The NS (v2) Sim ulator W orkshop br ought to you by Kevin - - PDF document
▶
SMART_READER_LITE
LIVE PREVIEW
The NS (v2) Sim ulator W orkshop br ought to you by Kevin - - PDF document
Jan 05, 2023 341 likes •923 views
The NS (v2) Sim ulator W orkshop br ought to you by Kevin F all La wrence Berk eley National Lab oratory kfal l@e e.lbl.gov http://www-nr g.e e.lbl.gov/kfal l AND Kannan V aradhan USC/ISI kannan@c atarina.usc.e du
SLIDE 1The NS (v2) Sim ulator W
rkshop
br
ught
to you by Kevin F all La wrence Berk eley National Lab
ratory
kfal l@e e.lbl.gov http://www-nr g.e e.lbl.gov/kfal l AND Kannan V aradhan USC/ISI kannan@c atarina.usc.e du http://www.isi.e du/~kannan Septem b er 18, 1997
SLIDE 2Audience and Outline
Audience
{ net w
rk
researc hers { educators { dev elop ers
T
pics
for to da y { VINT pro ject goals and status (Sally) { arc hitecture plus some history (Stev en) {
v
erview
f
ma jor comp
nen
ts (Kevin) { pro ject/co de status (Kevin) { details
f
ma jor comp
nen
ts (Kevin) { C++/OTcl link age and sim ulation debugging (Kannan) { top
logy
generation and session-lev el supp
rt
(Kannan) { m ulticast and reliable m ulticast (Kannan) { a complex link: CBQ (Kevin) { discussion and futures (Ev ery
ne)
NSv2 W
rkshop
kfal l@e e.lbl.gov Slide 2
SLIDE 3NSv2 Arc hitecture
Ob
ject-orien ted structure { ev
lution
from NSv1 (C++ with regular Tcl) {
b
jects implemen ted in C++ and \OTcl" { OTcl:
b
ject-orien ted extension to Tcl (from Da vid W etherall at MIT/LCS for V uSystem)
Con
trol/\Data" separation { con trol
p
erations in OTcl { data pass through C++
b
jects (for sp eed)
Mo
dular approac h { ne-grain
b
ject decomp
sition
{ p
sitives:
comp
sible,
re-usable { negatives: m ust \plum b" in OTcl, dev elop er m ust b e comfortable with b
th
en vironmen ts, to
ls
NSv2 W
rkshop
kfal l@e e.lbl.gov Slide 3
SLIDE 4Dev elopmen t Status
sim
ulator co de basis for VINT Pro ject
5ish
p eople activ ely con tributing to the co de base
ther
con tributions from Xero x P AR C, USC/ISI, UCB, LBNL
Some
appro ximate n um b ers: { 27K lines
f
C++ co de { 12K lines
f
OTcl supp
rt
co de { 18K lines
f
test suites, examples { 5K lines
f
do cumen tation!
See
main VINT and NS-2 w eb pages at: http://netweb.usc.edu/vint http://www-mash.cs.berkeley.edu/ns/ ns.htm l
to C++: { C++ has single class decl ) OTcl attac hes metho ds to
b
ject
r
class { C++ constructor/destructor ) OTcl init/destro y metho ds { this ) $self { OTcl metho ds alw a ys \virtual" { C++ shado w ed metho ds called explicitly with scop e
p
erator ) OTcl metho ds com bined implicitly with $self next { C++ static v ariables ) OTcl class v ariables { (m ultiple inheritance is supp
rted)
NSv2 W
rkshop
kfal l@e e.lbl.gov Slide 10
SLIDE 11OTcl Basics (con td)
use
instvar and instpr
c
to dene/access mem b er functions and v ariables
Example:
Class Counter Counter instproc init fg f $self instvar cnt set cnt g Counter instproc bump fg f $self instvar cnt incr cnt g Counter instproc val fg f $self instvar cnt return $cnt g Counter c c val ! c bump c val ! 1 NSv2 W
rkshop
kfal l@e e.lbl.gov Slide 11
SLIDE 12C++/OTcl Split Ob jects
Split
b
jects: implemen t metho ds in either language
new
and delete set c [new Counter] $c val
>
$c bump $c val
>
1 delete $c
Dene
instance v ariables in either C++
r
OTcl: Counter::Counter() f bind("cnt ", &value ); value = 10; ... g vs. $self set cnt 10 bind() simply uses Tcl T r ac eV ar NSv2 W
rkshop
kfal l@e e.lbl.gov Slide 12
