swift presentation
TRANSCRIPT
-
7/29/2019 SWIFT Presentation
1/76
Dept. of CSE, MMMEC Gorakhpur1
Dr. Udai ShankerDepartment of Computer Science & Engineering
M. M. M. Engineering College, Gorakhpur-273 010
India
SWIFT
A Real Time Commit Protocol
-
7/29/2019 SWIFT Presentation
2/76
Dept. of CSE, MMMEC Gorakhpur2
Outline of Presentation Distributed Real Time Database Systems (DRTDBS) - A Glance
Performance Issues
DRTDBS Model
Commit Protocol - SWIFT
Conclusions
Scope for Future Research Questions & Answers
-
7/29/2019 SWIFT Presentation
3/76
Dept. of CSE, MMMEC Gorakhpur3
Real Time Sys tem
Results be produced within a specified deadline period.
Correctness depends on
Logical results of computation, and also
Time at which results will be produced
Distr ibu ted Database Sys tem A Collection of Data Items Distributed over Distant Locations
DRTDBS
A Join of Real Time Systems & Distributed Database Systems
Time Constrained Distributed Database Systems
DRTDBSA Glance
-
7/29/2019 SWIFT Presentation
4/76
Dept. of CSE, MMMEC Gorakhpur4
Real Time Systems vs . DRTDBS
Real Time Systems Task Centric
Deadlines attached to tasks.
Distributed Real Time Databases Data Centric
Data has temporal validity, i.e., deadline attached to transactions.
Transactions must be executed by deadlines to keep the data valid, in
addition to produce results in a timely manner.
DRTDBSA Glance contd
-
7/29/2019 SWIFT Presentation
5/76
Dept. of CSE, MMMEC Gorakhpur5
Satellite
Imagery
News
Services
Need Summary
Report by 4 PM
Troop
Positions
Network
World Wide
Real-Time,DBs
Archival
DBs
The Problem
Scenario
Multimedia
DB
DRTDBSA Glance contd
-
7/29/2019 SWIFT Presentation
6/76
Dept. of CSE, MMMEC Gorakhpur6
Transactions
Perform task of setting the value of a real world object.
Are invoked to access databasesby all applications.
Must be scheduled to complete within their time constraint.
Satisfy database constraints.
DRTDBSA Glance contd
Notion of Transact ion
Partially ordered set of database operations A complete and consistent computation (i.e., they are designed to terminate
correctly, leaving the database in a consistent state)
Have dynamic runtime behavior (dependent on the state of the database,i.e., data values)
Data is a resource (transaction can be blocked in accessing data objects)
A transaction is said to commit if all changes can be successfully made to
the database and to abortif all changes cannot be successfully made to the
database.
-
7/29/2019 SWIFT Presentation
7/76Dept. of CSE, MMMEC Gorakhpur7
d t
v(t)
v0
Hard deadlineHard Real Time Transactions
If deadline missed, catastrophic consequence, either heavy economic or human
safety critical applications
life or environment threatening emergency situations
DRTDBSA Glance contd
-
7/29/2019 SWIFT Presentation
8/76Dept. of CSE, MMMEC Gorakhpur8
d t
v(t)
v0
firm deadlineFirm Real TimeTransactions
If deadline missed
Completing the transaction may generate harmful effects onthe system.
It can be, however late result is worthless. Exp-
Transactions attempting to recognize a moving object.
Arbitrage trading.
DRTDBSA Glance contd
-
7/29/2019 SWIFT Presentation
9/76Dept. of CSE, MMMEC Gorakhpur9d2 t
v(t)
v0
d1
Soft deadline
Soft Real Time Transactions
If deadline missed
Some value even after expiry of its deadlines
Value diminishes with time
DRTDBSA Glance contd
-
7/29/2019 SWIFT Presentation
10/76Dept. of CSE, MMMEC Gorakhpur10
Types of Transaction(Operat ions) Write-only Transact ions (Sensor Updates): Obtain state of the
environment and write into the database
Store sensor data in database (e.g., temperature)
Monitoring of environment
Ensure absolute temporal consistency
Update Transact ions (Ap pl icat ion Updates)
Derive new data and store in database
Based on sensor and other derived data
Read-only Transact ions
Read data, compute, and report (or send to actuators)
DRTDBSA Glance contd
-
7/29/2019 SWIFT Presentation
11/76Dept. of CSE, MMMEC Gorakhpur11
Perfo rmance Issues
Transaction Scheduling
Conflict Resolution
Execute Execute Conflict - Concurrency Control
Execute Commit Conflict - Commit Procedure
Deadlocks
Priority Assignment Policy
Data Invariance
Data Access Mechanism
Static Two Phase Locking
Dynamic Two Phase Locking
Performance Issues
-
7/29/2019 SWIFT Presentation
12/76Dept. of CSE, MMMEC Gorakhpur12
I/O & Disk Scheduling
Buffer Management
Communication Delays between Different Sites
Site Failures
Checkpointing and Logging for the Fault Tolerance & Failure Recovery
Predictability & Consistency
Security
CPU Scheduling
Performance Issues contd
-
7/29/2019 SWIFT Presentation
13/76Dept. of CSE, MMMEC Gorakhpur13
Network
ManagerSite 2 Site 3
Site N
Transaction
Manager
Transaction
GeneratorSink
C.C.Manager
Abort
Terminate
MemoryComputation
Commit
Database
Operation
Priority
Assignment wait Queue
ready queue
Blocked
Site 1
DRTDBS Model
-
7/29/2019 SWIFT Presentation
14/76Dept. of CSE, MMMEC Gorakhpur14
Model Assumpt ions
Firm Real Time Transactions
Processing of a transaction requires use ofCPU and data items
located at local site or remote site.
