Professional Documents
Culture Documents
utep
Sl
oynodc noutlng to
PaEll.l
H.tcrlganaous SGryeFr
lhPtln ttEnllc
llay 20, D88
EilEE 426
Pr!.
EPhEr{d.t
Table
of
Contants
I.
lntroductlon
Revieu
I
2
7
tl. III.
lu.
v.
of Litcraturc
Ih Slnulation
15
t7
l8
Ll3t of Ftgurrs
1.
2
Parallcl H.teroglneous
Qucue Splce l
I
8
3. 4.
5.
D.
lhe nlr.ll CooDut$ Progrm The Parallcl q!.uc llct.mganeous Sery.r Cdl?utrr Progr Slrulation
Run
9
L2 13
I Results
7.
g.
74
15
ABSIMCT:
Consider two
parallel
heterogcneous exponentiai
seryers.
Upon
its
type,
between
dispatcher sends a customr to the slower gueue. Thl3 system is sinulated and eyidence yas found to show that this optlnal po'licy has mrit. As a guidelin
for further work, a speciallzed case of thls problem l3 traced form concepfion to solution. Th! techniques used are erari ned for utllity ln solving the problem
at
hand,
iI.
II{TRODUCTIOI{:
ic routing
problem. Tt{o
parallel
heterogcneous
to a dispatcher that rust assign the customr to one of the queues. The routing problem is dyndrfllc in th scnse that the dlspatcher n6k!3 its declsion
strearn
bEsed on
optlmal routing
tim of the decision. tt ls proposrd that the stratrgy is of the threshold type. lncofiing traffic is sent to
first, faster, queue until the queue reaches a level that is greater than that of gueue 2. At that point a custoEr is sent to thc slorr gueuc 2.
the
iEPABTUFES
N,
AFBIVALS
DEPAATUFES
sinilar problerE. section II rvl ers appllcable literature for thls sp.cific problsn as we]l as the related problo of ltlple servcr slngle queues. In both cases
A substantlal body
devoted
of r.search is
to the thls
and
the analysis of thc horFgenaous serv.rs crse preccded the analysls ot thc hterogeneous servers because
crse.
Tha
nrltlpla
is
considered
This paper
will
ls optiDal.
sys-
sirulation of the
problem
tem
behavior.
III
to solve various
aspects of
II.
REVIE| OF LIIEMTURE:
Kinqman
[Klj
ana'lyzed tno
identical servers
each
arrlvll
the cus-
If
1/2.
Kino6;n proyed
is
less than
l.
He ana'lyzed
the rys-
in
continuous
tlrE
for
havy
traffic
conditions only.
th.[.
He
lel
tims.
for thrce different cost crlterlon In extendlng his re3ults, re are not concerned rith cost crit.rion involvlng storage costs. lhen averag. d.lay is thc only cost criterion, Chen providca an
He proposrd
optlmal pollcls
ll
servers.
Chenrs
Yinston
[I3],
[]r41, [l{5] analyzes three varlations of the two server sysa The
l.larkov decision
process.
technique
is due to Lippnan [13] and is descrlbed below. llinston [f5] proved that for a finite nunber of identical exponential servers the send to the shortqueue
est
number
vice during any tirE epoch. He proved this in both discrete and continuous
tim.
lJinston proved the discrete case using a new mthod reported by Lippnan
[13].
ing
The core
queue-
systems
tirEs.
The
dcfini-
transitlons
back
have constant
is satisfled if transition
lllowed.
is useful in finding the optirEl policy in terfis of an [[5] results to the case of a stochlstic arrlvof stochrstlc an'ival
proccss
al
and a service
rultiple parallel s.rvcr systrn with the basic Poisson and axponentirl 3tat'lstlcs, They do not use Lipp|lEnrs mthod. They first study nondeterfll ni sti c routing. Unlikr Chen they asslgn custoners in inverse proportlon to the srvlce rates. Larsen [11] strte3, aPp ently
chow and Kohlar
the
Chor{ and
Kohler's.
Chow
et
al
mi
propose
They make a
stake in assuming that balancing the load afiong processors ri/il.l result
in
the
lowest turnaround
tirE.
The
type. This can readlly be seen by examining the state transition diagrans of each policy. The state space is diyided into two parts (Flgure 2). If an
arrival encounters the
queue
system state
1. If
2.