SLIDE 13Example: a simple sim ulation
A
small but complete sim ulation script: { set up 4-no de top
logy
and
ne
bulk-data transfer TCP { arrange to trace the queue
n
the r1-k1 link { place trace
utput
in the le simp.out.tr
#
Create a simple four node topology: # s1 # \ # 8Mb,5ms \ 0.8Mb,50ms # r1
k1
# 8Mb,5ms / # / # s2 set stoptime 10.0 set ns [new Simulator] set node_(s1) [$ns node] set node_(s2) [$ns node] set node_(r1) [$ns node] set node_(k1) [$ns node] $ns duplex-link $node_(s1) $node_(r1) 8Mb 5ms DropTail $ns duplex-link $node_(s2) $node_(r1) 8Mb 5ms DropTail $ns duplex-link $node_(r1) $node_(k1) 800Kb 50ms DropTail $ns queue-limit $node_(r1) $node_(k1) 6 $ns queue-limit $node_(k1) $node_(r1) 6 set tcp1 [$ns create-connection TCP $node_(s1) TCPSink $node_(k1) 0] $tcp1 set window_ 50 $tcp1 set packetSize_ 1500 # Set up FTP source set ftp1 [$tcp1 attach-source FTP] set tf [open simp.out.tr w] $ns trace-queue $node_(r1) $node_(k1) $tf $ns at 0.0 "$ftp1 start" $ns at $stoptime "close $tf; puts \"simulation complete\"; $ns halt" $ns run NSv2 W
rkshop
kfal l@e e.lbl.gov Slide 13
SLIDE 14Example: a simple sim ulation (con t)
The
trace le pro duced lo
ks
lik e this: + 0.0065 2 3 tcp 1500
0.0
3.0
0.0065
2 3 tcp 1500
0.0
3.0 + 0.23344 2 3 tcp 1500
0.0
3.0 1 2
0.23344
2 3 tcp 1500
0.0
3.0 1 2 + 0.23494 2 3 tcp 1500
0.0
3.0 2 3
0.24844
2 3 tcp 1500
0.0
3.0 2 3 + 0.46038 2 3 tcp 1500
0.0
3.0 3 6
0.46038
2 3 tcp 1500
0.0
3.0 3 6 + 0.46188 2 3 tcp 1500
0.0
3.0 4 7 + 0.47538 2 3 tcp 1500
0.0
3.0 5 8 ... + 0.98926 2 3 tcp 1500
0.0
3.0 25 40 + 0.99076 2 3 tcp 1500
0.0
3.0 26 41 d 0.99076 2 3 tcp 1500
0.0
3.0 26 41
1.00426
2 3 tcp 1500
0.0
3.0 21 36 + 1.00426 2 3 tcp 1500
0.0
3.0 27 42 + 1.00576 2 3 tcp 1500
0.0
3.0 28 43 d 1.00576 2 3 tcp 1500
0.0
3.0 28 43
1.01926
2 3 tcp 1500
0.0
3.0 22 37 + 1.01926 2 3 tcp 1500
0.0
3.0 29 44 + 1.02076 2 3 tcp 1500
0.0
3.0 30 45 d 1.02076 2 3 tcp 1500
0.0
3.0 30 45
1.03426
2 3 tcp 1500
0.0
3.0 23 38
1.04926
2 3 tcp 1500
0.0
3.0 24 39
1.06426
2 3 tcp 1500
0.0
3.0 25 40 ... NSv2 W
rkshop
kfal l@e e.lbl.gov Slide 14
SLIDE 15The Sim ulator
Sim
ulator API is a set
f
metho ds b elonging to a simulator
b
ject:
Create
a sim ulator with: set ns [new Simulator]
What
this do es: { initialize the pac k et format (calls create packetformat) { create a sc heduler (defaults to a simple link ed-list sc heduler)
Sc
heduler: { handles time, timers and ev en ts (pac k ets), deferred executions (\A Ts") { Scheduler/List
link
ed-list sc heduler { Scheduler/Heap
heap-based
sc heduler { Scheduler/Calendar
calendar-queue
sc heduler { see Reev es, "Complexit y Analyses
f
Ev en t Set Algorithms", The Computer Journal, 27(1), 1984 NSv2 W
rkshop
kfal l@e e.lbl.gov Slide 15
SLIDE 16Using the sc heduler
Sc
heduler API is through Sim ulator
b
ject: Simulator instproc now ;# r eturn sche duler's notion
f
curr ent time Simulator instproc at args ;# sche dule exe cution
f
c
de
at sp e ci- e d time Simulator instproc run args ;# start sche duler Simulator instproc halt ;# stop (p ause) the sche duler Simulator instproc create-trace type files src dst ;# cr e- ate tr ac e
bje
ct Simulator instproc create_packet fo rm at ;# set up the simula- tor's p acket format
Example:
MySim instproc begin {} { ... set ns_ [new Simulator] $ns_ use-scheduler Heap $ns_ at 300.5 "$self complete_sim" ... } MySim instproc complete_sim {} { ... } NSv2 W
rkshop
kfal l@e e.lbl.gov Slide 16
SLIDE 17Sim ulator Timing
eac
h
b