Arrivals of transactions at a site are independent of the arrivals at
other sites and use Poisson distribution.
Each cohort makes read and update accesses.
Each transactionpre-declares its read-set (set of data items that
the transaction will only read) and update-set (set of data items
that the transaction will update).
Static two phase lockingwith higher priorityscheme is used forlocking the data items.
DRTDBS Model contd
-
7/29/2019 SWIFT Presentation
15/76Dept. of CSE, MMMEC Gorakhpur15
A lending transaction cannot lend the same data item in
read/update mode to more than one cohort.
Cohort already in the dependency set of another cohort cannotpermit another incoming cohort to read or update.
Database is in main memoryor in diskat all sites.
Communication delay is considered either0ms or100ms.
In case of disk resident database, buffer space is sufficiently largeto allow the retention of data updates until commit time.
Cohorts are executed inparallelway.
Operations performed by one cohort are independentof the results of the
operations performed at the other sites.
DRTDBS Model contd
-
7/29/2019 SWIFT Presentation
16/76Dept. of CSE, MMMEC Gorakhpur16
Symbols Meaningmi No. of Cohorts of ith Transaction
No. of Operations in Local Cohort of Ti
No. of Operations in jth Cohort of Ti
Tlock Time Required to Lock/Unlock a Data Item
Tprocess Time to Process a Data Item (Assuming read operation takes
same amount of time as write operation.)
Ncomm No. of Messages Exchanged Between Coordinator and a Cohort
Tcom Communication Delay, i.e., Constant Time Estimated for aMessage Going from One Site to Another
Noper_local Number of Local Operations
Noper_remote Max. No. of Remote Operations Taken Over by All Cohorts
localiN
jiN
DRTDBS Model contd
-
7/29/2019 SWIFT Presentation
17/76
Dept. of CSE, MMMEC Gorakhpur17
Proposed Method fo r Compu tat ion o f Deadl ine
Deadlines-
Expected Execution Times
Deadline (Di) of Transaction (Ti):
Di=Ai+ SF *Ri
Ai - Arrival Time of Transaction (Ti) at A SiteSF - Slack Factor
Ri- Minimum Transaction Response Time Given as
Ri=Rp+Rc
Rp- Time for Execution Phase
Rc- Time for Commitment Phase
DRTDBS Model contd
-
7/29/2019 SWIFT Presentation
18/76
Dept. of CSE, MMMEC Gorakhpur18
Global Transactions
Local Transaction s
gproceslockl TTTp sin2
c comm comR = N T
p local p jp l i l i comR =max{T N ,max(T N )+2T }
imj 1
localj
p lock process oper_localR = (2 T + T ) N
cR = 0
DRTDBS Model contd
-
7/29/2019 SWIFT Presentation
19/76
Dept. of CSE, MMMEC Gorakhpur19
Simu lat ion Detai ls
Event driven simulator was written in C language.
Each result was calculated as an average of 10 independent runs. In each
run, 20000 transactions were initiated.
Primary Performance Criteria
Proportion of Missed Deadlines (Miss Percentage, MP)
Miss Percentage Values
Normal Load- 0 to 20% Heavy Load- 20% to 100%
100gsinprocesforsystemthetosubmittednstransactioof.no
abortednstransactioofnumber=MP
DRTDBS Model contd
-
7/29/2019 SWIFT Presentation
20/76
Dept. of CSE, MMMEC Gorakhpur20
Two Simu lat ion Models
Main Memory as well as Disk Resident Distributed Real Time DatabaseSystem
Structure of simulation model and method for computation of deadlines
of global & local transactions are same as described previously
Each transactionis associated with
Health factor (HF) = TimeLeft/ MinTime
Where,
TimeLeft - Time left until Transactions Deadline
MinTime - Minimum Time required for Commit Processing
MinHF
1. Threshold that allows data held by committing transaction to be
accessed
2. Taken as 1.2 (Value of MinHF used in PROMPT)
DRTDBS Model contd
-
7/29/2019 SWIFT Presentation
21/76
Dept. of CSE, MMMEC Gorakhpur21
Simu lat ion Parameters
Parameters Meaning Settings
Nsite Number of Site 4
AR Arri val Rate 2-20 Transactions/Second
Tcom
Communication Delay 100 ms or 0 ms (Constant)
SF Slack F actor 1-4 (Unif orm Distribution)
Noper
No. of Operations in a
Transaction
3-20 (Unif orm Distribution)
PageCPU CPU Page Processing Time 5 ms
PageDisk Disk Page Processing Time 20 ms
DBsize Database Size 200 Data I tems/Site
Pwrite
Write Operation Probabil ity 0.60
DRTDBS Model contd
-
7/29/2019 SWIFT Presentation
22/76
Dept. of CSE, MMMEC Gorakhpur22
Commit Protocol
SWIFTStatic Two Phase Lock ing w ith Higher Pr ior i ty Based, Write-
Update Type, Ideal for Fast and Timel iness Commit Protocol
Introduct ion
Several factors contribute to difficulty in meeting real time constraint.
Data Conflicts Among Transactions
Prime Factor for System Performance Degradation
Data Conflicts Between Two Transactions
Execute-Execute Conflicts-Already Addressed
Execute-Commit Conflicts-Very Little Work
Designing of a Good Commit Protoco l - Imp or tant for DRTDBS
-
7/29/2019 SWIFT Presentation
23/76
Dept. of CSE, MMMEC Gorakhpur23
Related Work
Two Phase Commit (2PC) is still one of the most commonly used protocols in
the study of DRTDBS
N. Soparkar, E. Levy, H.F.Korth and A. Silberschatz. Adaptive Commitment
for Real-time Distributed Transaction. Technical Report TR-92-15,
Department of Computer Science, University of Texax, Austin, 1992.
