Sie sind auf Seite 1von 14

!e.

Adjust
railable
iand 5(XX)
l in thesc

I
a
[,"" ¿

Copocity Assignment I
n
¡ptcr.
I
t timc )istributed Networks
I },1. Ur
*¡tionr
I
§ E prcvious chapter we discussed capacity assignment for a very sim- t
¡aUnim, l¡¡l M-tcrmin¿ls at seven citirx feeding in to a úal processor located
!¡ toohri in,c :!ffral point. More generally, we have data mcssages flowing between
¡ nctwort lm¡§ :¡¡¡triffi-Étwork. This gives rise to the mesLtype network of Fig.
FEGats r [ü-r- -¡c -r.RPA network, TYMNEI, and the §ITA network, discussed in
to Da¡c '(Eh. ' l:l existing examples of this §pe of networt. ü
t thc rifr-- ;:ÉHsscd in conjunction with the simple con6guration of the previous
r¡@lg :,oriM oneis again willing to invoke the indcpcndene assumption
icdro
rfurc ácre. To be specific consider the hypothetizl network shown in
rity t" - Thk hee been taken from the book by L. Kleinrock previously
rü¡n:mr={' Y/ith the trafrc, its statistical characteristirx, and routes taken
Es
fh * gr( girtn between all pairs of five cities show-n- urc shall, repeating the
m1^,ú:[ .f, thc last chapter, indicate how one optimalt]¡ assigns capacity to
(rr r' -¡r scryEn links in the network. We shall then om¡rare this square-root
to@r.:r:, :ssienment for minimizittg zYerage time delay with two other
rr "*,:e.: r.trategies, proportional and equal assignment.
-:r :apaciry assignments depend on the routing strategies adopted. We
qffi. n'.q:gate the effect of routing briefly by chang:ag the routing strategy
:rumt.,¡i,f,: and noting the effect on capacity assigamert. Routing algorithms
'nuü ::re in detail in Chap. tl.
=esj&red
*¡¡r:r¡cl, Communicalion Nets, p' 23.
r--
TY A§SiGNMENT lN DiSÍ8tBu-:l
72 CAPACITY ASSIGNMENT IN DISTRIBUTED NETWORK§

TAtsLE :.'
TRAFFIC MATñ¡X, I.í'i, :: .
New Ycn
Destination:
i ,qri
Ne»'
York Chrcagc
:- --- 1-;r,' l:
\:'.r York J,j.1
- Jlcago 9.34

r"/ - -{o'üstoo

,esru
0.935 0.810
rf,r
l.tl 4 ¿ - ¡c Angeles 2.94 2.10 :c
: )¡ver 0.610 0.62S ..:
Los Ange

r#i
':-:r:r L, Kleinrock, Communicalion ly'eii. ; I
Houston
irFr!:iE-E ::affis gtl¿¡acteristics in either il::---':q

Fig. *1. Net*ort example, ir's (messages/sec) slpwn' eiüer direction mml=;: link will be the same, as will queui":-i :r
(shortest distarc routing)' lm :i i xrticular link. This enables us to :--.r-¡
}i rlmll :: ji¡ection only, reducing the prc:.:
r, :eiays over 14 one-way links to c'r"'.' i
-
This particular example will be discussed in this chaptef because rruüüisr* E
appeared in follow-up work in the litsature in discussions mil 1§i:Ee flow over seven links only, T:e ::::
relati¡§
áiri"r"", types of p.tfo..uoo criteria us€d in as§igning capacity' ]lluü@ "' i.
ffi include, in cont¡ast to the minimum time delay criterioa' minjmum I: :e:ermine tie capacity assignmenr ::: a
.qr"* il*" delay aod minimum of the maximum time dclay. lYe shaJl n@uuü¡tr ¿" ::ted earlier the independenc€ assl-;':-n

marize these pa¡rs extending Kleinrock's work to other cost


functim luü[umm]:;i,:r over any li:=-
link is statisticall¡
;F c,ompare, for the same network, the effect of using
different cost funq mrmrl -:::
:---e in the network. We must alsc
' io ¿s¡srmi',ing the capacity allocations for this network one must 'mmu
iu:ñ::¡ over each of the links. For thts ;-;
---* the shortest geographical rou::
to
the trafrc flowing beiween cities as well as the route assigoed mñ!ürfüü4§
F the network at one city and destined for another' As an
.:* .{lgeies-Chicago traffic routed i:--:
the traffic matrix of Table 4-1, indicating the average rate of messagE
-
"oi"riog
(messages/sec) between all city-pairs, is assumed given'f * :r .{:geies-New York traffic th¡ou-- I
' Náte'thai the traffc flowing has been assumed symmetrical §.Y ar
" l,*-.::-New York traffic through Ci-;¿
traffic: L.A. to N.Y. trafñc) for simplicity' Asymmetrical traffic flo¡r
in the same manner' The tolal average messatÉ ,!irm--¡r=-) . letting ?l& messagesf sec rep ie-:
easily be handled
emanating from any one city is of course given by summing ala : :, l.bie 4-1 showing messages enie;r:-;
appropria-te horizontal row. The average number of messages/sec
arÍl lülr, 1 ! -:3 average One-Way message i¿:e -:
city is found by summing the appropriate colug'n' Tht I

ali the entries in the table provides the total network trafrc input
"'parücot"r
All seven links in the nitwork of Fig' zl-1 are assumed to be full
that the capacitf in either direction is the same' In addiüon' becalx
symmsl¡1c¿1 nature of the traffic flow shown in Table
4-1, it is aPF¡-'l

