Upload
truongthu
View
213
Download
0
Embed Size (px)
Citation preview
A0AII? 915 NAVAL POSTGRADUATE SCHOOL MONTEREY CA F/6 9/2DISTRIBUTED VS. CENTRALIZED DATABASE SYSTEMS - TRANSACTION EXEC-ETC(U)
I NP5 2- D Z BADALU L A S S I F I D 5 - O 0
U lllIllEhEEEiEEEIIIIIIII
NPS52-82-007
2 NAVAL POSTGRADUATE SCHOOLMonterey, California
DISTRIBUTED VS. CENTRALIZED DATABASE SYSTEMS -TRANSACTION EXECUTION COST AND PERFORMANCE ANALYSIS
*Dushan Z. Badal
July 1982
uJ
Approved for public release; distribution unlimited Ti c |
8AUG oo2
i 2, ...
NAVAL POSTGRADUATE SCHOOLMonterey, California
Rear Admiral J. J. Ekelund D. A. SchradySuperintendent Acting Provost
The work reported herein was supported in part by the Foundation ResearchProgram of the Naval Postgraduate School with funds provided by the Chief ofNaval Research.
Reproduction of all or part of this report is authorized.
This report was prepared by:
DUSHAN Z. BADALAssociate Professorof Computer Science
Reviewed by: Released by:
DAVID K. HSIAO, Chairman WILLIAM M. TOLLESDepartment of Computer Science Dean of Research
J ............
UNrCLASSIEJED.._____SECURITY CLASSIFICATION :F THIS PAGE fWhen Data Entered)
READ INSTRUCTIONSREPORT DOCUMENTATION PAGE BEFORE COMPLETING FORM. REPORT NUMBER Z GOVT ACCESSION NO. 3. RECIPIENT'S CATALOG NUMBERNPS52-82-007 ,-
4. TITLE ,'nd Subtitle) S. TYPE Of REPORT & PERIOD COVERED
Distributed vs. Centralized Database Systems - Technical Reporttransaction execution cost and performance . PERFORMING ORG. REPORT NUMERanalysis
7, AU THNORf) 0. CONTRACT OR GRANT NUMIERP()
Dushan Z. Badal
9. PERFORMING ORGANIZATION NAME AND ADDRESS 10. PROGRAM ELEMENT. PROJECT, TASK
Naval Postgraduate School AREA & WORK UNIT NUMBERS
Monterey, CA 93940
I I. CONTROLLING OFFICE NAME AND ADDRESS 12. REPORT DATE
Naval Postgraduate School 13. July 1982
Monterey, CA 93940 NUMBER OF PAGES
)4. MONITORING AGENCY NAME & ADDRESS(II differont from Controling Office) IS. SECURITY CLASS. (of thle report)
UNCLASSIFIEDIS.. DECLASSIFICATION, DOWNGRADING
SCHEDULE
1. OISTRIBUTION STATEMENT (or thls Report)
Approved for public release; distribution unlimited
17. OISTRIBUTION STATEMENT 'ol the abstrect entered in Block 20, it different from Report)
IS. SUPPLEMENTARY NOTESA part of this report has been published in The Proceedings of the Second
International Symposium of Distributed Databases, September 1 - 3, 1982,Berlin.
It. KEY WORDS (Continue on reverse aide If necesa'y and Identify by black number)
Distributed databases, performance analysis, cost analysis, long-haul networkslocal area networks, distributed computing.
20. ABSTRACT (Cantlrvue an revereo aide it nocecaep and Identify by block number)
--The purpose of this paper is twofold. First, we investigate the impact ofconcurrency control on transaction execution cost and system throughput incentralized and distributed data base systems (DBS) based on slow and fast(local) networks. Second, we show that in terms of transaction execution costand OBS throughput there are some applications for which any distributed DBScan be more effective than any centralized DBS and vice versa. We also arguethat for other applications the decision in favour of distributed or central-ized DBS should be based on the comoarison of specific DBS systems.