K.Y. Lam, C-L. Pang, S.H. Son and J. Cao. Resolving executing-committing
conflicts in distributed real-time database systems. The computer Journal, 42(8), 1999, 674-692.
J. Haritsa, K. Ramamritham and R. Gupta. The PROMPT real time commit
protocol. IEEE Transaction on parallel and distributed systems, 11(2), 2000,
160-181.
Biao Qin and Yunsheng Liu. High performance distributed real time commitprotocol, Journal of Systems and Software, Elsevier Science Inc, 2003, 1-8.
Commit Protocol contd
http://d/2SC.ppthttp://d/2SC.ppthttp://d/2SC.ppthttp://d/2SC.ppthttp://d/2SC.ppthttp://d/2SC.ppthttp://d/2SC.ppthttp://d/2SC.ppthttp://d/2SC.ppthttp://d/2SC.ppthttp://d/2SC.ppthttp://d/2SC.ppthttp://d/2SC.ppthttp://d/2SC.ppthttp://d/2SC.ppthttp://d/2SC.ppt -
7/29/2019 SWIFT Presentation
24/76
Dept. of CSE, MMMEC Gorakhpur24
User
Coordinator
Site
Cohort
Site i
Cohort
site j
Cohort
Site n
Commit Protocol
Cohort
site k
Transaction Arrival
All Cohorts Processed
Transaction Commit
Execu
tionPhase
CommitPhase
Commit Protocol contd
-
7/29/2019 SWIFT Presentation
25/76
Dept. of CSE, MMMEC Gorakhpur
Two -Phase Comm it (2PC):Presumed Nothing 2PC protocol (PrN)
Assuming no failures, it works as follows:
Site at which transaction originates is coordinator;
Other sites as well as coordinator site, at which it executes, creates cohorts.
When an transaction wants to commit:
Coordinatorsends preparemsg to each cohort.
Cohortforce-writes an abor torpreparelog record and then sends a nooryesmsg to coordinator.
If coordinator gets unanimous yes votes, force-writes a commi t log
record and sends commi tmsg to allcohorts. Else, force-writes abor tlog
rec, and sends abor tmsg.
Cohorts force-write abor t /commit ted logrec based on msg they get,
then sendackmsg to coordinator.
Coordinatorwrites endlog rec after getting allacks.
Commit Protocol contd
-
7/29/2019 SWIFT Presentation
26/76
Dept. of CSE, MMMEC Gorakhpur26
Coordinator Cohort
Prepare
Vote YES
Commi t
Ack
Log Prepared
Log Comm it ted
Log Commi t
Log record is forced wr i t ten
VotingPhase
DecisionPhase
Two Phase Comm it (2PC) Protoco l
Log END
Commit Protocol contd
-
7/29/2019 SWIFT Presentation
27/76
Dept. of CSE, MMMEC Gorakhpur27
Propo sed Real Time Comm it Protoc ol - SWIFT
Data Ac cess Confl icts Reso lving Strategies
Types of Dependencies (Update Model & Read only)
Comm it Dependency (CD)
If a transaction T2 updates a data item read by another transaction T1, a
commit dependency is created from T2 to T1. Here, T2 is not allowed to
commit until T1 commits.
Abo r t Dependency (AD)
If transaction T2 reads or updates an uncommitted data item written by
transaction T1, an abort dependency is created from T2 to T1. T2 aborts, if
T1 aborts and T2 is not allowed to commit before T1.
Each transaction Ti, that lends its data while in PREPARED state to an
executing transaction, maintains two sets
CDS (Ti):Set of Transactions (Tj) commit dependent on transaction
(Ti)
ADS (Ti):Set of transactions (Tj) abort dependent on transaction (Ti)
Commit Protocol-SWIFT
-
7/29/2019 SWIFT Presentation
28/76
Dept. of CSE, MMMEC Gorakhpur28
Type of Dependenc ies in Different Cases o f Data
Conf l ic tThree Possible Cases of Conflicts
Case 1: Read-Update Con fl ict .
If transaction T2 requests an update-lock while transaction T1 is holding a
read-lock, a commit dependency is defined from T2 to T1. First, the
transaction identity (id) of T2 is added to the CDS (T1). Then T2 acquires
the update-lock.
Case 2: Update-Update Con fl ic t.
If both locks are update-locks and HF(T1) MinHF, an abort dependency
is defined from transaction T2 to transaction T1. The transaction identity (id)
of T2 is added to ADS (T1), and T2 acquires the update-lock; otherwise, T2
is blocked.
Case 3: Update-Read Confl ic tIf transaction T2 requests a read-lock while transaction T1 is holding an
update-lock and HF(T1) MinHF, an abort dependency is defined from T2
to T1. The transaction identity (id) of T2 is added to ADS (T1), and T2
acquires the read-lock; otherwise, T2 is blocked.