ject has a generic receiv e metho d NsObject::recv(Packet*, Handler* h = 0)
most
b
jects ha v e single neigh b
r
Connector::target
cut-thr
ugh
transfers; send pac k et directly to neigh b
r
without in v
lving
sc heduler Connector::send(Packet* p) f target
>recv(p);
g
b
arrier: { an y p
in
t that adv ances time in to future (i.e., dela y elemen t) { need in ter-ob ject \proto col" to decouple timing { barrier tak es non-n ull Handler { sc hedule dela y and in v
k
e handler
n
completion { example: queue/dela y
b
jects (later) NSv2 W
rkshop
kfal l@e e.lbl.gov Slide 17
SLIDE 18P ac k ets
pac
k ets are events (ma y b e sc heduled to \arriv e")
con
tain header section and (sometimes) data
header
section is a cascade
f
all in-use headers
all
pac k ets con tain a c
mmon
he ader: { pac k et size
used
to compute transmission time { timestamp, t yp e, uid, in terface lab el (for debugging, and m ulticast routing)
new
proto col agen ts ma y need to dene new headers NSv2 W
rkshop
kfal l@e e.lbl.gov Slide 18
SLIDE 19P ac k et Header F
rmat
hdrlen_ ip header body tcp header body rtp header body trace header body next_
Packet
at compile time at compile time size determined at compile time size determined at compile time size determined size determined points to next packet in either free list or in a PacketQueue bits() accessdata() packet data Size determined at simulator config time, stored in hdrlen_
Figure 1: A P ac k et Ob ject
header
con ten ts are constructed at sim ulator initialization time
p
erformed b y create packetformat NSv2 W
rkshop
kfal l@e e.lbl.gov Slide 19
SLIDE 20Connectors
Conne
ctor: simple in/out top
logy
b
ject with \drop target" /* * An NsObject with
nly
a single neighbor. */ class Connector : public NsObject { public: Connector(); inline NsObject* target() { return target_; } virtual void drop(Packet* p); protected: int command(int argc, const char*const* argv); void recv(Packet*, Handler* callback = 0); inline void send(Packet* p, Handler* h) { target_->recv(p, h); } NsObject* target_; NsObject* drop_; // drop target for this connector };
if
drop target undened, dropp ed pac k ets are freed NSv2 W
rkshop
kfal l@e e.lbl.gov Slide 20
SLIDE 21Error Mo dels
Err
r
Mo del: a (simple) parameterized lossy connector (can b e used as a base class for
ther
loss mo dels)
drops
pac k et
r
sets \error" bit (in common header)
error
units: pac k ets, bits, time Usage: # create a loss_module and set its packet error rate to 1 percent set loss_module [new ErrorModel] $loss_module set rate_ 0.01 #
ptional:
set the unit and random variable $em unit pkt # error unit: packets (the default) $em ranvar [new RandomVariable/Uniform] # set target for dropped packets $loss_module drop-target [$ns_ set nullAgent_]
if
drop target undened, dropp ed pac k ets are freed NSv2 W
rkshop
kfal l@e e.lbl.gov Slide 21
SLIDE 22Agen ts
A
gents: usually a proto col endp
in
t/en tit y (but ma y also b e used for implemen ting routing proto cols)
Where
they t in:
Agent
Port Classifier Local Addr
Link Link
Classifier Addr
Node
Node.entry
Agent Agent
What
they pro vide: { a lo cal and destination address (lik e an IP-la y er sender) { functions for helping to generate/ll-in in pac k et elds NSv2 W
rkshop
kfal l@e e.lbl.gov Slide 22
SLIDE 23Creating a new Agen t
The
Agent class: class Agent : public Connector { public: Agent(int pktType); virtual ~Agent(); virtual void timeout(int tno); protected: int command(int argc, const char*const* argv); void recv(Packet*, Handler*); ...
