1. Distributed Query Processing using different Semijoin operations.
Presented By:
Jamal Uddin Ahamed
Friday,March12,2004

2. Presentation Outline:
1.Overview.
2.Semijoin Operation.
3. Different semijoin operations.
a. 2 way semijoin.
b.Hash Semijoin.
c.Domain Specific Semijoin.
d. Composite semijoin.
4. References.
5.Questions and Answer.

3. 1.1 What is distributed database system?
A distributed database system is characterized by the distribution of the system components of hardware ,control and data. For this research, a distributed system is a collection of independent computers interconnected via point-to-point communication lines.

4. 1.2 Node Characteristics:
Each computer , known as a node in the
network, has a processing capability, a
data storage capability, and is capable
of operating autonomously in the system.
Each node contains a version of a
distributed DBMS.

5. 1.3 What is distributed query processing?
The retrieval of data from different sites in a network is known as distributed query processing.

6. 1.4 Phases of distributed query processing with a semijoin operator.
Initial Local processing (Selections and Projects are processed at each site.)
Semijoin processing ( A semijoin program) is derived from the remaining join operations and executed to reduce the size of the relations in a cost-effective way)
Final processing (all relations involved are transmitted to final site and all joins are performed there.)

7. 2.1 Semijoin:
A semijoin from Ri to Rj on attribute A can be denoted as Rj⋉ Ri .It is used to reduce the data transmission cost.
Computing steps:
Project Ri on attribute A (Ri[A] ) and ship this projection ( a semijoin projection) from the site of Ri to the site of Rj ;
Reduce Rj to Rj’ by eliminating tuples where attribute A are not matching any value in Ri[A] .

8. 2.2 Example: Example (semijoin s: R1—AR2): Site 2 Site 1 qs3 4 5 7 8 9 A C R2 B 1 2 6 R1
Site 1
Site 2 1 2 3
R1[A]
projection
Ship(3) 3 7
R2’
reduce qs
Ship(2)
Ship(6)
Benefit (s) = 6 -2 = 4
Cost (s) = 3
Cost effectiveness D(s) = B(s)-C(s) >0

9. 3.a.1 Definition of 2 way semijoin.
2-way Semijoin—an extended version of the semijoin
Definition: A 2-way semijoin (t) of Ri and Rj on attribute A can be denoted as
RiARj = {Ri—ARj, Rj—ARi }
So t reduces Ri and Rj to Ri’ and Rj’ respectively.

10. 3.a.2 Properties of 2 way semijoin.
Computing steps:
Send Ri [A] from site i to site j ;
Reduce Rj to Rj’ by eliminating tuples whose attribute A are not matching any of Ri [A] and at the same time partition Ri [A] to Ri [A]m (match one of Rj [A]) and Ri [A]nm(Ri [A]- Ri [A]m) ;
Send min(Ri [A]m , Ri [A]nm) back to site i ;
Reduce Ri to Ri ’ using Ri [A]m (or Ri [A]nm) .
Evaluation:
Benefit: B(t) = [S(Ri ) - S(Ri ’)] + [S(Rj) - S(Rj’)]
Cost: C(t) = S(Ri [A] ) + min[S(Ri [A]m ) , S( Ri [A]nm)]
If the benefit exceeds the cost (D(t) >0) then it is called a cost-effective 2-way semioin

11. 3.a.3 2-way semijoin example.1 2 3
R1[A]
projection
Ship(3) 3 4 5 7 8 9 A C R2 B 1 2 6 R1
Site 1
Site 2 3
R1[A]m 1 2
R1[A]nm
partition 7
R2’
reduce
Ship(1) 3 6
R1’
reduce qs
Ship(2)

12. 3.a.4 Semijoin Vs 2-way semijoin.
-It is an extended version of semijoin.
It has more reduction power than semijoin.
The propagation of reduction effects by the 2-way semijoin is further than by the semijoin.

13. 3.b.1 Hash-semijoin operator.
Main idea : use a search filter which represents the semijoin projection with a small bit array .
Definition:
The hash-semijoin of Ri and Rj is denoted Rj∝ Ri. It is computed as follow:
The Semijoin projection of Ri is represented as a bit array;
Shipping this bit array to the site of Rj ;
finally, the tuples of Rj are screened by the search filter.