The
deteninistic policies
THRESHOLD
FUNCTION
o l! o
ttJ
o SIZE OF OUEUE
1
Spacc
that
to inD'lerEnt
dynari
c routAt
ing.
have varylng
anEunts
of infomation avallable
statlstics.
this point
Apparently the nore infonflation a policy has and uses, the better
it
per-
fonns.
intend to
exa,ni
ne
his work
more
parallel
the service rates are identica'l and exponential. Their contribution is to show that the optimal policy is independent of customer arriv!.1 statistics. They
prove two optimal strategies, each applicab'le depending on the available infor_
mation.
the gueue lengths are observed then the optirEl po.licy is to route jobs to the shortest queue. lf queue lengths are not known it is optirnal to alternate jobs provided that
example
If
initial
by
that a po'licy of
sending jobs
service time
is not optirrBl.
policy does not balance the load between the servers contradicts som of the conjectures found in ear'lier papers, specifically Chot{ and Kohler.
at this point in ny research that t proposed a hypothesis to test by corrputer sinulation. At the end of the paper by Ephrerides et at,the authors consider the heterogeneous exponentirl server case. They state, ,,tt would be
was
It
interesting to find a 'sinDle, pollcy r{ith provable desirable behavior." I propose that the threshold po'llcy would be that sirDle policy. lhlle I was collecting data fron the sirulation
extend previous rsults. Bertsekas [81] proves the existnce
using
dynanic progra,nming. He restricts the result to the tuo server case with poisson
tim3,
The cost
function
used
is the total
cost.
to
the
in Jigwe 2.
one
is optinEl.
arDng
He does
all
dynami
traffic statistics.
servers.
is left to do.
this by using
LipFnan's
is of th threshold type. Thy prove technlque [13]. The optimrl policy is proven for both
iterativ.
nEthod
stlte
to
be a useful approach
exani ned
ln greater
dpth.
tried to not only report tha r.lrvant work that has been done. I hrye also tried to hlghltght som of thc succassful techniques usrd to arriyc at thc rasaarch$s rasults. At sqlt point co0?uter sirulation ceases to be useful and rnalys'ls of the systa rust proceed by folloring onc of these or slnllar rjgorous mthods.
have
IIT.
TH SII.IULATIOil:
A.
Hypothesis: There
will
be minimized
for
tlethod:
chose
to tinE step the siDlation rather than event step, This made the
tim.
arrival
process was
cess. A knom
slot
In
queue (M/M/l) was sinulated to denrnstrate the validlty of the size used in th sirulatlon. -Ihe progran was run long enough to achieve
a reasonable steady
any one
state.
ThJs was
tim slot only one event lras allowd. Upon occurrence of any event of interest the progran advanced to the next tim slot. At each tlm slot the progran lookcd for an arrival. lf therc was an arrival, total arrlval
guue
thr dispatcher
of ties
decided which
to
send
th. arrivrl. If
old then
custor:r. If
there was
depirture. Sirultaneous departures are not allowed. At the end of the arrival proce3s averagr queue slzc is con?uted as well as averaga delay. other statlstics are not calculatcd becausc of the lncrease in run tlnE that such calculations would produce.
The averagc delay
for customrs ranaining in the queue. the ll4lL queue was silllllatad in order to valldate the choice of expectcd arrlvals and thc choicc of slot slze. Agrin sirultaneous arrivlls and
allolred. Flgure 3 contains th prograrn llsting for th l'1lM/l sirulation. Figure 4 contains the progrdn llstlng for the parallel queue
departure3 are not
heterogeneous sePYer systen.
'
r{^iTtN
', EXaa
a6
' fils Pnoeiri struuEs Ar r/r{/t QUEUE ' nls strutrrllr Is uso r0 vrl.to^rl ff '
rxo rx ctiolcE
0F sLoT stzE,
ot0tcE
6PESE0
rutEi F lrtlv
Ls
, i.LrlflA , EXPEIqIE|) ' slrul Tl0l SLOT SIZE ' IIITIr! ct ITIOI
'
tr ln8Mts
'C:Uil,PFl' Foi rPPEl0 As ,l ir, D, il, ur80r r 0, rut r cltsE,t FoiC.11050 oPEI 'C;\itl.tnr' fol lPPElo rs ,1
PFIt{t 11.
xr.f,Sl,l/C:IlT.0:l-0
RAootilzE TliR
FOlJ.lr0i
lF ,(l <
Fl)nct-1t00
r ri
J
E0T0
x. t+ 5lt
1:
tl .xl i t:
e0T0
50
occlntEl)
dt,
ooqXto
$!li.Jor/(tr0)'Isl,r
. l), lsl| /
'
'
t{EIl
426
20,
lt$
ru
sR|Eis Tr sEprurE
Q|EUES
Prot,
EPHETII0ES
'
rBls PRtrilr
slir.ltls
A HErlm6t|Eous ^IG0RITH'
R0IJTIN6
rLeonITHi.