basic
tasks to create a new agen t: 1. decide its inheritance structure 2. create the class, recv, and timeout functions (if needed) 3. dene OTcl link age functions (Kannan will explain ho w later) 4. write the necessary OTcl co de to access y
ur
agen t
hardest
part ma y b e understanding the OTcl/C++ in teraction (fortunately , m uc h
f
this is shielded from y
u
if y
u
so c ho
se)
NSv2 W
rkshop
kfal l@e e.lbl.gov Slide 23
SLIDE 24Example: the Message Agen t
pro
vides a v ery simple place to store a message
P
ac k et header (from message.h): struct hdr_msg { char msg_[64]; /* per-field member functions */ char* msg() { return (msg_); } int maxmsg() { return (sizeof(msg_)); } };
OTcl
link age (for class creation, from message.cc): static class MessageHeaderC la ss : public PacketHeaderC la ss { public: MessageHeaderCl as s( ) : PacketHeaderCl as s( "P ac ke tH ead er /M es sa ge ", sizeof(hdr_msg )) {} } class_msghdr; NSv2 W
rkshop
kfal l@e e.lbl.gov Slide 24
SLIDE 25Example: the Message Agen t (con t)
The
class denition, constructor and v ariable link age: static class MessageClass : public TclClass { public: MessageClass() : TclClass("Agen t/ Me ss ag e" ) {} TclObject* create(int, const char*const*) { return (new MessageAgent()) ; } } class_message; class MessageAgent : public Agent { public: MessageAgent() ; int command(int argc, const char*const* argv); void recv(Packet*, Handler*); protected: int
ff_msg_;
}; MessageAgent:: Me ss ag eA ge nt () : Agent(PT_MESSA GE ) { bind("packetSi ze _" , &size_); bind("off_msg_ ", &off_msg_); } NSv2 W
rkshop
kfal l@e e.lbl.gov Slide 25
SLIDE 26Example: the Message Agen t (con t)
Main
functions: void MessageAgent::recv(Packet * pkt, Handler*) { hdr_msg* mh = (hdr_msg*)pkt->access(off_ msg_ ); ... process packet ... } int MessageAgent::command(int argc, const char*const* argv) { Tcl& tcl = Tcl::instance(); // call into interp if (argc == 3) { // $obj send msgtext if (strcmp(argv[1], "send") == 0) { Packet* pkt = allocpkt(); hdr_msg* mh = (hdr_msg*)pkt->access(off _ms g_); const char* s = argv[2]; int n = strlen(s); if (n >= mh->maxmsg()) { tcl.result("message too big"); Packet::free(pkt); return (TCL_ERROR); } strcpy(mh->msg(), s); send(pkt, 0); return (TCL_OK); } } return (Agent::command(argc, argv)); // for inheritance } NSv2 W
rkshop
kfal l@e e.lbl.gov Slide 26
SLIDE 27TCP Agen ts
ns
has sev eral v arian ts
f
TCP a v ailable: { Agen t/TCP
a
\taho e" TCP sender { Agen t/TCP/Reno
a
\Reno" TCP sender { Agen t/TCP/NewReno
Reno
with a mo dication { Agen t/TCP/Sac k1
TCP
with selectiv e rep eat (follo ws RF C2018) { Agen t/TCP/V egas
TCP
V egas { Agen t/TCP/F ac k
Reno
TCP with \forw ard ac kno wledge- men t"
The
ne-w
a y TCP receiving agen ts curren tly supp
rted
are: { Agen t/TCPSink
TCP
sink with
ne
A CK p er pac k et { Agen t/TCPSink/DelAc k
TCP
sink with congurable dela y p er A CK { Agen t/TCPSink/Sac k1
selectiv
e A CK sink (follo ws RF C2018) { Agen t/TCPSink/Sac k1/DelAc k
Sac
k1 with DelAc k
The
t w
-w
a y exp erimen tal sender curren tly supp
rts
nly
a Reno form
f
TCP: { Agen t/TCP/F ullTcp NSv2 W
rkshop
kfal l@e e.lbl.gov Slide 27
SLIDE 28Base TCP Agen ts
TCP
(T aho e), TCP/Reno, and TCP/NewReno
Common
features: { computations all in pac k et units w/congurable pac k et size { fast retransmit { slo w-start and congestion a v
idance
{ dynamic R TT estimation and R TX timeout assignmen t { sim ulated (constan t) receiv er's adv ertised windo w
T
aho e TCP: { p erform slo w-start
n
an y loss (R TO
r
fast retransmit) { no fast reco v ery