¡lbid.. p. ll. Th:sspeciEc example is used' asnotedearlier' because er:=


have appeared in the hte=rure' AJthough lhe.eltriT shown a¡e i" --Ys,:-
leng:- -:
easill,ü converred ro :is tbits/sec) bl multiplling by average messá8e
erar,ple. if messages qe:: ali on the average 1000 bi¡s.long'
1tl
eltnes
l::t :-
par?S---:
o, ,fu. The case of m¡-ual message lengrhs is considered in a late¡
-SSIGNMENT IN DiSTHIBUTED ¡IET!!Oñ(S 73
b ¡¡erwon
TA8 LE 4.1
TRAFFIC MAfñIX. NETWORK, FIG, 4.i'
Destination:
Ne*' Los
York
IL
Chicago Hous¡on Angeles
3-r5
Denver
I
.;. . \'q¡k 2.91 o.610

It1
9.34 0.935
I- cago 9.34 0.820 2.40 0.628
::'.IstOO 0.935 0.820 0.60E 0.131

-:s .{ageles 2.94 2.& 0.608 0.753


llqi'er 0.610 0.628 0.131 o.?53

:-:': L Kieinrock, Communication Nets,p.22 r


r-ai- :ialfic characteristics in either direction of trafrc flow along a
r- =: link will be the same, as
will queueing delays at the nodes at either
rcction :ru*-
: : particular link. This enables us to focus attention on the traffic in
Ír".-
:r -: di¡ection only, reducing the problem from that of'calculating
I
rcause it links to only 7. we shall thus focus on one-
r relating
acity.
.i-".* :3 jelays over 14 one-way
r!... -_-.:¿ge flow over seven links only. The total ¡etwotk flow is obviously I
imum
,- " :"f --:5.
-:
:etermine the capacity assignment for each of the seven links we
I
/e shall
unctions
r r u3 r :oted earlier thi independence assumption: traffic queueing up for
:L:.Ljs.cn ovef any link is itatistically independent'of traffic appearing
t functis¡r. j: 1:.- eise in the network. We must also determine the average message
: must .. : t,- ; rig over each of the links. For this purpose as§ume the touting of
to
an
- r ¿!?": :"kes the shortest geographical route:
I
oessage
' cs Angeles-Chicago traffic routed through Denver
- :s Angeles-New York traffic through Houston
N.Y. toL.r
-
; flows
l<arer-New York traffic through Chicago.
ssaSe : :-:'ically, letting yr& me§sages/sec represent the appropriate entry in
g along - - . ., -.. Table +isiláwing messages entering city and destined j for city
= "-,f links of
c arriving - : I : r ar ihe average one-way message rate in each of the seven
i
The sum
rt. 11 : f+s ,* Yn, : 3.i5 messages/sec, ooe-wav
ülduplex
l¿ : ?¿¡ -l- ?¿r
: 3.55 messagesisec, one-§av
cause of : : 0.i3 messages/sec' one-\\a)'
Pparent - f s3
: ?:¡ * ?n *'lst: 3.64 messagesfsec' cne'\\a\'
xtensions o' : lzt : 0'82 messages/sec' one-\\'a)'
i,'sec, rhe¡-
th in birs. F ia:'ltr *.vot : 3.88 messages/sec, one-\\'ay
I t e muJn; )
''
- : "t^.
I l
-l-
\
..r.,
l tL
: 9.95 messagesfsec, cne-l¡ av
raph.
74 CAPACITY ASSIGNMENT IN DISTRIBUTED NETWORKS CAPACITY ASSIGNMENT IN DISTRiBUT:: IIE

The total average message traffñc through each of the links is of'course :-ffic. Say the Chicage-N.Y. messag:r :.-
twice the one-way figures, as noted earlier. The total one-way link traffic is rrlle the Denvcr-N.Y. messages are :,r- ¡
a¡m Table 4.1 we have
,sl
¿ : É tt -- zs.tzmessages/sec
I _ 0.61(500) - 9._1¿ llt
It",, on the average. The total number of messages/sec entering the entire network tb 9.95
on the average is m üe average. (More generally, p.e \¡,:-.:
frr 'r |: E?¡* rñage lengths in the two different dirs-¡.-m
:aÉGe capacities to account for thc hea',.c: ti
where we sum over all the entries in the matrix of Table zl-1. This is just 38.3 With both the link traffic dema¡C i
Tfl rnessages/sec on the average. The one-way traffie ís thus y':yl2:i9.15
s
. e d€fined in a typical case, we are no* r:_ &
;y messages/sec. The average number of links traversed by a typical message is u ¡nk gap¿sis allocation on the ¡nessatE --r!
;tl ' (25.12119.15): 1.3.
The link capacity assignments in bps (bits/sec) depend not only on the
rúelrrÍlent, in the
@k,
sense of minimizing the
is the same as that found prwi*rx,*.
:r
r
iink trafiñc itbe demand or offered load, in nnessages/sec), but on the messap d.ate that with message statistlx Érssttur:
lengths as well. Recall from Eq. (3-1) in the previrous chapter, that the 6 ,ulysis of Chap. 3 holds troe here a. n
)r average time delay on the ¡th link due to messrlge transmission and buf[ering
while waiting for transmission to begin, is given by
+s!isr. the optimum capacity asdglc=:
:e" i,
:

r : t1 t
arrlq)t: 7, Ct . r
-]-- -
!1,C, X, Z -
(This assumes Poisson rnessage arrivals, exponentiallydistributed messaF I¡a-- úat C: E =;
C, is the overall capa;:r
lengths, and an infinite buffer for the queue, as noted earlier.) What shall wt
use for I f ¡.t,, the average message length in bits, if the messages flowing com '¡* rhlJe pc=F?,lt4. The avera= -,r
fa; from various sources and nodes ? We shall essentially ignore the issue here I Lmnrjú-r is again given by ES. (34),
simply assuming all messages in the network have the same average lengtt
l/p. This enables us to get reasonable numbers quickly. In more compler r:lti:7:
situations different types of messages may be flowing over a given link
(Recall the network examples of Chap. 2 in which different types of messagrr{ *nu& u minimu¡n time delay, found usrr_s -,:

of varying lengths may be flowing in the same direction over any link. Thest ffii" s gr;en by Eq. (!6),
could be different message categories; control or data messages; ACf : (E. ;-i
messages if not embedded in a message block; varying messagp lengtl* -'¿O
inbound to and outbound from a computer, if not blocked into fixedJen¿:it
i: a:91.v these equations to the network

c
packets, etc.) Some sort of average thus has to be taken. One simple pcs-
sibility is to weight message lengths in link i by the relative number a
;,ave the same average lengü I p. t
fi¡F(iage capacity in messageslsec is 6.¡
messages ol the particular length ffowing through the link. Thus, we defi-u
E*s¡ft und.er question we then hart
I : s- blUi!
!,- tH' E !¡* - ). 25.12
p:
lrc:_W:
b then rather lightly loaded. r-X
-§-liDL i I ¡tl7,F:o :ñfic in any one direction only.)
E ¡--,-qas (3-l), (3-5), and (3-6) cal ;.J
The summation shosn over link i refers to those messages from sou¡ix. i¡
destination I that are routed through link i. As an example, consider li¡:¡
- uficuiions for the network of Fr¿ !
*¡r
Chicago-N \'.. in Fig. -1-1. This carries Denver-N.Y. and Chicago-\ I =s¿hantcapacity allocation a¡,: :-
F
TWORKS :}:4":rf.!, ASSIGNMENT IN OISTEIEUTED NETWORKS 75

f course ::-=c Sa¡- the Chicago-N.Y. messages are 200 bits long, on the al'erage,
ffic is *:-.* :'3 Denver-N.Y. messages are 500 bits lon-e. tising the traffic data
'*:¡: lfablc &l we have
I 0.61(500)-7 9.3«200): 218 birs
network E:---85-
m :É average. (More generally, we would have to account for different
ilñ**lF ]€agths in the two different directions. For duplex lines one would
§r;rx capacities to account for the heavier direction of traffic.)
ust 38.3
+
7 n¡ both the link traffic demand l, and the average message length 1
t9.15 :e:red in a typical in position to determine the effect
cáse, we are now a
ssage is r ^,r r capacity allocation on the message time delay. Tlre optirunt capacity
*qg:ent, in the sense of minimizing the average time delay throughout the J
' on the sm.rk is the same as that found previously in Chap. 3. (A little thought ryill
message
miie:. t-b¿t with message st¿tistks assumed independent from link to link,
hat
lffering,
the
u: r-=l1sb of Chap. 3 holds true here as well.) We thus have, as in that
T,,'s,:. the optimum capacity asqigpment for üirk í given by the same I
7, , c(l - p)JT¡tr,
. (t'ttqt:- V,- -EJx,tp-- I
nessage
hall we
d*:=-' -¡at C: ZC, is the overall " oirUof
the network that rc keep t
rg come r " -¡rJe pc=FG,ltt).The average time delay for messages in the
here by rs-*,:n:r is again given by Eq. (3-4),
: Iength
om.plex
_1
T : 14 X,r,
:n link.
ressages ¡*:-ir --x minimum time delay, found using the capaci§ assignment of Eq.
. These
''- s grven by ES. (3-O,
; ACK ;F JL'l tt)'
lengths roia_
-(»1,@!_l
l-length
rle pos-
l: apply these equations to the network of Fig. 4-1 assume first that all
rber of =HH !ñ have the same average length l/p. Assume further that the overall
:¡s.r¡,:rir Elessage capacity in messages/sec is ñxed at pC : 192 messages/sec.
define ; -r :rE Detwork under question we then have a traftc intensity factor
I 25.12
P: ft:'#: o'13
(+t)
* (Recall again that we a¡e @n-
-twork is then rather lightly loaded.
¡ce / to Iquaüons (3-l), (3-5), and (3-6) can now be used to determine the
¿l-l to minimize message time
link 7, =ra:r1!' allocations for the network of Fig.
_N.Y. -
lxa The resultant capacity allocation and time delal's. by link, are indicated
:¡:ra TY ASSIGNMENT lN DISTRIBUTED \:-".:=(
CAPACITY ASSIGNMENT IN DISTRIBUTED NETWOBKS
rr*¿-f, capacity assignments is even Srea:.::-:-
in Table 4-2. Also shown are capacities ar$ time delays obtained using trvo
:fl: cáse. while the average time delai .::-:-3:
other capacity assignments :
-'::. aote that the equal assignment st:::i;--' *r
,,i 1. An equal assignment strategy in which the total capacity C is simply ti :-¿e time delay per message, does red';¡ -.-: :'
divided equally amons all the links, independent of traffic on the tink. (In this rq: ald heavy traffic links. Thus, compa:::.- - ¡'
case, then, ilC¡: pC,il : 19217 :27.4 messages/sec capacity allotted to : -: :esu1ts, the light traffic link 3 has hac r: r , iT
each link.) :::É: :¡om 206 msec to 36.6 msec. The cc::: :,: :,:r
: -: heavy traffic link ? is proportiona-,." i--
2. A proportional assignment strategy* in which C, is proportional to
r
Í,
-- --.ec. As no¿ed above we shall discu.: :- -
the traffic demand A,; i.e., we'let .
*-,";re that eliminates time delay pe:...: :, :(:
s1
pr
rt
C',lor", :-C7, (+2) rffi-::: *ith link tirne delays equalizeci rt:: -; " ;
]3 *:at is aow the effect on capacitl' as::_i:--.::
The corresponding time delays for these two alternative ways of assigning ,*un-, ti :hanges in the routing strategl I -1-":: , -.:
capacity are found by using Eqs. (3-1) and (3-4). : :ffi,- -r Chap. trl it does pay to.onsi:=: -:,. '-r
E-: rrrrrxlSe we rnake ".iust one change in tle ¡ - : -"::l
TABLE 4-2 i :e =.culations above. Assume that Lc. .j-:,.. :,-
,imr¡re: :i way of Denver and ehieagc. :ei::: -::-
F cApActTy ASSTGNMENTS, NETWORK OF FtG. 4-1 (FULL-DUPLEX)
-; xiefF*for the reader to eheek:i,3 -'--:*:
Demand, l"¡ Capacity A liocation í C-: {one-rvay, messáges/seq}
rur{T:'-rir ior ihis alternate route ne¡x'o¡k
one-way, equal
. mes- Sqilare T¡ assign- T¡ propor- i¡: 6.i T.ñ;;: -s
sages/sec) root (msec) ment (msec) tional &z: ü'6i
1 3.15 28 ñ.4 2'1.4 41.3 24 A¡: 0'13 (no c:,:5
a
3.55 30 37.8 27.4 41.9 2'7.5 X* : 6'58
¡i 3 0.13 5 2A6 27.4 36.6 I
4 3.@ 30 38 27.4 42.1 28 ),s: 0'82 (no c:':¡-:
5 0.82 13.5 78.8 27.4 37.6 6.3 xd . 0.93
6 3.88 31.5 27.4
lr : 12.89
36.2 42.5 30
H 7 9.95 54 22.6 27.4 51.3 16.5
f.io : 42 msec f"quar : 57.6 msec tñ p¡oP
-
"lMmle i:;s are indicated in the network diar':- :
54.8 msec