Is 0t tr THtsll4o ttPE,
It'10000 D . 1000
L,Jl,o .
I|PUT'lrlruSH0-o't
L.rz l0t Nl - t.5 I 0t llt2. .t t 0 oPEr '!:\123\HElml5.Pir' Foi at rt ^ppEN0 Pilft rt, I, 0, 1 , lrt80 r 0, ,1 . t, rlE.0
closE
FOic.lrc50 0Pl c:u23\fln011-Pir' tLl .0. rzf .0r t.0 nr[ortzE llti FotJ.lT0 I
FoRd!'1T00
lF
IUO
,l
Fff lttro al
,t
'
r 1:
CltTo
l)lln ll{ lF
^rtlv
l,
oCc1rtrt
QoEtf,
xl . xl I
l0 IF
,(1 <
IF Xl > 12 1
l:
T{ TtEt t2
e0T0
X2
50
[r.t;lt+It
IF f,ro
ttEt
4n
l0
t
6010
tF Ql
tol
EprY
i0
ifigl
50
'lF
'
'
DEI I)IIFIIE
q2
If
lE
lrruil
oCCllSEo
lE
qt
rcTEim
Q2
12
50 FXT
.xlt/ (rr D) l llslll: ll^vl.[T / (t.D) t2sur. xzl / (r.0) + r2sur: /wa. tzt I $.0)
xlcll
'
'
cl|furE
F08
qf,lf,
tzf.tzf +xZ. U2+1, I l?'ru?) l.(I1T.t21)/(lr0) r6!i - lSUr r I tilft ,1, x1^v5r a, t2*tat p, i, v, tslJi / c CL0!E rl
rJfi
C
SIUE
I0 ulE
^5
IIIIII
LEF
'
lxcutr Exttglt!
rlrf c
0F oUY
cutturEls
' ,(l ve . AvEillE Ql sIzE llrls iul ' X1 . CUiIrrIYE AVEi.l.t qr SIZE
Th. Parallel
Quaue HeteroEneou3
C.
Equipmnt:
The progran was
t.iritten in the
BASIC
languagc. lhe
QUickBASIC
4.0.
is
conputer
Even so
it
too
executes 5,000-15,000
BASIC
lines of
it
was
accunulated over
D.
Results:
detendned that
it
was
necessary and a 0
of
l/0 is
Lanbda
' ,9:
l1U
= 1.0,
Percentagr Error
20.171 27 .0W
-7
Expect.d
M-1000 D=10000
Queue 0el ay;
observed
10.81 12.70
Size;
slze3
8.34 10.29
.3* 2.9*
that from the data that thc tength of the slrulation ls tbre
crlti-
10
The
first
i|.1000 and 0-10000. As figure 5 shor{s the average queue size was not well
behaved which
indicates that the average computed de'lly fop the system is also
worthless,
The second run was rEde
at
tl=lOoOO and
Iith
higher average
size.
The
ls still
a
state.
But we do see
function. Ihrt
we would
ty.
For the
thlrd run
changed
lpparently mved
it
7).
However
ir
E.
Conclusions:
Though the
an
optlrEl threshold policy, the r.su'lts indlcate that thls night be so.
results are ccrtainly strong cnough to rDtivat. further
probleo
si lation.
Thc bas.lc
is to progran th. sirutation in 'C' using Sorlandrs Turboc, Though the ilprovemnt in sped of executlon n|ay be no
tily next step
is cotDutatlon spled.
to 32 btt mrkstations. In addition, given rbna proccsslng powar I would generlte y oyn randon nu[6ars rlong the lines of Coates, Jrnacek, and Lever [C3].
be transportable
wlll
11
Q2 SlzE
AVERAGE DELAY
UJ
o
lU
(5
trJ
,/-...\.'
I
a IL o
N
UJ
f, IU f
IU
THRESHOLD
o rU a
(E lrJ
Lu
(\,
E
c ;l
a lr o
o
N
lU
f UJ f,
lU
L 5
10
THRESHOLD
UJ
o
(5
ltJ
tu
lU
:)
trJ
E
Itt
o LL o
IU
f,
N
U'
.l
THRESHOLD
IV.
of the useful
technrques
far to
so'lve
probiem
is in the process of
being solved.
I will
ou
list
the techniques I
that
to
achieve.