Reno
TCP: { fast r e c
very:
inate cw nd b y dup ac k coun t un til new A CK { slo w-start
n
R TO {
n
fast retransmit: cw nd cur w in=2 , ssthr esh cw nd
\Newreno"
TCP: { mo dest mo dication to Reno TCP {
nly
exit fast reco v ery after A CK for highest segmen t arriv es { helps reduce \stalling" due to m ultiple pac k et drops in a windo w NSv2 W
rkshop
kfal l@e e.lbl.gov Slide 28
SLIDE 29Other TCP Agen ts
TCP/Sack,
TCP/Fack, and TCP/Vegas
Selectiv
e A CK TCP: { SA CK sim ulation based
n
RF C2018 { A CKs carry extra information indicating receiv ed segmen ts { requires SA CK-a w are sink { sender a v
ids
sending redundan t info { default to 3 \SA CK blo c ks" (for using timestamps, see RF C2018)
blo
c k con tains start/end sequence n um b ers
blo
c k con taining most recen tly receiv ed segmen t alw a ys presen t { regular A CK n um b er still giv es nal sa y
F
ac k TCP: { \forw ard A CK" TCP (exp erimen tal, see SIGCOMM '96) { use SA CK info for estimate
f
pac k ets in the net w
rk
{
v
erdamping algorithm (to limit slo w-start
v
ersho
t)
{ ramp do wn algorithm (for transmission smo
thing)
V
egas TCP: { con tributed co de from T ed Kuo (NC State Univ) { not directly supp
rted
at this time NSv2 W
rkshop
kfal l@e e.lbl.gov Slide 29
SLIDE 30TCP Agen t P arameters
Common
conguration parameters and defaults for TCP agen ts: Agent/TCP set window_ 20 ;# max b
und
n
window size Agent/TCP set windowInit_ 1 ;# initial/r eset value
f
cwnd Agent/TCP set windowOption_ 1 ;# c
ng
avoid algorithm (1: standar d) Agent/TCP set windowConstant_ 4 ;# use d
nly
when windowOption != 1 Agent/TCP set windowThresh_ 0.002 ;# use d in c
mputing
aver age d window Agent/TCP set
verhead_
;# !=0 adds r andom time b etwe en sends Agent/TCP set ecn_ ;# TCP should r e act to e cn bit Agent/TCP set packetSize_ 1000 ;# p acket size use d by sender (bytes) Agent/TCP set bugFix_ true ;# se e do cumentation Agent/TCP set slow_start_restart_ true ;# se e do cumentation Agent/TCP set tcpTick_ 0.1 ;# timer gr anularity in se c (.1 is NONST AND ARD) Agent/TCP set maxrto_ 64 ;# b
und
n
R TO (se c
nds)
Agent/TCP set dupacks_ ;# duplic ate A CK c
unter
Agent/TCP set ack_ ;# highest A CK r e c eive d Agent/TCP set cwnd_ ;# c
ngestion
window (p ackets) Agent/TCP set awnd_ ;# aver age d cwnd (exp erimental) Agent/TCP set ssthresh_ ;# slow-stat thr eshold (p ackets) Agent/TCP set rtt_ ;# rtt sample Agent/TCP set srtt_ ;# smo
the
d (aver age d) rtt Agent/TCP set rttvar_ ;# me an deviation
f
rtt samples Agent/TCP set backoff_ ;# curr ent R TO b acko factor Agent/TCP set maxseq_ ;# max (p acket) se q numb er sent NSv2 W
rkshop
kfal l@e e.lbl.gov Slide 30
SLIDE 31TCP Sink Agen ts
Sinks
for
ne-w
a y TCP senders
T
yp es { standard sinks, dela y ed-A CK sinks, SA CK sinks
Standard
sinks: { generate
ne
A CK p er pac k et receiv ed { A CK n um b er
v
erloaded in \sequence n um b er" pac k et eld
Dela
y ed-A CK sinks: { same as standard, but with v ariable dela y added b et w een A CKs { time to dela y A CKs sp ecied in seconds
SA
CK sinks: { generates additional information for SA CK capable sender { congurable maxSackBlocks parameter NSv2 W
rkshop
kfal l@e e.lbl.gov Slide 31
SLIDE 32Tw
-W
a y TCP (\F ullTCP")
most
TCP
b
jects are
ne-w
a y (and require a source/sink pair)
real
TCP can b e bi-directional
sim
ultaneous t w
-w
a y data transfer alters TCP dynamics con- siderably
(new{
still undergoing debugging)
the
TCP/FullTcp agen t: { follo ws closely to \Reno" TCP implemen tation in 4.4 BSD { b yte-orien ted transfers { t w