14. 3.b.2 hash semijoin example.R2 R1 1 B
H(x)=X
Hij((Ri))Bij
S#(R1) 1 3 4 8
projection S#
Phone 2
222 3
333 4
444 5
555 6
666
Ship(Bij) Rj S#
Name 1
Cindy 3
Jemal 4
Sunny 8
Maggie
reduce 3
333 4
444

15. 3.b.3 Semijoin Vs Hash Semijoin.
Advantages:
Hash-semijoin is more cost-effective than semijoin
The search filter in the hash-semijoin achieves considerable savings in the cost of a semijoin operation
Limitation:
Only works on execution tree
Tightly related with the hash functions

16. 3.c.1 What is horizontally partitioned database
We can call a distributed database system is horizontally partitioned (or fragmented) if the relations can be split horizontally into several disjoint sets of tuples, which are called horizontal fragments.

17. 3.c.2 Horizontally partitioned database system.(Example)
EMP1: 1D-no 10
EMP
E-no
E-name
D-no
101
johnson 01
103
jordan 03
105
erving
E-no
E-name
D-no
101
johnson 01
103
jordan 03
105
erving
109
jabbar 12
110
sampson 14
141
chang 16 
EMP2: 11D-no 20
E-no
E-name
D-no
109
jabbar 12
110
sampson 14
141
chang 16

18. 3.c.3 Horizontally partitioned database system.(Properties)
A fragmented relation Ri can be constructed by performing a union operation on all its fragment.
Ri = Uk Rik
There is commutative rule between the binary operations join and union for fragmented relations: a join between two fragmented relation R1 and R2 is equivalent to a union over the joins between each fragment of R1 and each fragment of R2.
Mathematically:
(U R1k)[A=B] (U R2m)= U(R1k[A=B] R2m)
k m k.m

19. 3.c.4 Why can’t we use regular semjoin between two fragment to reduce the size of fragments?(Continue)
We consider a joint Ri[A=B] Rj between two fragmented relations Ri and Rj. We want to reduce the size of Rik, a fragment of Ri , by semijoin before it is sent to the final processing site. We cannot perform the semijoin
Rik A=B] Rjm
between Rik and any fragment Rjm of Rj without considering the other fragment Rjm of Rj , because the join operation dictates that no tuple of a relation can be eliminate before it is compare with all tupls of the other joining relation which may be contribute to the join.

20. Example: sal: 101E-no 105 EMP1: 1D-no 10 EMP2: 11D-no 20
E-name
D-no
101
johnson 01
103
jordan 03
135
erving
E-no
Sal
D-no
101
1000 12
102
2000 03
105
3000 11
D-no 01 03 12 14 16
EMP2: 11D-no 20
sal: 105E-no 110
E-no
E-name
D-no
109
jabbar 12
110
sampson 14
141
chang 16
E-no
Sal
D-no
107
1000 12
2000 03
110
3000 11

21. 3.c.5 Definition of Domain Specific Semijoin.
The domain-specific semijoin operation, Rik( A=B] Rjm, where A and B are the joining attributes and Rik, Rjm are two fragments of the joining relation Ri and Rj respectively, is defined as follows:
Rik( A=B] Rjm ={r|r Rik ; r.A  Rjm [B] U(Dom[Rj.B]-Dom[Rjm.B])}
Where Rik is the restricted fragment and Rjm is the restricting fragment. We also called Ri the restricted relation and Rj is the restricting relation of the domain-specific semijoin.

22. 3.d.1 Definition of Composite Semijoin.
Composite Semijoin: a semijoin in which the projection and the transimssion involve multiple columns (attrs).

23. 3.d.2 Example of Composite Semijoin.R2 R1 A1 A2
Non-join Attr 1 aa - bb 2 cc 3 A1 A2
Non-join Attr 1 cc - aa 2 bb 3 A1 A2
Non-join Attr 1 aa -
No False loop!!

24. 3.d.3 Semijoin Vs Composite Semijoin.
Composite semijoins in a query processing algorithm is likely to result in substantial RT reduction.
Composite semijoins should not always be used. If it results greater RT, ignore it.
Strategy with composite semijoins is at least as good as that without composite semijoins.

25. References:
Using 2-way semijoin in distributed query processing. By Hyunchul Kang and Nick Roussopoulos.
Improving distributed query processing by hash-semijoins. By Judy Tseng and Arbee Chen.
Domain Specific Semijoin:A new operation for distributed query processing. By Jason Chen and Victor Li.
Composite Semijoin in distributed query processing. By William Perrizio and Chun Chen

26. Comments
&
Questions??
Thank You!