OD pm 1473 EDITION O I OV 65 IS OBSOLETE UNCLASSIFIEDN 0102- LF- 01.5601 SECURITY CLASSIFICATION OF THIS PAGE (fle. Date 111tnod)
Distributei vs. Centralized Database Systems
transaction execution cost and performance analysis
Cushan Z. Radal
Computer Science Department
Naval Postgraduate School
Monterey, CA 93940
Nbstract
Tne purpose of this paper is twofold. First, we investiate the
impact of concurrency zontrol on transaction execution cost anti system
throurhput in centralized and distributed Jata base systems (FAS) hasee
on slow and fast (local) networks. Second, we show that in terns if
transaction execution cost and D3S throughput there are some apolica-
tions for *hich any distributed CBS can he more affective than -ny cen-
tralized CBS and vice versa. t,;e also argue that for other applications
the -'ecision in favour of -istri'uted or centralized 2-S shoul , he 'ase
on the compnarison of specific OBS systems. Accession ForINTIS ($!A&IDTIC TABUnaunouico d
Dist ribiit ! on/Availnbility Codes
Avnil and/orDist Spccal
1. Introduction.
Distributed database systems (D-OBS) are alleged to provide numerous
advantages over centralized database systems (C-CBS). The usual argu-
ments in favour of the D-DBS are:
a. im-nproved user attitude - the distribution of data and process-
inq gives users greater control and autonomy over the Jata processing
b. improved reliability and availability - because of the parti-
tioning of data and replication of processors and data
c. improved extensibility and modularity
d. decreasing cost of hardware should make D-DPS cost effective
e. D-DBS could provide better performance and perhaps lower tran-
saction execution cost because they have inherent concurrent execution
capabilities not available in C-CBS
'Xe expect that in many applications the last consideration, i.e.
transaction execution cost and performance, will be the principal factor
when dec.ding between C-DBS and D-OBS. Therefore in this paper, we
analyze and compare the transaction execution cost and the performance
of C-CBS and D-0BS. We are here interested in two goals. First, -we
want to investigiate the importance or the impact of concurrency control
on transaction execution cost and system throughput in C-OBS and O-DBS
based on slow and fast (sometimes called local) networks. Second, we
ir3 interested in a simple and robust analysis which explains certain
intuitive notions about the preferability or suitability of D-DBS or C-
DBS for some applications. Our analysis is general , i.e. it is not
meant to represent any particular concurrency control mechanism, C-DS
2
or D-OBS. The analysis however can become specific by substituting
proper values, representing specific concurrency control mechanisms or
DBS systems, into the derived formulas.
2. Previous work.
It appears that there are no published papers dealing with compara-
tive analysis of distributed and centralized DBS systems. However,
recently some work has been done on the performance analysis of distri-
buted DBS (BAD 90, mOL 79, RIE 79) and there are two publications (BUC
79, STE 79) somewhat related to this paper. The paper by Fucci anl
Streeter (BUC 79) deals with the cost and performance analysis of a sin-
gle processor with multiple remote terminals. The paper arso contains a
short sketch of a very simple comparative cost analysis of distributed
and cen, *',ized DBS. The paper by Stewart (STE 79) describes a discrete
simulation modelling tool developed for the performance analysis of IBM
SNA system configurations. It is obvious from the paper that this per-
formance analysis tool can be used for already implemented systems as
the simulator input requires detailed system implementation information
e.g., number of instructions per program, number of I/C, a priority
interrupt mechanism, a nriority dispatcher, CPIJ lefinition, etc. How-
ever, it is indicated in the paper that the simulator could be use. for
the performance analysis of distributed OBS - presumably once it has
been implemented.
3
3. D-PBS vs. C-DBS: Transaction Execution Cost Analysis.i
In our analysis, we consider C-DBS with remote users and a D-DBS
with similar capabilities in which data and processing power are distri-
buted to the remote users. Figure 1 shows both CBS configurations we
analyze in this paper. (RU is remote user and LU is local user).