-w
a y data supp
rted
{ most
f
the connection establishmen t/teardo wn { symmetric:
nly
ne
agen t t yp e used for b
th
sides NSv2 W
rkshop
kfal l@e e.lbl.gov Slide 32
SLIDE 33F ullTCP P arameters
P
arameters and defaults: Agent/TCP/FullTcp set segsperack_ 1 ;# se gs r e c eive d b efor e gener ating A CK Agent/TCP/FullTcp set segsize_ 536 ;# se gment size (MSS size for bulk xfers) Agent/TCP/FullTcp set tcprexmtthresh_ 3 ;# dupA CKs thr esh to trigger fast r exmt Agent/TCP/FullTcp set iss_ ;# initial send se quenc e numb er Agent/TCP/FullTcp set nodelay_ false ;# disable sender-side Nagle algorithm Agent/TCP/FullTcp set data_on_syn_ false ;# send data
n
initial SYN? Agent/TCP/FullTcp set dupseg_fix_ true ;# avoid fast rxt due to dup se gs+acks Agent/TCP/FullTcp set dupack_reset_ false ;# r eset dupA CK ctr
n
!0 len data se g s c
ntaining
dup A CKs Agent/TCP/FullTcp set interval_ 0.1 ;# delaye d A CK interval NSv2 W
rkshop
kfal l@e e.lbl.gov Slide 33
SLIDE 34T rac Sources
Sources
(\applications") used to driv e agen ts
curren
tly used
nly
b y TCP
T
yp es: { T elnet
sim
ulates c haracters t yp ed b y a user { FTP
bulk
data transfer
OTcl
In terface: $src start ;# start sending p ackets $src stop ;# stop sending p ackets $src attach-agent ;# asso- ciate agent with sour c e $ftpsrc produce npkts ;# send npkts num- b er
f
p ackets $ftpsrc producemore npkts ;# send npkts mor e p ackets
API
is still under some dev elopmen t
sources
nly
used b y TCP at this time NSv2 W
rkshop
kfal l@e e.lbl.gov Slide 34
SLIDE 35T elnet T rac Source
ma
y sp ecify interval
if
zero, pic ks randomly among 10000 measured in terarriv als (TCPLIB)
if
nonzero, uses scaled exp
nen
tial for in terarriv als
pac
k et size constan t (but a v ailable via bind call) NSv2 W
rkshop
kfal l@e e.lbl.gov Slide 35
SLIDE 36CBR and UDP Agen ts
CBR
Agen ts: { stands for \constan t bit rate" (not really used
nly
this w a y) { non-connection-orien ted sending agen t { sends pac k ets at p erio dic in terv al
r
quasi-p erio dically { constan t-size pac k ets
UDP
Agen ts: { v ery similar to CBR agen ts { uses TrafficGenerator class for pac k et sizes/times NSv2 W
rkshop
kfal l@e e.lbl.gov Slide 36
SLIDE 37R TP and R TCP Agen ts
R
TP
\Real-time"
(transp
rt)
proto col (RF C 1889) { implemen ted as Agent/CBR/RTP
b
ject { sp ecial \R TP" header (con tains seq n um b er and srcID) { sends data p erio dically similar to CBR sources { resets faster when mo ving from high to lo w rate
R
TCP
con
trol proto col for R TP { implemen ted as Session/RTP
b
ject { sends at rate based
n
n um b er
f
ther
senders { rep
rts
kno wn sources and stats NSv2 W
rkshop
kfal l@e e.lbl.gov Slide 37
SLIDE 38Other Simple Agen ts
the
LossMonitor agen t: { monitors arriv als
f
pac k ets { lo
ks
for sequence n um b er holes { pro vides coun ters for:
nlost
n
um b er
f
holes in n um b er space
npkts
pac
k et arriv als
bytes
b
yte arriv als
lastPktTime
time
f
last arriv al
expected
next
seq n um b er exp ected
the
Message agen t: { v ery simple agen t { allo ws for including text \messages" in pac k ets { curren tly limited to at most 64 b yte (short) messages NSv2 W
rkshop
kfal l@e e.lbl.gov Slide 38
SLIDE 39T ap Agen ts and the \Real W
rld"
allo
ws the sim ulator to in teract with a real net w
rk
(curren tly exp erimen tal)
T
ap Agen ts: { for no w uses 1600 b yte buer as \header" (ie. ether frame + slop) { bi-directional agen t b et w een sim ulation and net w
rk
{ uses abstract \net w
rk"
b
ject { receiv es
ne
pac k et p er ev en t (handled through Tcl I/O)