Several factors bear notice. First note that in order to minimize the
overall time delay the light traffic links receive less capacity than do rhe
heavy traffic links. (Compare links 3 and 5 with 7, as an example.) The tim:
delays incurred on the light traffic links are thís rnuch higher than thc:c
incurred on the hearl'traflic areas. Link 3, for exampie. has an avera_se riEL
delal' of 106 msec r¡ ilh the square-root assignment -strat€gy, tr,hile link 7 has a
time delal of 12.6 msec. The light user is thus penalized in far or of ihe hear .
user. \\-e shall discurs shortll'an alternare time delal criterion in uhich rL.
penaltl does not o.-cur. a// links inrohin_e. on the a\erage, the same tir:
de1a1'.
Sec'trnC. nLrle rh¿t the proportional capacin as:i-qnment scheme erag-
geriliis tile djst:ncircr bet§'een lighr and hear¡ links e\en more. The oisp";-:.
Iq. +2. Alternate routing, lt's (messagr. rÉ :iM
.I:::
_. ISSiGNMENT IN DISTRIBUTED NETWORKS 7?
N ETWOR KS

: using ttvg
'yrsl,
-: =:ac1r) assignments is even greater than in the square-root assi-en-
-
rfim: r:r-:: ¿hrle the average time delay incu¡red is i¡creased some*'hat.
*':::*:. -,-::
:hat the equal assignment st¡ategy, while increasing the or.erall
J is simplr ---: delay per message, does reduce the differecce in time delays on
üui,,,-:-lE
-k. (In this
.riotted tc
Jltm r - r - 3ar) traffic links. Thus, comparing square-root and equal assign-
m r"-:
:..r:: r
=. *:s. the light traffic link 3 has had its average rime delay reduced six
-: - 106 msec to 36.6 msec. The corresponding increase in time delay
I
rfr -r j r:;',]'traffic link 7 is proportionally less, going from 22.6 msec to a
-:r-j .{s nofed above we strall discuss shortly a capacity assignment
&Í::-]*,.-:::at eliminates time delay penalties incurred by light users al- ¿
.r ,!ri -: i - ;.ih link time delays equalized throughout th.e aetwork.
{4-2)
* --": ,s now
sstgmng
the effect on caoaeify assignments and corresponding time
- :' :-anges in the routing strategy? Although routing will be discussed
-,::r : Chap, il it does pay to consider one example at this point. For
I
:,-::,:,ie we make
"just one change in the shortest distance strategy used in
t= ularicns above, Assume that Los Ailgetres-Hew York traflfic is now
.,: " :
: --: : ; E av of Denver ancl Chieago, rather than f{ouston, as iil Fig. 4.1.

r--t*-for the reader to eheek the foilo*ving one-way traffic flows


I
es)
" "-.-: :-r¡ this alternate route network:
propor- Ár: 6.1 :nessages/sec
I
tional

z4
&z: 0.51
A¡: 0.i3 (no change)
I
27.5
I A*: 6.58
I

28
,_
hs- 0.82 (no change)
6.3
30
Lc: 0.93
/0.) xt: 12"89

i
-.,l: *i :-*s are indicated in the network diagram of Fig. 42.
msec

rtze the New York


Co rhe
- e trme
those
re time
-hasa
hear-v
:h this
: ¡ime s - -'FE

Houston

I-e" +2. Alternate routing, lis (messages/sec) showri, ei:ber direction.