Q.UQ
4
i .7
In this paper the cost is defined in terms of the number of exe-
cuted instructions, the amount of generated I/O and the number of mes-
sages, or any subset of these. !e model the average cost of one tran-
saction processing in the C--3S as consisting of three oarts as Follows:
Cc =Ccsys +Ccom +Csyn (I)
where
ecsys is the average cost of executing one transaction in the
C-CBS without a concurrency control (CC)
Ceom is the averaqe (communication) cost of submitting the tran-
saction from the remote user to the C-CDBS
is the average CC cost, i.e. the averaqe cost due to syn-
chronization of the transaction
Since any transaction in any OBS is either conflicting (i.e., try-
ing to acquire resources already acquired by some other transaction) or
nonconflicting then Csyn consists of two types of costs - the cost ((cc)
associated with nonconflicting transaction and the cost (Coonfl) when
the transactions interfere, i.e., conflict. weL can express Cs as fol-syn
lows:
1syn DI Cconf 1 ,7cc (2)
5
where
DI, the degree of interference, is the ratio of the number of
conflicting transactions to the number of all transactions
Cconf1 is the average CC conflict resolution cost per conflict-
ing transaction
DI*Cconf1 is the average conflict resolution cost per transac-
tion
Ccc is the average CC no-conflict cost per transaction
Transactions in D-DBS can oe divided not only into conflictinq and
nonconflicting (as in C-DBS) but they can be also divided into local and
nonlocal or global transactions r epending whether they execute at one
site only (local transaction) or more than one site (global transac-
tions). Thus the cost of transaction processing in the distributed DBS
(D-DBS) can be modeled as the cost of local and nonlocal (or global)
transaction executions weighted by the terms reflecting the number of
local and global transactions in the system. Then the cost CA of the
transaction processing in D-DBS is:
Cd = X*(Clsys + CIsyn) + (l-X)*(Cgsyn + Cgsys (3)
whereClsys is the average cost of local transaction execution without
considering concurrency control
clsyn is the average CC cost per local transaction, i.e., one
which needs to access data only at one site
Cgsys is the average cost of global transaction execution
without considering concurrency control
Cgsy n is the average CC cost per global transaction, i.e. one
which needs to access data at more than one site of the D-DBS
X is the ratio of the number of local transactions to all tran-
sactions
l-X is the ratio of the number of global transactions to all
transactions
Let
CgsYS = Clsys + Ceata (3a)
where
Cdata is the average (communication) cost of data transfers dur-
ing the global transaction execution
C1 syn can be further decomposed in a similar manner to Csy n in the case
of C-DBS:
C1Sy n = D11 * Clconfl + CIcc (4)
where
DI1 is the degree of interference of local transactions at each
site of the D-DBS, i.e. ratio of the number of conflicting local tran-
sactions (i.e., local transactions conflicting with local transactions
7
*--!~i
but not with global transactions) to the total number of local transac-
tions
Clcc is the average CC no-conflict cost per local transaction
Clconfl is the average CC conflict resolution cost per local
conflicting transaction.
Cgsy n can also be decomposed in similar manner as follows:
Cgsyn = gcc + 01g * Cgconf1
where
Cgcc is the average CC no-conflict cost per global transactio
"gconfl is the average CC conflict resolution cost oer q1obal
transaction
DI is ratio of the number of conflicting transactions (i.e.,
global transactions conflicting with global and local transactions) to
the number of global transactions
Substituting into (1) and (3) from (2), (3a), (4), and (5) we obtain
Cc =csys +Ccom +DI* CconfI + Ccc ( )
= C!Sys X*(DI*lconfl + Clc c ) (l-X)*(CI *Cgconf+ C+cc +
Clata) (7)
i3
In order to simplify (f) and (7) we assume that the cost of synchroniz-
ing local transactions in O-CBS is proportional to the cost of synchron-
izing the transactions in C-DBS, i.e., we assume that:
Clconfl 1 * C conf
Clc c K 1 * Ccc)
where K, # 1 if the following, applies:
a) processors in D-DBS can not support the same CC mechanism at
the same cost as C-CBS, e.g., number of I/O is different, etc.
b) D-OBS uses CC mechanism different from the one used in C-7PS
c) when both of the above apply
We can also assume that the cost of local transaction execution
without synchronization in 2-rBS is proportional to the cost of transac-
tion execution without synchronization in C-CBS, i.e., we assume:
C2sys 2 csys (?)
where '2 # I if processors in D-DBS execute local transactions (without
CC) at a different cost compared to C-CBS.