the
Scheduler/RealTime class { sp ecial v ersion
f
(curren tly List-based) sc heduler { ties sim ulated time to real-time { for no w, pun ts if sim ulation gets far b ehind { (can still do in teresting things!) NSv2 W
rkshop
kfal l@e e.lbl.gov Slide 39
SLIDE 40Net w
rk
Ob ject
abstraction
f
a (real-w
rld)
net w
rk
base
class for sp ecic net w
rk
t yp es (e.g. IP net w
rk)
used
b y
ther
to
ls
in Katz/McCanne/Brew er's MASH pro ject (see http://www-mash.cs.berkeley.edu/mash/ index. html)
Net
w
rk
class: { requires so c k et system API (UNIX
r
WinSo c k) { supp
rts
a basic send/recv in terface { separate send/recv \c hannels" (i.e. so c k ets) { non-blo c king
ptional
{ framew
rk
supp
rts
m ulticast, addr/iface selection, etc
IP
Net w
rk
(Network/IP class) { m ulticast group mem b ership { lo
pbac
k
n/o
con trol { implemen ts m ulticast and unicast con trols for IP net w
rks
NSv2 W
rkshop
kfal l@e e.lbl.gov Slide 40
SLIDE 41T rac Generator
generate
trac according to distributions
r
traces
generally
used for CBR/UDP agen ts
Exp
nen
tial { exp
nen
tially distributed
n/o
times { parameters:
n
time,
time,
rate, pac k et size { what these mean:
burst
for exp
time
with mean
n
time
b
e silen t for exp
time
with mean
time
while
bursting, send at rate rate
use
appropriate in ter-departure time giv en rate/size
P
areto { pareto distributed
n/o
times { (man y aggregated together can b e LRD) { parameters:
n
time,
time,
rate, shap e, pac k et size { what these mean:
lik
e exp
,
except pareto using shap e parameter NSv2 W
rkshop
kfal l@e e.lbl.gov Slide 41
SLIDE 42T race-Based T rac Generator
generate
trac according to trace le
t
w
classes:
Tracefile and Traffic/Trace
trace
le uses small binary format: { rst 32-bit eld: in ter-pac k et time (microsecs) { second 32-bit eld: pac k et size (b ytes) NSv2 W
rkshop
kfal l@e e.lbl.gov Slide 42
SLIDE 43Queue Managemen t and P ac k et Sc heduling
buer
management: ho w to hold and toss (mark) pac k ets
p
acket sche duling: what pac k ets get to depart when
Buer
managemen t: { Drop-tail (FIF O) { Random Early Detection (RED)
P
ac k et sc heduling: { FIF O { CBQ (includes priorit y + round-robin) { Round-robin (DRR) { V arian ts
f
F Q (WF Q, SF Q) NSv2 W
rkshop
kfal l@e e.lbl.gov Slide 43
SLIDE 44Queue Handlers
Dequeued
pac k ets are
ften
sen t do wnstream to delays
dela
ys (usually) cause t w
actions:
1. the pac k et is sc heduled to arriv e do wnstream at time t + d 2. the queue b ecomes un blo c k ed at time t 3. t is transmit time, d is prop dela y time
so,
dela ys represen t a commonly-o ccurring sche duling b arrier
Queue
parameters: Queue set limit_ 50 ;# max p acket c
unt
in queue Queue set blocked_ false ;# queue starts
blo
cke d Queue set unblock_on_res um e_ true ;# queue is unblo cke d af- ter r esume
con
trol
f
blo c king can b e useful for queue banks (e.g. CBQ) NSv2 W
rkshop
kfal l@e e.lbl.gov Slide 44
SLIDE 45Drop T ail and RED Queues
Drop-T
ail Queues (Queue/DropTail class) { simple FIF O, drop-tail queues { drop from tail when
ccupancy
reac hes qlim
RED
(Random Early Detection) Queues (Queue/RED class) { active buer managemen t tec hnique { t w
thresholds:
minth and maxth { also a maxim um probabilit y maxpr
b
{ compute aver age queue
ccupancy
v
er time { if a v erage exceeds maxth (or qlim ) drop a pac k et { if a v erage is under min th, allo w pac k et to en ter queue { b et w een, scale drop probabilit y linearly
n
[0; maxpr
b]
NSv2 W
rkshop
kfal l@e e.lbl.gov Slide 45
SLIDE 46RED Queue P arameters
bytes