. EF;FTtBES
ABP VALS
DEPAFTUBES
Queue
problem discussed
shows
The
buffer is srved by
two
tims.
first
posed by Larsen
[Ll]
who conjectured
the
c
threshold
dynarni
programing and LipprEn.s method [13] mntioned prevtously. falrand vided an alternate proof using
[tl]
pro_
len using two different cost criterlon, discounted cost and average
extended the
cost.
and
He
l5
ffriya]s,
exponential faster
server and Erlangian slower server. For the di3counted cost criterlon using
tilE model he was able to shoy that the optinlal policy is of the threshold type. For average cost criterion he was on'ly
both the continuous and discrete
In light of the
work
success
that
approaches
to the parallel
system wou'ld be
Strategi es:
l.
oynami
all
tlm.
n as probabillstic argupnts.
proposed by Coucoubertls and HaJ'ek.
2. 3.
knot
Iterative mthods
such as
that recently
l{i I estones:
l. 2. 3.
Prove
that the threshold pollcy is thc optirEl policy for tro seryers opttnality for nbrc Acneral arrival
and departure
and
statis-
ti cs.
Produce
16
v.
colicLUsIoil:
is being conducted in the area of dynard c routing, yet there is roon in this field for more endeavor. I have atterDted, through this
Substantial work
paper, to acquaint nyself with the current state
of research.
Obviously som
relatively
new and
I will
haye
tor.
At this point
was able
17
Bl bl
lography
[81]
Cltffs, N.J., 1987. [C1] Chen, P. S., optinal file allocatlon in rulti-levcl storage systens,
Models. Prentic-Hall,
Englewood
tC21
Choyt, yuan-Chleh and Kohler, tJalter H., ibdets for dyndri c load balancing in a heterogeneous llrl'ltiple processor system, IEEE Transactions on Corputers. C-28, 5 ( ay 1979), 354-361.
[C3] [C4]
Coates, R. F. f., Janaceck, G, J., and Lever, K. V., llonte Carlo Sirulation and randon nunDer generatlon, IEEE Joumal on Selected Araas in Comunications. JSAC-5, (January 1988), 58-66.
Courcoubeti 3, C. A. lnd Reiman, i'1. l.,optinal control of a queueing systen t lth sirultaneous service requireflents, IEEE TransEctions on sirultaneous ervrce requ requrreflnts, Autonati c Control. AC-32, 8 (Augu3t 19871, 7L7-727.
!l!
and
tlalrand,
J.,
slrDle
dynrnl
c routing
[K1]
J. F. c.,
in plrallel,
Anna'lr
of
eth.nati cal
13L4-1323.
[11]
Larsen, R. 1., Contro'l of nrltlple exponentlal servers wlth applicrtion to corDuter syster6, Ph. 0. thcsls, Univ, of I'laryl rnd, Conputer Science
and tunar, P. R., heterogeneous servers, (August 1984), 696-703.
@.
N-29,
queuelng
687
LipFEn, S. -7t0.
system3,
4.,
[Rlj
[S1l
Ricart, G., Efficient synchronization algorltlllls for distributed systens, Ph. D. Thcslt, unlv. of ilaryl and, Con?utcr Scilnce DePartlEnt, 1980.
Sasaki, G. and HaJek,8., optirEl dynrni c routing in single cotrEdity networks by lteraalve nEthoas, :E Transactlons on comrnicatlons. c0l1-35, l1 (t{ovclldcr 1987), rr99-1206.
[vl] vinlotis, I., optirbl control of a nultiserver tlaryland,system with - - heterogeneous s;ryers, il. sc. Thesls, Unlv. of quuelng Electrlcal
Engin..ring Departmnt,
1985.
18
l-Hll - '
Ha'lrand. J., A note on "ODtlmal control of a queueinE systen with two heterogineols servers," svstems & Control Letiers. 4-, (1984r, l3L-134.
assignmnt of customrs to parallel servers, Journal of Applied Probabilitv. 15, (1978), 406-413.
fr3l - '
finston.
f.. oDtirbl control of dlscrete and continuous tim nEintenrnce systems'with v;riable servic ratcs, 0!gIEf-E!5-8C5!i!fC!. vol. 25, 2 (Harch-Apri I L977), 259-258t977),
469-483.
[fl4] finston, U., Assignmnt of custoners to servers in a hetrogeneous - ' queueini syiten with $ritching, operations Research. vol. 25, 3 (l'lay-June
[f5] - -
linston, Y., Optinallty of thc shortest line disclPline, Journal of Applled Probabilitv. 14, (1977), 181-1e9.
19