Commit Protocol-SWIFT contd
-
7/29/2019 SWIFT Presentation
29/76
Dept. of CSE, MMMEC Gorakhpur29
If (T2 CD T1){
CDS (T1) = CDS (T1) {T2};T2 is granted update lock;
}
else{if ((T2 AD T1) AND (HF(T1) MinHF)){
ADS (T1) = ADS (T1) {T2};T2 is granted the requested lock (read or update);
}elseT2 will be blocked;
}
Processing of A ccess of Data Items in Conf l ic t ing
Mode by Lock Manager
Commit Protocol-SWIFT contd
-
7/29/2019 SWIFT Presentation
30/76
Dept. of CSE, MMMEC Gorakhpur30
Mechanics of Interact ion between Lender and
Borrow er Cohor ts
If transaction T2 has borrowed the data item locked by transaction T1,
following three scenarios are possible:
Scenar io 1:T1 receives decision before T2 is going to start processing
phase after getting all its locks.
If the global decision is to commit, T1 commits.
All cohorts in ADS (T1) and CDS (T1) will execute as usual and the sets
ADS (T1) and CDS (T1) are deleted.
If the global decision is to abort, T1 aborts. The cohorts in the
dependency sets of T1 will execute as follows:
All cohorts in ADS (T1) will be aborted;
All cohorts in CDS (T1) will execute as usual;
Sets ADS (T1) and CDS (T1) are deleted.
Commit Protocol-SWIFT contd
-
7/29/2019 SWIFT Presentation
31/76
Dept. of CSE, MMMEC Gorakhpur31
Scenario 2:T2 is going to start processing phase after getting all locks
before T1 receives global decision.
T2 is allowed to send a WORKSTARTED (discussed later) message to itscoordinator, if it is commit dependent only; otherwise it is blocked from
sending the WORKSTARTED message (So, the coordinator cannot initiate
the commit processing operation) and has to wait, until
Alternative 1: either T1 receives its global decisions, or
Alternative 2: its own deadline expires,
whichever occurs earlier.
In case of alternative 1, the system will execute as in scenario 1, whereas in
the case of alternative 2, T2 will be killed and will be removed from the
dependency set of T1.
Scenario 3:T2 aborts before T1 receives decision
In this situation, T2s updates are undone and T2 will be removed from the
dependency set of T1.
Commit Protocol-SWIFT contd
-
7/29/2019 SWIFT Presentation
32/76
Dept. of CSE, MMMEC Gorakhpur32
Basic Idea of Protocol
Sending of WORKSTARTED Message Just Before Start of Processing
Phase
Allowing Commit Dependent Only Borrower to Send Its WORKSTARTED
Message Instead of being Blocked
Checking of Completion of Processing & Removal of Dependency Before
Sending YES VOTE Message
(A )Execution Phase is divided in
Locking Phase
Processing Phase
Commit Protocol-SWIFT contd
-
7/29/2019 SWIFT Presentation
33/76
Dept. of CSE, MMMEC Gorakhpur33
Cohort Execut ion
During the locking phase, the transaction locks the data items.
Just before the start of processing phase, the cohort sends a
WORKSTARTED message to its coordinator.
After the receipt of WORKSTARTED messages from all its cohorts, the
coordinator sends VOTE REQ message to all its cohorts at time t
calculated as follows:
t = Max {ti+ Processing_t imei} - Tcom
where,
ti= Arrival time of WORKSTARTED message from cohorti
Processing_t imei= Processing time needed by the cohort i
Tcom= Communication Delay from one site to another
Commit Protocol-SWIFT contd
-
7/29/2019 SWIFT Presentation
34/76
Dept. of CSE, MMMEC Gorakhpur34
(B )
One of the following two decisions is taken based on the types of
dependency
T2 sends WORKSTARTED message to its coordinator if it is only
commit dependent on other cohorts.
Free from Cascaded Aborts (Abort of T1 (lender) does not
cause T2 (borrower) to abort)
T2 is not allowed to send WORKSTARTED message to its coordinator
if it is abort dependent on other cohorts.
Coordinator cannot initiate commit processing.
It has to wait until either T1 receives its global decisions or its
own deadline expires, whichever occurs earlier.
Commit Protocol-SWIFT contd
-
7/29/2019 SWIFT Presentation
35/76
Dept. of CSE, MMMEC Gorakhpur35
(C)
On receipt of Vote Req. message, one of the following decisions is taken
Ifcoordinators VOTE REQ message
Cohort sends a YES VOTE message, only if
No Dependencies
It has finished its processing
If it is still dependent on any cohort or has not finished its processing
YES VOTE message is deferred.
Borrower sends deferred YES VOTE message, after
Completion of Processing, and
Removal of Dependencies
This may be either due to abort or commit of the lender.
Commit Protocol-SWIFT contd
-
7/29/2019 SWIFT Presentation
36/76
Dept. of CSE, MMMEC Gorakhpur36
Algor i thmif (T1 receives global decision before, T2 is going to start processing phase after
getting all locks)
{ONE: if (T1s global decision is to commit)
{ T1 enters in the decision phase;
All cohorts in ADS (T1) and CDS (T1) will execute as usual;
Delete set of ADS (T1) and CDS (T1);
}
else //T1s global decision is to abort
{ T1 aborts;
The cohorts in CDS (T1) will execute as usual;
Transaction in ADS (T1) will be aborted;
Delete sets of ADS (T1) and CDS (T1);
}}
else if (T2 is going to start processing phase after getting all locks before, T1
receives global decision)
{ Check type of dependencies;
Commit Protocol-SWIFT contd
-
7/29/2019 SWIFT Presentation
37/76
Dept. of CSE, MMMEC Gorakhpur37
if (T2s dependency is commit only)
T2 sends WORKSTARTED message;
else
{ T2 is blocked for sending WORKSTARTED message;while (! (T1 receive global decision OR T2 misses deadline))
{
if (T2 misses deadline)
{ Undo computation of T2;
Abort T2;
Delete T2 from CDS (T1) & ADS (T1);
}
else GoTo ONE;
}
}
}
else //T2 is aborted by higher transaction before, T1 receives decision
{ Undo computation of T2;
Abort T2;
Delete T2 from CDS (T1) & ADS (T1);
}
Commit Protocol-SWIFT contd
-
7/29/2019 SWIFT Presentation
38/76
Dept. of CSE, MMMEC Gorakhpur38
SWIFT is compared w ith proto cols-
PROMPT
2SC
SWIFT- Preliminary Version - One (SWIFT-PV-1)
SWIFT- Preliminary Version - Two (SWIFT-PV-2)
SWIFT-PV-1 Basic concept of sending the WORKDONE message only, ifcohort is commit dependent on other cohorts.