78 CAPAC¡TY ASSIGNMENT IN DISTRIBUTED NETWORKS .:::3 -Y ASSIGNMENT IN DIS'I'F 3 -':: ''-,*'
The total message flow for the network is now 7 : Z X¡: 28.1
: , ": ::e so-called Cheb¡ sher :: :-::.': -' :-
.-:
messages/sec (one-way), while the average number of links tiaversed by a 'r.¡: :urns out to be user-orien¡ed. Fc: . -.-- '
:
typical message is ñ ).ly :1.a7. The increased flow of traffic (from 25.12
-
) ¡,, :r",: -.:e iargest time dela¡ on ec! oii:: : r- ¡:
to 28. I messages/sec) and larger number oflinks traversed is of course due to :,*','j.at is minirnized,subjectto ihe.-:: .-tr r.:.
changing from the shortesÉ distance routing strategy. The overall minimum .::r.¡i-rr-r: ::rin-max grite¡icn TLe : ::. -. . ,!
time delay, with square-root capacity assignment, doesn't change very much :{ ¿ -:,ig ihis criterjcn,: io be eq-. - :. - i
in this case, however. A calculation shows that f,¡,: 43.I msec (compared to ":i -: .- ,ir€ ft --- 1 r'ase henc: :.- - - - -:
i
42 msec in the previous case), while the time delay with proportional capacity
- - :' - :: l :i rjeslgn.
assignment is 63.2 msec {compared to 54.8 msec previousiy).
¡ Although we have chosen to discuss link eapacity assignment for net- ]-:e rii:i:ni¿.ati,:* o{ T't.k: ct Eq - -- ::
works in the context of simple and relatively artificial examples, the techniques *'-- ., c-,;t r.ising',,rie lag:a:_.: - - -,É
outlined have been utilizsd in design studies for the ARPA network.* The -: -.--:rr=¡;ied
lrt:¡nxm cap*City;.ssigrim=- : '.' .-l'
t iril
approach used has been extended as v¿ell to inch¡de non-linear capacity-cost
Al
relations, nodal processing delays, and the rninimization of time delays other ,-.t
frtk\
- V,
than the ayerage time delay discussed thus far. trn the recnainder of this ¡:,f-:
chapter we consider a relatively simple yet useful extension of Kleinrock's
)r' is.
\) \./. -
work whicfu equ*Iizes the disparity between light and heavy users of a net- td
!t¿,-'.! - -
work, as noted earlier.*+ The reader is referred to the literature for other i -.
extensions that indude nodal processing delay and nonlinear cost functioni
that appear in this case.t In the next chapter we consider algorithms for taking
',r :-='- ic¡i : I l+,ehavethesquai.-:r. -:r;ú

ffi into account the nontíaear capacity-cost relations typical ofreal networks, but
l:. : ¿ramo'ter ,

specialized to the case of centralized computer-communication networks only.


The work on capacity assignment discussed thus far can be generalized
I 5,t with very little addition¿l effort in ttre following two ways:tt
i uq:i, *-r '". lhe miniruurr possibie netwLrii : - ,.
1. A fixed cosÉ D : f to be constrained. The s.tmmation is
d,C, is '-r. *i:', c-Bn possibly be assigned on l:i r

Fltl lu::¿: : ,:,n that link), and drA,l p, is the ::: : '

again over the illinks of tnfietwork, but associated with each choice of link
capacity C, is a cost d,Cr linearly proportional to it. (lf 4:1 we have the
:B '-!:.i.gnable pcrtion of the total cosi j
r- : ;neck let d,: l. This isjust the ;",: :
case previously dllcussed.)
i¡r :-, :"¡int in this chapter. Then
2. Choose C, to minimize
D* : E {).,ltt,): PC a--.: :
Tk, :lf f;rr,>rf"r
(+3)
ü*w:a[
I
.-É rarameters introduced earlier. -4.].,: -:
For k : 1 we have the average time clelay criterion already considered, for
k : 2 we have a ¡aean-squared delay to be minimized, and for & oc, we {yLrttltrz: I ¿-=
-
tI-. Kieiorock, *Models for Computer Networks," Proc. IEEE Infernationsl Confer- :-:'= Eq. (4-d)
enc¿ añ Commt¿nicatiors. Ju¡e 1969, pp. 21.9 to 21.16.
,,.'rrr-X,, tTTY:
*tB. &{eister, H. R. Mueller, and H. R. Rudin, Jr., *New Optimization Criteria for '
u..
lt¡ I. i
Message-Srvitchin-e Ne¡sorks," IEEE Trans. on Comnunicatio¡ts Technology. COM-19, no.
3, June 1971, 256-4J. u;:e Iq. (3-5).
iB. Meister, lI. R. Vueller, and H. R. Rudin, Jr., "On the Oprimization of Message. e--_: t .- oo: Eq. (+a) results in
Swirching Net*orks,- IEEE Trans. on Communications, CA}d-?fi. no. l, Feb. 1971, 8-14. -
tiMeister, foIueller. and Rudin, Jr,, *Optimization Criteria." cJ-):1,+l
' Ft P>: -=-.
TORKS
gTY ASSIGNMENT IN DISTFIBUTED NETWORKS 79
:28.1
rb so+lled Chebyshev or min-max criterion. It is this last criterion
Ibya ¡r¡s out to be zser-oriented. For with & + co a tirrle thought indicates
t 25.12 L- hrgest time delay on aDy of the links dominates. It is then this time
due to
:.tnu,lxl
much
' I is minimized, subject to the constraint on cost- This gives rise to the
fim min-max criterion. The time delay on s¿g! link turns out in this
lrsiag this criterion) to be equal. Light users are thus not penalized as
I
red to re in the & : I case, hence the suggestion that this is a user-oriented a
pacity d to desip.
¿
:r net- Jh srinimization of IG) af Eq. (4-3) subject to a ñxed cost constraint is
eiques It crried out using the l-agrange multiplier techaique of Chap. 3. The
* The
¡+ost
u optimum capacity assignment for arbitrary fr turas out to be I
A,
o ter /-<kJ
eJ
-
l
(#)
,f this PT