'We also assume, somewhat arbitrarily, that the legree of global and
local transaction interference in C-DBS is proportional to the leoree of
transaction interference in C-CRS as follows:
DII =(1K1"01
I.1 <3
07
where large X3 reflects either good D-DBS design anl/or applications
with strong locality. We feel that tne above assumption, whether right
or wrong, becomes irrelevant for applications featuring very low d!egrees
of transaction interference. We note here that most of present applica-
tions seem to be in that class.
Finally we assume that the cost of synchronizing global transactions in
D-OBS is proportional to the cost of synchronizing the transactions in
C-DSS, i.e., we assume that:
Cqcc = Ko *Ccc (I!)
Cgconf1 = .o*Cconf1
We will refer to a D-DBS which uses slow network (slow compared to
secondary memory channels) as a slow D-CPS for which I< >, 1. We will
refer to a D-CSS which uses fast network (comparable to seconrlary neory
channels ) as a fast (local) D-7DBS for which go l. Substitutin, (9),
(9), (10) an , (11) into (:) and (7) we aet
Cc =csys +Ccom +D*confl +Ccc (12)
Cd = 2*Ccsys + X*(<3*DI*R.**Ccon fl + l*Ccc) +
(l-X)*((l-K3)*DI*,)*Cconfl + Ko*Ccc + C,-ata) (13)
We would like to know when the cost of transaction execution is
larger in C-DeS compared to the cost of transaction execution in O-DBS.
Let
10
.... • . 1- 7 _ . i . , .. . e '
CC> CdjCA
Substituting from (12) and (13) into (14) we obtain
(I-K2)*Ccsys +Cco m + (1-K1 *K3*X)*DI*Cconfl + (1-KI*X)*Cc >
(1-X)*(2*Ko*(I-K 3)*DI*Ccon fI + Ko*Ccc + Cdata) (15)
If X.l, i.e. when almost all transaction in D-IBS are local
then (15) reduces to
(l-K2)*Ccsys + Ccom + (l-Kl*K3)*DI*Cconfl + (l-Kl)*Ccc > 0 (!q)
When almost all transactions are local (or equivalently when D-CDBS is
well designed) we introduce only negligible error by assuming that 1-3 =
1, i.e., assuming that the degree of interference in C-DBS is the same
as in C-DBS. Then (16) can be rewritten as
ccom > (K2-1)*Cc s ys + (Kl-1)*Csy n (17)
Let's assume that K1 = 1, i.e. C-OBS and D-09S execute under the same C
mechanism and the CC cost is the same in both. Tnen (17) reduces to
Ccom >/ (K-l)*Ccsys (18)
If we also assume that K2 = 1, i.e. the cost of transaction execution
without CC in O-DBS and C-DBS is the same then
11
"co >0c (19)'com
which is always true.
The above result says that in applications where (a) almost all
transactions are local, (b) the same CC mechanism is used in C-CBS and
D-DES, (c) local processors in D-DBS do not impose any processing cost
penalty compared to C-DBS processor, then the D-DBS regardless of the
speed of its network and regardless of the degree of interference will
result in lower transaction execution cost. This result which has been
derived from the analysis of our model is in a complete accord with our
intuition as it is to be expected if the model is realistic.
,'e come to the same conclusion when K2 = 1 and K1 < 1, (i.e. local
processors in D-CBS do not impose any transaction processing cost
penalty compareel to C-DES processor and D-D0S executes under lifferent
CC mechanism which has lower overhead cost).
Let's assune that V! = :-,' i.e. the local processors in C-2S
impose the same cost penalty on transactions and CC programs processing.