SWIFT-PV-2 Sending of WORKSTARTED message is considered before
the start of processing phase.
SWIFT Combination of concepts of SWIFT-PV-1 and SWIFT-PV-2.
Commit Protocol-SWIFT contd
-
7/29/2019 SWIFT Presentation
39/76
Dept. of CSE, MMMEC Gorakhpur39
Fig. 4.1: Miss % with (RC+DC) at Communication Delay=100msNormal & Heavy Load
Transaction Arrival Rate (no. per second)
2 4 6 8 10 12 14
Miss
%
0
20
40
60
80
100
PROMPT
2SC
SWIFT-PV-1
Simulat ion Resu l tsMain Memory Resident Database
Commit Protocol-SWIFT contd
-
7/29/2019 SWIFT Presentation
40/76
Dept. of CSE, MMMEC Gorakhpur40
Fig. 4.2: Miss % with (RC+DC) at Communication Delay=100msNormal & Heavy Load
Transaction Arrival Rate (no. per second)
2 4 6 8 10 12 14
Miss%
0
20
40
60
80
100
PROMPT2SC
SWIFT-PV-2
Main Memory Resident Database
Commit Protocol-SWIFT contd
-
7/29/2019 SWIFT Presentation
41/76
Dept. of CSE, MMMEC Gorakhpur41
Fig. 4.3: Miss % with (RC+DC) at Communication Delay=100msNormal & heavy Load
Transaction Arrival Rate (no. per second)
2 4 6 8 10 12 14
Miss%
0
20
40
60
80
100
PROMPT
2SC
SWIFT-PV-1
SWIFT-PV-2
SWIFT
Main Memory Resident Database
Commit Protocol-SWIFT contd
-
7/29/2019 SWIFT Presentation
42/76
Dept. of CSE, MMMEC Gorakhpur42
Fig. 4.4: Miss % with (RC+DC) at Communication Delay=0msNormal & Heavy Load
Transaction Arrival Rate (no. per second)
10 15 20 25 30 35 40 45 50
Miss%
0
20
40
60
80
PROMPT
2SC
SWIFT-PV-1
Main Memory Resident Database
Commit Protocol-SWIFT contd
-
7/29/2019 SWIFT Presentation
43/76
Dept. of CSE, MMMEC Gorakhpur43
Fig. 4.5: Miss % with (RC+DC) at Communication Delay=0msNormal & Heavy Load
Transaction Arrival Rate (no. per second)
10 15 20 25 30 35 40 45 50
Miss%
0
20
40
60
80
PROMPT
2SC
SWIFT-PV-2
Main Memory Resident Database
Commit Protocol-SWIFT contd
-
7/29/2019 SWIFT Presentation
44/76
Dept. of CSE, MMMEC Gorakhpur44
Fig. 4.6: Miss % with (RC+DC) at Communication Delay=0msNormal & Heavy Load
Transaction Arrival Rate (no. per second)
10 15 20 25 30 35 40 45 50
Miss%
0
20
40
60
80
PROMPT
2SC
SWIFT-PV-1
SWIFT-PV-2
SWIFT
Main Memory Resident Database
Commit Protocol-SWIFT contd
-
7/29/2019 SWIFT Presentation
45/76
Dept. of CSE, MMMEC Gorakhpur45
Fig. 4.7: Miss % with (RC+DC) at Communication Delay=0msNormal Load
Transaction Arrival Rate (no. per second)
3 4 5 6
Miss
%
0
5
10
15
20
25
PROMPT
2SC
SWIFT-PV-1
Disk Resident Database
Commit Protocol-SWIFT contd
-
7/29/2019 SWIFT Presentation
46/76
Dept. of CSE, MMMEC Gorakhpur46
Fig. 4.8: Miss % with (RC+DC) at Communication Delay=0msHeavy Load
Transactional Arrival rate (no. per second)
6 9 12 15 18
Miss%
0
20
40
60
80
100
PROMPT
2SC
SWIFT-PV-1
Disk Resident Database
Commit Protocol-SWIFT contd
-
7/29/2019 SWIFT Presentation
47/76
Dept. of CSE, MMMEC Gorakhpur47
Fig. 4.9: Miss % with (RC+DC) at Communication Delay=0msNormal Load
Transaction Arrival Rate (no. per second)
3 4 5 6
Miss%
0
5
10
15
20
25
PROMPT
2SC
SWIFT-PV-2
Disk Resident Database
Commit Protocol-SWIFT contd
-
7/29/2019 SWIFT Presentation
48/76
Dept. of CSE, MMMEC Gorakhpur48
Fig. 4.10: Miss % with (RC+DC) at Communication Delay=0ms
Heavy Load
Transactional Arrival rate (no. per second)
6 9 12 15 18
Miss
%
0
20
40
60
80
100
PROMPT
2SC
SWIFT-PV-2
Disk Resident Database
Commit Protocol-SWIFT contd
C i l