rock"s
a net-
l»lt#l'/(*+')l(l+"
I
Luo:lJ__i_t_rt__t__J
otLg
f10f¡§
aking wk :
<*rl
I we have the square-root capacity assignment result found
f
s, but
:¿rameter' ,
I
:-, only.
:1ized Do: D - D*, with D*:T.d-L
7'il¡ )
, ; --- ": the minimurr possible network cost. For the smallest capacity (in
-. -::: ,'an possibly be assigaed on link i is as already noted l,/p, (the
"r -r-r ¡n that iink), and d,1,1 p, is the corresponding cost. Do is then the
I
-?- "r-istg¡able portion of the total cost 0. {
'-r : :heck let d,: i. This isjust the case considered in Chap. 3 and up
: :L1itt in this chapter. Then
,* : T {X,lA,) : pC and D, - C{t - p)
+3)
-" - j :: parameters introduced earlier. Also from Eq. (+5)
lor
we lyLtr'1rrz:»,m
nfer-
c:: -:: Eq,(zl-4)
.t) - 7, - J(TJu) C(l -
v' p)
a for -V-r-Ei@l- r
!, oo.
:x'':-i.:¡ Eq. (3-5).
,iE:3ge- ;.-: É oor Eq. (+a) results in
L ,-r+. -
r.)-Lt,
vr 1 Do (+6)
- V,- -,8@i j n)
80 CAPACITY ASSIGNMENT IN DISTRIBUTED NETWORKS
r. ::,ACITY ASSIGNMENT IN DIsTR;ts--:: ",
-i
Note that with equal average message lengths (p,: p) on all links, this is \¡re that except for the hghr i::-:
essentially a constani pius proportional assignment strategy. It is intermediate ,il l. the min-max capaciiy assignme-.-., .:. i.
between Íhe equai assignment and proportional assignment cases. The :: -:i assignment case. This is becau¡; -; t: l
l.
I
n,,'
corresponding time delay on each link lurns out to be ::. _sir¡ in rhis example (p :0.1 3 \- . ,-
T(6)
¡i
1.'4 (+7) ,sl = : pclN, just the equal assigni--3: : :.: ,.:
L"", L/.,
- ua
^ i tl '--.: :3rm ls negligible.) For p .* l, pC = : ü
while the min-max time delay incurred by each message, on the average, is , i:ment strategy noted earlier. T:: : :*:
Fr, just -jr:ence on the light and heavy traffic i-- r.. ¡
T(-) : ñT*)' : *Z!t (+8) :;.. : :he equal assignment approach
fa,, : : ¡ _:r
D"*f F¡ i -:..;Jvely iow time delay), while pel. - : . :
with ñ : 717 t}..e ayerag§ number of links traversed by a n:essage. r, :,:;tii assignments equalize these
t* a :: :]=:
tVe cal apply these resuits to the network of Fig. 4'1 to compare with lie effect on the various use¡: .. :-.r.r
FF'I Specifically, we agai'n
the capacity assipment strategies previously discussed. ;=;:rcail."" in Fig. 4-3, taken from rhe :;:::-,:r
let p, - p a.ra tate pC : lg2messagesfsec.Letd,: 1' Then D.: C(l - p') rir" :: c¿se ? : A"25 a¡¡d
¡tC : i00.5 ¡¡e::¿.gg¡
as already noted and :- :,:"::d (i : 25.12messages/sec). k c.- . -:
)r" trc*t : x,+ &fA (+si :"r,¡l: :fe delay characteristic possible

:
i..: : -. ¡
:i --a'.ivel.v smali increase in avera_se :.:: :ñ
with N the total number of links (? in this case). we again take P : 0.1':" ü,::-: ¿ xnalty but this time delay c:..:.:;:r
Then for this examPle, 61?a-: deiay, cur.re.
: )u¡ * 23.8 messagesfsec l:: stress in this chaptcr as in Chap. -r - ",
tr . ttc:*)
N
ri':ver_»:m
q(*t J= : 42 -r""
1q"* -: averages"of them. The min-m".l :--,:
:

500
and
Fp
hI,
: Tt*) : 1.3(42): 54.6 msec

The capacity assignments and corresponding time delays, by link, are lis¡ed
;H in Table zl-3.
TABLE 4.3 300
- ne
CAPACITY ASSIGNMENTS, NETWORK OF FIG. 4.1 :E:av
(includes min-max assignment) -=¿c )
200
l¡ (mes- pcÍ1t pct*, ltQ Tt
sageslsec, m€s- Tltt (mes- f:*' (equal
(msec) sages/sec) (msec) assignme::
Link one-way) saeps/sec)

1 3.15 ?3 ñ.4 27 42 21.4


2 3.55 30 37.8 27.4 42 n.4
3 0.13 5 206 u42 27.4
q 3.U 30 38 27.5 42 21.4
t
0.82 13.5 78.8 24.6 42 27.4 Average
5
3.88 36.2 27.7 27.4
33.8 ^)
6 F¡. {-3. Time delay characteristic as furr.*.:c :
7 9.95 s{ 22.6 42 21.4
ñ 1. Hcster et al., IEEE Trans, Communko:¡,x,: f,
f i-) : 54.6 msec r aq!¡ ._
f-:" : 4l msec -
( -:- :
<? "h@ i971.)
t
.,
ETWORKS
ASS|GNMENT IN DISTRIBUTED NEÍWORKS 81

rks, this is
¡mediar;
. e
. : ::La: except for the light traffic links (3 and _í) and rhe heavy rraffic
:.ses. Tht
-: ::i:n-max capaciay assi_gnments are almost the same as those tor the
- , _;::rent case. This is because lle have picked a reiatively low traffic {
ihis example {p:0.13). Note from Eq. (+9) rhar for p(1,
(4-'.r,t : -,j..\', just the equal assignment case. (For p ( l, 1,, K FC,,andthe J
- .. negiigible.) For p * l, ttc:*t :1,,, essentially the proportional
: strategy noted earlier. The min-max approach does make a
¡n the üght and heavy traffic links, however. Note from Table 4-3 a
{,+g ::-:a1 assignment approach favors the light user too much (proüding
, .-::.,, low time delay), while penalizing the heavy user. The min-max ¿
* -. .;signments equalize these two extremes.
.
-pare witl

: c(r -" p:
we agail -'
-':" sfect on the various users of changing the criterion is shown
= . in Fig. +3, taken from the references.* The curves are sketched I
i -;1!e p : 0.25 and pC : 100.5 messages/sec as the total capacity to be
- :: ). : 25.12 messages/sec). lt demonstrates the /crge improvement in
(+e -:- :e ia,,- characteristic possible for the lightest load Qink 3), at the cost )
-:--:' '.31), smaii increase in average time delay. The heaviest
load (link ?}
,:0.13 " . :enaity but this time delay characteristic is süll always below the
"-.:
-::
:e1a¡'curve.
:i:ess in this chapter as in Chap" 3 has been on minimizing link time
I
: r!erages"of them. The min-max criterion, for example, minimizes I
500

Lightest load {línk 3)


f.
t
|]-r are listec

-:ng
Jelay
-rsec )
Tt
§qual
s.i:gnment)

¿ 41.:
¡oad {l¡nk 7}
" 41,9


t
t
:
365
42.'.
37.r-
l, ¡
I
Average
i**
Mín-max

.¡ ¿
s7.: Fg. .l-3. Time delay characteristic as function of criterion. (Courtesy of
B. Mcister et al., IEEE Trans. Communications Tecbology, COM-19, no,
. :¡ur) :
:
-,6
msec
:, Juoc 1971.)

KL
:SS 3\\'rE¡r; il'l : S-:3u-il \:-"
f Sa cAPACITY ASSIGilME}{T lN DISTRIBuTED
I{ETWORTS
r -.; ihe caPacit¡ oi er;h line :: ::': -
: ¡: the network is to
be 5 sec wrth ::
thctimcdclayonanl'liat.InthencxtchaptcfindiscrrssiogGapscityalloca- §1:all that
or"tta ti' ¿*t' wc focus on minimizingtbc
tions whcc capacity it;;;;i
P,i, ;;ilñ;;*gt ti-t debv lrom sou?cc to &s'tl,lr,t¡on' PC=81,
- --'.: shortest-path (least number oi '::"- :
PROBTEMS !.:=: traffic routed via Paris' ar: :2
::¡ied via FranKort'
l' conaccting ñvc Europcan citics arcrt
{.1 A map of a proposcd data nctwork 1 -'lag the same total capacity as in (1) susrg
aDDears in Fig.
p¿'t' iü'nentrator is uscd in Gtch of thc citilr to -*'¡:for the network if the routing
from terminals locatcd within tbc
sütistically multiplcx nc*gt' .:-= ton¿oo-Rome via Frankfort' anc
?

r
lr''
vicinity of cach- Thc tcrminals
message/mia. cach' düñ
output mcsssgc§ on thc averagc
;iüst b"sth is lü!0 bis' Thc
of I
tre6c
thc avrragp numbcr of ncssrsÉr!ú¡"
matrix for this aetwork, indicating
-. London. ComPare with (1)'
' ::
I '- . .:le capacities in the network are : :' ' :
.:-- effective P,and averag3
capa-ity,
ü;" any tso citics, is as follows: "':-.
_ 11,"_r\,€.
1¡¡ (ME§SAGE§/MI!{)

F' l+ tz
,¡ t ,Ol0?f.il
2 40loato
lolo?olon
h 3
1
t
¡t1¡
Ilolol0
lO

lb ponm:t thc citics ¡¡ ¡hown il thc up'


kFi¿
Fullduplcx

F 4r,srg :: ',he generalized capacity assignme:.