Then from (17) we get
Ccom > (K!-l)*(Csy + Ccsys) (2)
From (20) the only scenario on which we can make a general observa-
tion on is when '1 >> 1, i.e., when local orocessors in 7--AS i-nose
heavy processing cost penalty compared to C-CBS processor. Then from
(20) we obtain
12
Cco m >> (Csyn + Ccsys) (21)
The conclusion from (21) is that even if KI >> 1 C-OBS can still
result in lower transaction processing cost in applications where users
ship large amounts of data per transaction between remote terminals and
C-CBS processors, and terminal communication lines are slow and costly
and transactions are not computationally intensive.
we note here that our analysis could be made specific by substitut-
ing either measured or assumed values for the parameters in our formu-
las. We avoid loing so as no commercial D-DBS is operational today.
Therefore, we rather attempt to use a few reasonable assumptions so that
Ae can simplify our formulas and then make oeneral observations.
Let's assume that DI o0, i.e. very few transactions interfere
(that seems to be valid assumption for most applications). Also assume
that Xl = K2 = l,i.e. local processors in D-O8S Io not impose any cost
penalty on transactions and CC proarams processing compared to C-5BS
processor. Then from (15) we obtain
Ccom > (I-X)*(Cdata + Ccc*(O-l)) (221
For fast D-DBS Ko 1 and (22) reduces to
Ccom > (1-X)*Cdata (23)
For slow :--3S o >> 1 and (22) reduces to
Ccom > (!-X)*(Cdata + Ko*Ccc) (21)
13
The observations we offer on (23) and (?4) is that -when comparinq
transaction execution cost in C-08S and fast D-DBS then the CC mechanism
problem is not very important. However, it becomes very important when
comparing cost of transaction execution in C-DES and slow D-rBS.
4. C-CBS vs D-DBS: Performance Analysis.
The second issue we want to investiiate in this paper is the impact
of concurrency control on throuqhput of C-DRS and D-OBS based on slow
and fast networks. We are also interested in identifyinq applications
for which we can say that any D-DBS will likely outperform any C-BS and
vice versa.
The C-_BS throughput can be lerive& by considering the fact that
system transaction processing rate is 2ecreased by synchronization of
transactions. Thus we can express C-CBS throughput %c as follows:
Qc = ISC - Sccc + SCconfl)] 125)
where
SC is the basic transaction processingi rate of C-CBS which does
not have any concurrency control
Ccc is the fraction of basic transaction processing rate %
used for synchronization of transactions
Srconfl is the fraction of S. used for resolution of conflicts.
14
We model D-OSS as a set of n processors whose physical dependercy
due to an underlying network and whose logical and functional enendency
due to ilobal transactions are reflected only as a lecrease of the
throughput of every site in tD-DBS. Thus, the throughput of D-.rBS can be
,erived in terms of each node processing rate and its decrease ,ue to
the synchronization of local and global transactions as follows:
.d n*{SL-(SLcc + DIL*SLconfl) ] + (l-Y)*[CO*SL - (SOcc + DIG*Sconfl)]}
where
Y is the ratio of local transactions to all transactions
n is the number of sites or nodes in the D-t)BS
SL is the transaction processing rate of each of n nodes without
concurrency control
Skcc is the fraction of SL used for the synchronization of tran-
sactions local to one site
DIL is the -egree of local transaction interference, i.e., the
ratio of local interferring transactions to all local transactions
S -onf! is the fraction of SL used for the resolution of local
transaction conflicts
S#Cc is the fraction of SL used for the processing, of synchron-
ization messages of global transactions
3Gconfl is the fraction 3f SL used for the resolution :f 71obal
transaction conflicts
DIG is the degree of global transaction interference, i.e., the
ratio of interferring global transactions to all global transactions
I_ =
C. is the ratio of D-DBS average network Aelay to !'-DRS local
processor I/O time
We further assume that
SGcc = Col*SLcc
SGconfl = C02*SLconfl
where
C01 = C0*McC l 00 *Mcc
C0 2 C0 0 *Mconfl
where
is the ratio of average network delay to CC nessaoe set-up
time
Scc is CC no-conflict overhead, i.e., the average number of mes-
sages a iiven CC mechanism requires for synchronization of nonconflict-
inq transactions
'confl is CC conflict overhead, i.e., the averaoe number of mes-
sages a given CC mechanism requires for the resolution of transaction
conflicts. (An example of CC conflict and no-conflict overhead analysis
of several CC mechanisms can be found in (BAD ?1)).