-
7/29/2019 SWIFT Presentation
49/76
Dept. of CSE, MMMEC Gorakhpur49
Fig. 4.11: Miss % with (RC+DC) at Communication Delay=0ms
Normal Load
Transaction Arrival Rate (no. per second)
3 4 5 6
Miss
%
0
5
10
15
20
25
PROMPT
2SC
SWIFT-PV-1
SWIFT-PV-2
SWIFT
Disk Resident Database
Commit Protocol-SWIFT contd
C i P l
-
7/29/2019 SWIFT Presentation
50/76
Dept. of CSE, MMMEC Gorakhpur50
Fig. 4.12: Miss % with (RC+DC) at Communication Delay=0msHeavy Load
Transactional Arrival rate (no. per second)
6 9 12 15 18
Miss%
0
20
40
60
80
100
PROMPT
2SC
SWIFT-PV-1
SWIFT-PV-2
SWIFT
Disk Resident Database
Commit Protocol-SWIFT contd
C i P l
-
7/29/2019 SWIFT Presentation
51/76
Dept. of CSE, MMMEC Gorakhpur51
Fig. 4.13: Miss % with (RC+DC) at Communiction Delay=100msNormal & Heavy Load
Transaction Arrival Rate (no. per second)
2 4 6 8 10 12 14
Miss%
0
20
40
60
80
100
PROMPT
2SC
SWIFT-PV-1
Disk Resident Database
Commit Protocol-SWIFT contd
C i P l
-
7/29/2019 SWIFT Presentation
52/76
Dept. of CSE, MMMEC Gorakhpur52
Fig. 4.14: Miss % with (RC+DC) at Communiction Delay=100msNormal & Heavy Load
Transaction Arrival Rate (no. per second)
2 4 6 8 10 12 14
Miss%
0
20
40
60
80
100
PROMPT
2SC
SWIFT-PV-2
Disk Resident Database
Commit Protocol-SWIFT contd
C i P l
-
7/29/2019 SWIFT Presentation
53/76
Dept. of CSE, MMMEC Gorakhpur53
Fig. 4.15: Miss % with (RC+DC) at Communication Delay=100msNormal & Heavy load
Transaction Arrival Rate (no. per second)
2 4 6 8 10 12 14
Miss
%
0
20
40
60
80
100
PROMPT
2SC
SWIFT-PV-1
SWIFT-PV-2
SWIFT
Disk Resident Database
Commit Protocol-SWIFT contd
C i P l
-
7/29/2019 SWIFT Presentation
54/76
Dept. of CSE, MMMEC Gorakhpur54
Part ial Read On ly Optim izat ion
If some of its cohorts have only read operations, no need to be involved in
the second phase of protocol because it does not matter whether thetransaction is finally committed or aborted to ensure its atomicity at that
cohort site.
Cohort may send a read-only WORKSTARTED message to its coordinator
indicating that it is no longer needed by the cohort to participatefurther.
Minimizes intersite message traffic, execute-commit conflicts and log writesconsequently resulting in a better response time.
1% to 5% Improvement in Transaction Miss Percentage
Poss ible Cases o f Data Conf l icts
Update-Update & Update-Read are only possible conflicts with arriving cohorts
Commit Protocol-SWIFT contd
C i P l
-
7/29/2019 SWIFT Presentation
55/76
Dept. of CSE, MMMEC Gorakhpur55
Dependency Requ irement
Abort Dependenc y (ADS)
If transaction T2 reads or updates an uncommitted data item updated bytransaction T1, an abort dependency is created from T2 to T1. T2 aborts, if T1
aborts and T2 is not allowed to commit before T1.
Type of Dependencies in Cases o f Data Confl icts
Case 1: Update-Update Confl ic t
If both locks are update-locks and HF(T1) MinHF, an abort dependency is
defined from T2 to T1. Transaction identity (id) of T2 is added to ADS (T1), and
T2 acquires the update-lock; otherwise, T2 is blocked.
Case 2: Update-Read Confl i ct
If T2 requests a read-lock while T1 is holding an update-lock and HF(T1)
MinHF, an abort dependency is defined from T2 to T1. Transaction identity (id)
of T2 is added to ADS (T1), and T2 acquires the read-lock; otherwise, T2 is
blocked.