u fixed cost linearly propi
rm*.--::-1¡'. *itt
cefin:
ilñ=..-on eacb-link, and time delay
given b'r i
ll.-=t- capacity on each link is
rl--
4r.:-::'r. ;riterion: Let k' oo in Eq'
*ui;--..... on each link is given b;' Eq' ':'5
,,* ' - ,t',,. and given bY Eq' (a-7)'
{r;-: :-: min-max results of Eqs' (4-6) t'
:+ -. f:nd the capacity assignments a:'-
'v, -::-1 ;:mparing with the minimur:' "."
**,' * , .;... ttitn the entries
of Table '1-: '
*',*". '.,,-on in that table as well' j:
is:- : :.:e 5'citY EuroPean o€t\\ors
ñomc riltrx]iirr-
-. :.signment, by link' using a :: --'
Fla¡ Pat

k
83
xETWORTS
; : * ::: :apacit¡ of each line if the minim-m avera-se time dela1. I

dloca- ir¡"--r nerqork is to be 5 sec rvith total capacity C fixed. (Ibre:


izingrhc
;--51 thet

'¿C: tAF¡l
&\ (I

- re tr rrtest-path (least number of links) reuting, with London-


i :re trafñc routed via Paris, and Parl-Amsterdam traffic
dticr -:rl::'j tia Frankfort . .,,e+:i
citirx to
üthin -uug:he same total capacity as in (l) abottn find the;mihimum time
thc uea.¡ for the network if the routing strategf now routes messages
ofl :-:a
Thc tnfrc ¿
X-,rndon-Rome via Frankfort, and Paris-Amsterdam trafrc
;¡cion. Compare with (l). f
t :E capacities in the network are chosen to be 1200 bps. Find the
;ci =paciry,
effective p, and average time del;ay, andcompare with
::!\ e. J
:*T', ::- capacity allocations and corresponding average time delays
*u:e .{-2 under the three different capaci§ assignment strategies
;-=-:. ,
*:in,r rhe routing strategy for the network of Fig. 4-1 that routes Los
.r*e;-\e*' York traffic by way of Denver and Chicago. Verify the I
lr-1,i: tráffic flows shown in Fig. ¿1-2. Sho.w that the minimum
,r"a4r :irle delay for this network with the flbws shown is 44 msec.
r-, ¡=,:ies propagation delay. Indicate how you might take this into
r' :::: ¿ad c¿lculate the time delay including propagation effects.
r.i,-: :E o possible propagation delay estimat¡s: l0 msec/1000 miles,
; -- :.¡ec 1000 miles.
*g :: the generalized capacity assignment aralysis of this chapter.
",u::-11'. with a fixed cost linearly proportiaaal to the sum of the
r:r:rr:€§ on eacl¡link, and time delay defined a. in Eq. (4-3), show the
*-:::- capacity on each link is given by Eq t4-4).
.:-:¿r ,-riterion: Let k- oo in Eq. (4-4). Siow that the capacity
r--:.ri on each link is given by Eq. (a-6) wh:le the link time delays
: : - !-:-iland given by Eq. (a-7).
:: :-¡ min-max results of Eqs. (4-6) to Eq. rr-9¡ to the network of
= -. Find the capacity assignments and corr:.ponding time delays,
comparing with the minimum avera-si delay results. These
-{-.
: ¿e:ie ivith the entries of Table 4-3. Chec.: :re equal assignment
- *
* --. :her\ro in that table as well.
s:- :r rire 5-city European network of Probl;:¡ -1.1. Determine the
--::-. assignment, by Iink, using a min-ma.r ;::terion, for the same
CAPACITY ASSIGNMENT IN DISTRIBUTED NETWORKS
-
total capacit! as in part (l) of that problem'\§ompare the resultan:
ur..ug.',i*. delays, Uort'itt link aná overall áverage' with the minio;
assignment result
l** mum time d.t"y ..ruit.-óátnputt with the equal
part (3) of Problem 4.1'
L-

)t

Fil
i¡rolized Networks:
tFn
Delqy-
)r" Trode-offs

ul :r 'r:..-iry assignment discussion '¡: :--: -


H luüluiie: ¿ :;chnique in which average m;:::= :

r,mu :i--¿--it) held fixed. A variation oi:. .' 'i

- §:-rles cost linearly proportiona- : : ':


- :: :he average time delaY with :--. - '

IT - ::3ropri&te design approach ]n r-:-


' -: :.al\\'ork operation. It enables t:''.--. :
h6 : .'r- , r complex networks, it allo.'t. :::
;:-..iueS chosen from a fixed nu;::":
*- -::,::ment, among others, and is :e":
ll, . j
-' - -,-1 ¡ccount, as was done in the ':t:
''::rilrL, -": : - - .:h rs at ihe sam.e time unre a1,.; : -

:: '- . pla,vs the overriding role arir .

r - -, - .-r:ar. Examples wouid i¡iciuc. ^


1r' - - -.. lo the capacity, the reia'.''. ;
-:.:i the cost of a Particuiar lr:' '
..:..:les: caPacitY, use for r"il..'
"" : 1:- --a not part of a larger piCr':-.;
i: ::: . etc. (In the United St:.:
- - ::ate to state, and intrasi":=

\1, Gerla, "New Li¡e Tarif. ":. - -: --


. National ComPuter Coi':"::-: ' ll

Das könnte Ihnen auch gefallen