window if appropriate). If the ACK is a new ACK, it calls newack(), and otherwise checks to see if it is a duplicate of the last ACK seen. If so, it enters fast retransmit by closing the window, resetting the retransmission timer, and sending a packet by calling send_much.

newack() - this function processes a “new” ACK (one that contains an ACK number higher than any seen so far). The function sets a new retransmission timer by calling newtimer(), updates the RTT estimation by calling rtt_update, and updates the highest and last ACK variables.

Functions for Managing the Retransmission Timer These functions serve two purposes: estimating the round-trip time and setting the actual retransmission timer. rtt_init - this function initializes srtt_ and rtt_ to zero, sets rttvar_ to 3/tcp_tick_, and sets the backoff multiplier to one.

rtt_timeout - this function gives the timeout value in seconds that should be used to schedule the next retransmission timer. It computes this based on the current estimates of the mean and deviation of the round-trip time. In addition, it implements Karn’s exponential timer backoff for multiple consecutive retransmission timeouts.

rtt_update - this function takes as argument themeasured RTT and averages it in to the runningmean and deviation estimators according to the description above. Note that t_srtt_ and t_rttvar are both stored in fixed-point (integers). They have 3 and 2 bits, respectively, to the right of the binary point.

reset_rtx_timer - This function is invoked during fast retransmit or during a timeout. It sets a retransmission timer by calling set_rtx_timer and if invoked by a timeout also calls rtt_backoff.

rtt_backoff - this function backs off the retransmission timer (by doubling it).

newtimer - this function called only when a new ACK arrives. If the sender’s left window edge is beyond the ACK, then set_rtx_timer is called, otherwise if a retransmission timer is pending it is cancelled.

Tracing TCP Dynamics

The behavior of TCP is often observed by constructing a sequence number-vs-time plot. Typically, a trace is performed by enabling tracing on a link over which the TCP packets will pass. Two trace methods are supported: the default one (used for tracing TCP agents), and an extension used only for FullTCP.

One-Way Trace TCP Trace Dynamics

TCP packets generated by one of the one-way TCP agents and destined for a TCP sink agent passing over a traced link will generate a trace file lines of the form:

+ 0.94176 2 3 tcp 1000 ------ 0 0.0 3.0 25 40
+ 0.94276 2 3 tcp 1000 ------ 0 0.0 3.0 26 41
d 0.94276 2 3 tcp 1000 ------ 0 0.0 3.0 26 41
+ 0.95072 2 0 ack 40 ------ 0 3.0 0.0 14 29
- 0.95072 2 0 ack 40 ------ 0 3.0 0.0 14 29
- 0.95176 2 3 tcp 1000 ------ 0 0.0 3.0 21 36
+ 0.95176 2 3 tcp 1000 ------ 0 0.0 3.0 27 42

The exact format of this trace file is given in Chapter: "Trace and Monitoring Support", Section "Trace File Format". When tracing TCP, packets of type tcp or ack are relevant. Their type, size, sequence number (ack number for ack packets), and arrival/depart/drop time are given by field positions 5, 6, 11, and 2, respectively. The + indicates a packet arrival, d a drop, and - a departure. A number of scripts process this file to produce graphical output or statistical summaries (see, for example, ~ns/test-suite.tcl), the finish procedure.

Two-Way Trace TCP Trace Dynamics

TCP packets generated by FullTcp and passing over a traced link contain additional information not displayed by default using the regular trace object. By enabling the flag show_tcphdr_ on the trace object (see section refsec:traceformat), three additional header fields are written to the trace file: ack number, tcp-specific flags, and header length.

Commands at a glance

The following is a list of commands used to setup/manipulate TCP flows for simulations:

set tcp0 [new Agent/TCP]

This creates an instance of a TCP agent. There are several flavors of TCP-sender and TCP-receiver (or sink) agent currently implemented in ns. TCP-senders currently available are: Agent/TCP, Agent/TCP/Reno, Agent/TCP/Newreno, Agent/TCP/Sack1, Agent/TCP/Vegas, Agent/TCP/Fack.
TCP-receivers currently available are: Agent/TCPSink, Agent/TCPSink/DelAck, Agent/TCPSink/Sack1, Agent/TCPSink/Sack1/DelAck.
There is also a two-way implementation of tcp called Agent/TCP/FullTcp. For details on the different TCP flavors see earlier sections of this chapter.

Configuration parameters for TCP flows maybe set as follows:

$tcp set window_ <wnd-size>

For all possible configuration parameters available for TCP see Section "Other Configuration Parameters" of this chapter. The default configuration values can also be found in ns/tcl/lib/ns-default.tcl.

Following is an example of a simple TCP connection setup:

set tcp [new Agent/TCP] ;# create tcp agent
$ns_ attach-agent $node_(s1) $tcp ;# bind src to node
$tcp set fid_ 0 ;# set flow ID field
set ftp [new Application/FTP] ;# create ftp traffic
$ftp attach-agent $tcp ;# bind ftp traffic to tcp agent
set sink [new Agent/TCPSink] ;# create tcpsink agent
$ns_ attach-agent $node_(k1) $sink ;# bind sink to node
$sink set fid_ 0 ;# set flow ID field
$ns_ connect $ftp $sink ;# active connection src to sink
$ns_ at $start-time "$ftp start" ;# start ftp flow

For an example of setting up a full-tcp connection see Section "Simple Configuration".


Previous Chapter "UDP Agents" | Index | Next Chapter "SCTP Agents"