Commit Protocol-SWIFT contd
C i P l
-
7/29/2019 SWIFT Presentation
56/76
Dept. of CSE, MMMEC Gorakhpur56
Fig. 4.16: Miss % with (RC+DC) at Communication Delay=0msNormal & Heavy Load
Transaction Arrival Rate (no. per second)
10 15 20 25 30 35 40 45 50
Miss
%
0
20
40
60
80
SWIFT
SWIFT with Partial Read Optimization
Simulat ion Resu l tsMain Memory Resident Database
Commit Protocol-SWIFT contd
C it P t l
-
7/29/2019 SWIFT Presentation
57/76
Dept. of CSE, MMMEC Gorakhpur57
Fig. 4.17: Miss % with (RC+DC) at Communication Delay=100msNormal & heavy Load
Transaction Arrival Rate (no. per second)
2 4 6 8 10 12 14
M
iss%
0
20
40
60
80
SWIFT
SWIFT with Partial Read Optimization
Main Memory Resident Database
Commit Protocol-SWIFT contd
C it P t l
-
7/29/2019 SWIFT Presentation
58/76
Dept. of CSE, MMMEC Gorakhpur58
Fig. 4.18: Miss % with (RC+DC) at Communication Delay=100ms
Normal & Heavy load
Transaction Arrival Rate (no. per second)
2 4 6 8 10 12 14
Miss%
0
20
40
60
80
SWIFT
SWIFT with Partial Read Optimization
Disk Resident Database
Commit Protocol-SWIFT contd
C it P t l
-
7/29/2019 SWIFT Presentation
59/76
Dept. of CSE, MMMEC Gorakhpur59
Fig. 4.19: Miss % with (RC+DC) at Communication Delay=0ms
Normal Load
Transaction Arrival Rate (no. per second)
3 4 5 6
Miss%
0
5
10
15
20
SWIFT
SWIFT with Partial Read Optimization
Disk Resident Database
Commit Protocol-SWIFT contd
C it Pr t l S T
-
7/29/2019 SWIFT Presentation
60/76
Dept. of CSE, MMMEC Gorakhpur60
Fig. 4.20: Miss % with (RC+DC) at Communication Delay=0msHeavy Load
Transactional Arrival rate (no. per second)
6 9 12 15 18
M
iss%
0
20
40
60
80
100
SWIFT
SWIFT with Partial Read Optimization
Disk Resident Database
Commit Protocol-SWIFT contd
Commit Protocol SWIFT
-
7/29/2019 SWIFT Presentation
61/76
Dept. of CSE, MMMEC Gorakhpur61
Announcement of the abort of a cohort can be directly sent to its siblingas well as its coordinator.
No need for coordinator to send the abort message to rest of its cohorts.
1% to 3% Improvement in Transaction Miss Percentage (very limited)
Effect o f Perm it t ing Cohorts to Commun icate With
Each Other of the Same Transaction (CCST)
Commit Protocol-SWIFT contd
Commit Protocol SWIFT
-
7/29/2019 SWIFT Presentation
62/76
Dept. of CSE, MMMEC Gorakhpur62
Fig. 4.21: Miss % with (RC+DC) at Communication Delay=0msNormal & Heavy Load
Transaction Arrival Rate (no. per second)
10 15 20 25 30 35 40 45 50
Miss%
0
20
40
60
80
SWIFT
SWIFT with CCST
Main Memo ry Resident Database
Commit Protocol-SWIFT contd
Commit Protocol SWIFT
-
7/29/2019 SWIFT Presentation
63/76
Dept. of CSE, MMMEC Gorakhpur63
Fig. 4.22: Miss % with (RC+DC) at Communication Delay=100msNormal & heavy Load
Transaction Arrival Rate (no. per second)
2 4 6 8 10 12 14
Miss
%
0
20
40
60
80
SWIFT
SWIFT with CCST
Main Memory Resident Database
Commit Protocol-SWIFT contd
Commit Protocol SWIFT
-
7/29/2019 SWIFT Presentation
64/76
Dept. of CSE, MMMEC Gorakhpur64
Fig. 4.23: Miss % with (RC+DC) at Communication Delay=100msNormal & Heavy load
Transaction Arrival Rate (no. per second)
2 4 6 8 10 12 14
Mis
s%
0
20
40
60
80
SWIFT
SWIFT with CCST
Disk Resident Database
Commit Protocol-SWIFT contd
Commit Protocol SWIFT
-
7/29/2019 SWIFT Presentation
65/76
Dept. of CSE, MMMEC Gorakhpur65
Fig. 4.24 Miss % with (RC+DC) at Communication Delay=0ms
Normal Load
Transaction Arrival Rate (no. per second)
3 4 5 6
M
iss%
0
5
10
15
20
SWIFT
SWIFT with CCST
Disk Resident Database
Commit Protocol-SWIFT contd
Commit Protocol SWIFT d
-
7/29/2019 SWIFT Presentation
66/76
Dept. of CSE, MMMEC Gorakhpur66
Fig. 4.25: Miss % with (RC+DC) at Communication Delay=0msHeavy Load
Transactional Arrival rate (no. per second)
6 9 12 15 18
M
iss%
0
20
40
60
80
100
SWIFTSWIFT with CCST
Disk Resident Database
Commit Protocol-SWIFT contd
Commit Protocol SWIFT d
-
7/29/2019 SWIFT Presentation
67/76
Dept. of CSE, MMMEC Gorakhpur67Fig. 4.26: Break-up of Miss % with (RC+DC) at Communication Delay=100
Transaction Arrival Rate (no. per second)
0 2 4 6 8 10 12 14 16
Miss%
0
10
20
30
40
50Total Transaction Miss %
Transaction Miss % During Processing Phase
Transaction Miss % Other Than Processing Phase
Impact o f Early Send ing of WORKSTARTED MessageMain Memory Database with Commun icat ion Delay of 100ms
Commit Protocol-SWIFT contd
Commit Protocol SWIFT d
-
7/29/2019 SWIFT Presentation
68/76
Dept. of CSE, MMMEC Gorakhpur68Fig. 4.27: Break-up of Miss % with (RC+DC) at Communication Delay=0ms
Transaction Arrival Rate (no. per second)
10 15 20 25 30 35 40 45 50 55
Miss%
0
10
20
30
40
50
60
70
Total Transaction Miss %
Transaction Miss % During Processing Phase
Transaction Miss % Other Than Processing Phase
Main Memo ry Database with Communicat ion Delay of 0ms
Commit Protocol-SWIFT contd
Commit Protocol SWIFT d
-
7/29/2019 SWIFT Presentation
69/76