We also assume that
SL = C1*SC
SL-C 1_ 2 cc
p1:-*S
5tconfl C2*SCconfl
DIL = C3*DI
DIG = (I-C 3 )*DI
We are interested when
Oc > (27)
Substituting into (27) from (25) and (26) and using the above
assumptions we obtain
Sc - (Scc + DI*Cconf ) > n*{Y*(C*S C - (C2*SCcc + C3*C2*DI*SCconfl)] +
(I-Y)*C0*C!*SC - (C01*C 2*SCcc + (I-C3)*Co2*C2*I*Scconfl)]} (28)
In order to simplify (28) we will assume that DI-O, C1 = C, and C
= 1. Thus we consicer applications which have very few conflictinr
transactions. We also assume that C-DBS and D-DS either use the same
CC -echanisms or if they are different then the .ecrease of local pro-
cessor througihout is the same as if they both use the same CC echah-
isms (C1 = C2 ). Finally we assume fast D-D8_S where D-LfBS local proces-
sor 1/O speed is the same as D-DES network speed (CO = 1). Sustitutina
these assumptions into (28) leads to
S(29)- *(l - n*C,) > 1 - n*r *(C01 + Y*(I - C1))3Ccc
We consider three cases 'when 1 > n*Cl, 1 = n*C1 and 1 < n*CI. When 1 >
n*CI, i.e., SC > n*SL then (29) can be rewritten is
17
SC 1 -n*C*(Y + C0 1 *(1 - Y)) (3")
SCcc 1 - n*C1
Powever, for any D-C8S
SC (31)> 1
and therefore
(1 - n*C1 *(Y + Cl*(I - Y)) (32.)
1 -n*C1
As we assumed 1 - n*C > 0 (32) reduces to
C01 . 1 (33)
.hus if SC > n*SL , then > 2, only if Cn, < 1.
'hen n*C1 = I then (29) corsiderina (31) reduces to
C'I >1 (34)
Thus if SC = n*SL then Z d > 9d only if Co1 > 1, else <
When 1 < n*C 1 , i.e., when S, < n*SL then (29) considerina (31) r-.ucesLI
to
.hus f S, < n*, en c > " Only if C1 > , se > "
From (33), (34) and (35) we can conclude that when comparing
18
J
throughput of C-DBS and 0-DBS based on fast network then the CC mechan-
ism (i.e. its CC overhead) and its efficient implementation (i.e. effi-
cient message processing) are quite important. This observation is not
entirely intuitive in the light of our assumption of very Few conflict-
ing transactions, i.e., DIDO. Of course if there are many conflictinq
transactions one would expect CC mechanism to be important for 7-DES
throughput.
An interesting observation can be made on (35). Even if D-DBS
transaction processing rate without CC mechanism is larger than transac-
tion processing rate without CC in C-D3BS, a D-DBS with CC mechanism can
perform worse than C-DRS with CC mechanism if Col < 1, i.e., if either
CC mechanism has high no-conflict overhead or it has slow CC message
processing.
Let's consider applications where YWl, i.e., there are very few
global transactions or equivalently almost all transactions are local.
In such case we can also assze that C, = 1. Substituting these assumtp-
tions into (27) we obtain
SC*(l - n*CI) > (Scc + DI*Ccconfl)*(l - n*C 2 ) (39
ie consider (36) when 1 > n*Cl, 1 = n*C1 and. 1 < n*C1
',hen 1 > n*C1 then from !39) and (31) we obtain:
C2 < Cl (37)
Thus when SC > n*SL then 2c > Td only when C2 < CI, i.e., when
19
O8S and C-OBS use different CC mechanism or they use the same one but
the D-DBS local processor throughput decrease due to synchronization of
nonconflicting transactions is smaller than the lecrease in local pro-
cessor transaction processing rate compared to C-DBS transaction oro-
cessinq rate.