do computations in b ytes instead
f
pac k ets (requires assignmen t
f
a mean pac k et size estimate)
thresh
min
thresh
maxthresh
max
thresh
mean
pktsize
used
for computing estimated link utilizations during idle p erio ds
q
weight
w
eigh t giv en to instan taneous queue
ccupancy
for EWMA
wait
RED
should force a w ait b et w een drops
linterm
recipro
cal
f
maxprob
setbit
mark
instead
f
drop
drop-tail
drop
new pkt instead
f
random
ne
NSv2 W
rkshop
kfal l@e e.lbl.gov Slide 46
SLIDE 47T race and Monitoring Supp
rt
Tw
main
items: tr ac es and monitors
T
races
write
an en try for some ev en t (often pac k et arriv als/departures/drops) { Trace/Enque
a
pac k et arriv al (usually at a queue) { Trace/Deque
a
pac k et departure (usually at a queue) { Trace/Drop
pac
k et drop (pac k et deliv ered to drop-target)
Monitors
k
eep statistics ab
ut
arriv als/departures/drops (and
ws)
{ SnoopQueue/Out
n
utput,
collect a time/size sample (pass pac k et
n)
{ SnoopQueue/Drop
n
drop, collect a time/size sample (pass pac k et
n)
{ SnoopQueue/EDrop
n
an "early" drop, collect a time/size sampl e (pass pac k et
n)
{ QueueMonitor
receiv
e and aggregate collected samples from sno
p
ers { QueueMonitor/ED
queue-monitor
capable
f
distinguish- ing b et w een \early" and standard pac k et drops { QueueMonitor/ED/Flowmon
p
er-o w statistics monitor (manager) { QueueMonitor/ED/Flow
p
er-o w statistics con tainer NSv2 W
rkshop
kfal l@e e.lbl.gov Slide 47
SLIDE 48T race File F
rmat
File
format for traces generally
f
this form: + 1.45176 2 3 tcp 1000
1
256 769 27 48 + 1.45276 2 3 tcp 1000
1
256 769 28 49
1.46176
2 3 tcp 1000
1
256 769 22 43 + 1.46176 2 3 tcp 1000
1
256 769 29 50 + 1.46276 2 3 tcp 1000
1
256 769 30 51 d 1.46276 2 3 tcp 1000
1
256 769 30 51
1.47176
2 3 tcp 1000
1
256 769 23 44 + 1.47176 2 3 tcp 1000
768
3 52 + 1.47276 2 3 tcp 1000
768
4 53 d 1.47276 2 3 tcp 1000
768
4 53
Fields:
arriv al/departure/drop, time, trace link endp
in
ts, pac k et t yp e, size, ags,
w
ID, src addr, dst addr, sequence n um b er, uid
Man
y
f
these elds are from the common pac k et header: struct hdr_cmn { double ts_; // timestamp: for q-delay measurement int ptype_; // packet type (see above) int uid_; // unique id int size_; // simulated packet size int iface_; // receiving interface (label) static int
yd and Jacobson, "Link-sharing and Resource Managemen t Mo dels for P ac k e t Net w
rks",
T
N,
Aug 1995
rewrite
from CBQ co de in ns-1
pac
k ets are mem b ers
f
classes
classes
ma y con tain a priority and a b andwidth al lo c ation
classes
ma y b
rr
w
un used bandwidth from
ther
classes
pac
k ets are sc heduled using a round-robin sc heduler according to the classes they b elong to: { pac k et-b y-pac k et RR { w eigh ted RR { high-to-lo w priorit y NSv2 W
rkshop
kfal l@e e.lbl.gov Slide 55
SLIDE 56CBQ Implemen tation
Queue (arbitrary) Queue (arbitrary) SD SD Connector CBQClass CBQClass Classifier (e.g. hash) SI SI SO SO CBQ (scheduler) QMon QMon
CBQLink
head_ drophead_ classifier_ queue_
Ma
jor comp
nen
ts: { classier (maps pac k ets to classes) { classes (holds class state) { sc heduler (sc hedules pac k et departures)
Implemen
ted as a sub class
f
link: CBQ link NSv2 W
rkshop
kfal l@e e.lbl.gov Slide 56
SLIDE 57Router Mec hanisms
Flo
yd and F all, "Router Mec hanisms to Supp
rt
End-to-End Congestion Con trol", LBNL TR, F eb 1997
p
rt
from ns-1 v ersion based
n
new Flo wMon and CBQ
Connector RED Queue RED Queue SD CBQClass CBQClass Classifier SI SI SO SO CBQ (scheduler) FlowMon head_ drophead_ classifier_ queue_