Dept. of CSE, MMMEC Gorakhpur69Fig. 4.28: Break-up of Miss % with (RC+DC) at Communication Delay=100
Transaction Arrival Rate (no. per second)0 2 4 6 8 10 12 14 16
Miss%
0
20
40
60
80
Total Transaction Miss %
Transaction Miss % During Processing Phase
Transaction Miss % Other Than Processing Phase
Disk Resident Database with Commun icat ion Delay of 100ms
Commit Protocol-SWIFT contd
Commit Protocol-SWIFT td
-
7/29/2019 SWIFT Presentation
70/76
Dept. of CSE, MMMEC Gorakhpur70
Commit Protocol-SWIFT contd
Fig. 4.29: Break-up of Miss % with (RC+DC) at Communication Delay=0ms
Transaction Arrival Rate (no. per second)
4 6 8 10 12 14 16 18 20
Miss%
0
20
40
60
80
100
Total transaction Miss %
Transaction Miss % During Processing Phase
Transaction Miss % Other Than Processing Phase
Disk Resident Database with Commun icat ion Delay of 0ms
Conclusions
-
7/29/2019 SWIFT Presentation
71/76
Dept. of CSE, MMMEC Gorakhpur71
ConclusionsSWIFT-A New Real Time Comm it Protoc ol
Performances Comparison with 2SC and PROMPT When
Communication Delays-Negligible or Large 5% to 10% Improvement in Transaction Miss Percentage
Performances Comparisonfor Partial Read-Only Optimization
1% to 5% Improvement in Transaction Miss Percentage
Impact of Permitting Communication between Cohorts of SameTransaction (Sibling)
Up to 3% Improvement in Transaction Miss Percentage
Scope for Future Research
-
7/29/2019 SWIFT Presentation
72/76
Dept. of CSE, MMMEC Gorakhpur72
Scope for Future Research
Performance Evaluationof Proposed Priority Assignment
Policy and Commit Protocols on DRTDBS by
Analytical Methods
Experimentation in Actual Environment
Experimentation in Replicated Environment
Performance Evaluation of Proposed Commit Protocols
using 1PC Protocol
3PC Protocol
Performance Evaluationof Proposed Works in
Hard Real Time Environment Soft Real Time Environment
Scope for Future Research contd
-
7/29/2019 SWIFT Presentation
73/76
Dept. of CSE, MMMEC Gorakhpur73
Scope for Future Research contd
Performance EvaluationofSWIFT in
Mobile DTRDBS
An Obvious Extension of Our Work for
Multiprocessor Environment
Fault Tolerance and Reliability Aspects
Impact of Communications in between Cohorts of
Same Transaction (Siblings) on Overall System
Performance.
Extension of Our Research Work
for Grid DatabaseSystems
References
-
7/29/2019 SWIFT Presentation
74/76
Dept. of CSE, MMMEC Gorakhpur74
1. Udai Shanker, Manoj Misra and Anil K. Sarje, SWIFT-A New Real Time Commi t
Protoco l, International Journal of Distributed and Parallel Databases, Springer
Verlag (online on May 26, 2006).
2. Udai Shanker, Manoj Misra and Anil K. Sarje, Distr ib uted Real Time Database
Systems: Back grou nd and L i terature Review, International Journal of Distributed
and Parallel Databases, Springer Verlag (under second review).
3. Udai Shanker, Manoj Misra and Anil K. Sarje,Dependency Sensi t ive Distr ibu ted
Comm it Protocol, Proceedings of the 8th International Conference on Information
Technology (CIT 05), Bhubaneswar, India, Dec. 20-23, 2005, pp. 41-46.
4. Udai Shanker, Manoj Misra and Anil K. Sarje,A Memory Eff ic ient Fast Distr ibuted
Real Time Commit Protoc ol,Proceedings of the 7th International Workshop on
Distributed Computing (IWDC 2005), Indian Institute of Technology Kharagpur, India,
Dec. 27-30, 2005, pp 500-505.
5. Udai Shanker, Manoj Misra and Anil K. Sarje,Optimiz ing Distr ibuted Real-Time Transact ion Proc essing Dur ing Execut ion Phase,Proceedings of the 3rd
International Conference on Computer Application (ICCA2005), University of
Computer Studies, Yangon, Myanmar, March 9-10, 2005, pp 1-7.
References contd
-
7/29/2019 SWIFT Presentation
75/76
Dept. of CSE, MMMEC Gorakhpur75
6. Udai Shanker, Manoj Misra and Anil K. Sarje, Some Performance Issues in
Distr ibuted Real Time Database Systems, Proceedings of the VLDB PhD
Workshop, The Convention and Exhibition Center (COEX), Seoul, Korea, Sept. 11,
2006.7. Udai Shanker, Some Performance Issues in Distr ibuted Real Time Database
Systems, PhD Thesis, Department of Electronics & Computer Engineering, Indian
Institute of Technology Roorkee, Roorkee-247 667, India, June 2006.
8. Gray Jim and Reuter A., Transaction Processing : Concepts andTechnique,
Morgan Kaufman, San Mateo, CA, 1993.
9. Gray Jim, Notes on Database OperatingSystems, Operat ing Systems: an
AdvancedCourse, Lecture Notes in Computer Science, Springer Verlag, Vol. 60,
pp. 397 - 405, 1978.
10. Lam Kam - Yiu, Concurrency Control in Distr ibuted Real - Time Database
Systems, PhD Thesis, City University of Hong Kong, Hong Kong, Oct. 1994.
11. Ulusoy Ozgur, Concurrency Con trol in Real - t ime DatabaseSystems, PhDThesis, Department of Computer Science, University of Illinois, Urbana-Champaign,
1992.
Questions and Answers
-
7/29/2019 SWIFT Presentation
76/76
Thank You