'.Then 1 = n*C i , i.e., when SC = n*SL then (36) always holds and thus
> -
" hen 1 < n*C 1 then from (36) and (31) we get
C2 > C1 (30)
Thus when SC < n*SL then q. > Q. only if C2 > C1 . If C2 < C, then
> c. The implication here is that d >2. if either D-DBS uses more
efficient CC mechanism than C-CBS or D-BS uses the same CC mechanism
but its implementation is more efficient compared to C-OBS transaction
processing, i.e., if
SLcc SL
-CC(
As can be seen from (36), (37) and (38) CC nechanism is a signiFi-
cant issue when comraring performance of C-DBS and D-CBS systems. More-
over it is important even for the performance of D-DRCS based on Last
network and for applications where either there are few inter~errin'
transactions or where there are few ilobal transactions.
5. Conclusions.
20
In this paper we have investigated transaction execution cost and
system throughput in C-DBS compared to D-BS systems based on slow and
fast networks. The conclusions reached in this paper indicate that the
efficiency of CC mechanism is of great importance when comparing C-OS
and fast or slow D-CBS throughput. The same observation acplies when
comparing transaction execution cost in C-CBS and slow 0-DS. It seems
that CC mechanism is not important when comparing the cost of transac-
tion execution in C-DBS and fast D-DBS.
In this paper we have also indicated for which applications any D-
DBS is likely to be a better solution than any C-DBS and vice versa.
5. Acknowledgement.
The author would like to acknowledge support of the VPF Foundation
Research Programs for this work.
References
rBAD8O1
Badal, D. Z., "The Analysis of the !ffects if (.oncurrenc- Control
on Distributed 1atabse System Performance", Proceedings of the '-th
International Conference on Very Large 71ata lases, Montreal, October
1920, 37 -3- 2.
(BADK]
Badal, D. Z., "Concurrency Control Overhead or Closer Look at
21
Blocking vs. Yonblocking Concurrency Control Mechanisms", Proceedings of
the 5th Berkeley Conference on Distributed Data vanaoement and Com-uter
'etworks, San Francisco, February 19R1.
rBUC79]
Bucci , G. and Streeter, D. N., ' Methodoloqy for the 'esiqn of
jDistributed information Systems", CACm 22, 4 (April 19 9), 233-?5.
[MOL791Garcia-4olina, H., "Performance of U~ate Alcorithms for Repi-
cated rata in a Distributed Database," Ph.D. dissertation, Devnartnent of
Computer Science, Stanford University, June 1979.
rRIE79JJRies, D. R., 'The Effects of Concurrency Control on Tatahase
manaaement System Performance," Ph.D. dissertation, Computer Science
! epart. ent, University of California, :-erkeley, April l1'9.ESTE79 ]
Stewart, H. M. "Performance analysis of Complex Communications
Systems", IBM System Journal 12, 3 (1970), 35<-73.
2L
120
INITIAL DISTRIBUTION LIST
Defense Technical Information Center 2Cameron StationAlexandria, VA 22314
SDudley Knox Library2
Code 0142Naval Postgraduate SchoolMonterey, CA 93940
Office of Research AdministrationCode 012ANaval Postgraduate SchoolMonterey, CA 93940
Chairman, Code 52Hq 2Department of Computer ScienceNaval Postgraduate SchoolMonterey, CA 93940
Dushan Badal, Code 52Zd 5Department of Computer ScienceNaval Postgraduate SchoolMonterey, CA 93940
Robert B. GraftonOffice of Naval ResearchCode 437800 N. Quincy StreetArlington, VA 22217
David W. MizellOffice of Naval Research1030 East Green StreetPasadena, CA 91106
Vinton G. CerfDARPA/IPTO1400 Wilson BlvdArlington, VA 22209
CDR R. OhlanderDARPA1400 Wilson BlvdArlington, VA 22209
ral. D. AdamsDARPA1400 Wilson BlvdArlington, VA 22209
23
______