| Database Systems: The Complete
Book
| Solutions for Chapter 15 |
|
Solutions for Section 15.2
Solutions for Section 15.3
Solutions for Section 15.4
Solutions for Section 15.5
Solutions for Section 15.6
Solutions for Section 15.7
Solutions for Section 15.8
Solutions for Section 15.9
Solutions for Section 15.2
Exercise 15.2.1(a)
The iterator for pi_L(R) can be described informally as follows:
- Open():
-
R.Open()
- GetNext():
-
Let t = R.GetNext().
If Found = False, return.
Else, construct and return the tuple consisting of the elements of list
L evaluated for tuple t.
- Close():
-
R.Close()
Exercise 15.2.1(b)
Here is the iterator for delta(R):
- Open():
-
First perform R.Open().
Then, initialize a main-memory structure S that will represent a
set of tuples of R seen so far.
For example, a hash table could be used.
Initially, set S is empty.
- GetNext():
-
Repeat R.GetNext() until either Found = False or a
tuple t that is not in set S is returned.
In the later case, insert t into S and return that tuple.
- Close():
-
R.Close()
Exercise 15.2.1(d)
For the set union R1 union R2 we need both a set S that
will store the tuples of R1 and a flag CurRel as in
Example 6.13.
Here is a summary of the iterator:
- Open()
-
Perform R1.Open() and initialize to empty the set S.
- GetNext()
-
If CurRel = R1 then perform t = R1.GetNext().
If Found = True, then insert t into S and return
t.
If Found = False, then R1 is exhausted.
Perform R1.Close(),
R2.Open(), set CurRel = R2,
and repeatedly perform t = R2.GetNext()
until either Found = False (in which case, just return), or
t is not in S.
In the latter case, return t.
- Close():
-
Perform R2.Close()
Exercise 15.2.4(a)
Read R into main memory.
Then, for each tuple t of S, see if t joins with
any of the tuples of R that are in memory.
Output any of those tuples that do join with t, and delete them
from the main-memory set.
Exercise 15.2.4(b)
Read S into memory.
Then, for each tuple t of R, see if t joins with
any tuple of S in memory; output t if so.
Exercise 15.2.4(e)
Read R into main memory.
Then, for each tuple t of S, find those tuples in memory
with which t joins, and output the joined tuples.
Also mark as ``used'' all those tuples of R that join with
t.
After S is exhausted, examine the tuples in main memory, and
for each tuple r that is not marked ``used,'' pad r with
nulls and output the result.
Return to Top
Solutions for Section 15.3
Exercise 15.3.2
The formula in Section 15.3.4 gives 10000 + 10000*10000/999 or 110,101.
Return to Top
Solutions for Section 15.4
Exercise 15.4.2(a)
An iterator for delta(R) must do the first sorting pass in its
Open() function.
Here is a sketch:
- Open():
-
Perform R.Open() and repeatedly use R.GetNext() to
read memory-sized chunks of R.
Sort each into a sorted sublist and write out the list.
Each of these lists may be thought of as having its own iterator.
Open each list and initialize a data structure into which the
``current'' element of each list may be read into main memory.
Use GetNext() for each list to initialize the main-memory
structure with the first element of each list.
Finally, execute R.Close().
- GetNext():
-
Find the least tuple t in the main-memory structure, and output one
copy of it.
Delete all copies of t
in main memory, using GetNext() for the
appropriate sublist, until all copies of t have disappeared, then
return.
If there was no such tuple t, because all the sublists were
exhausted, set Found = False and return.
- Close():
-
Close all the sorted sublists.
Exercise 15.4.2(c)
For R1 intersect R2 do the following:
- Open():
-
Open R1 and R2, use their GetNext() functions to
read all their tuples in main-memory-sized chunks, and create sorted
sublists, which are stored on disk.
Initialize a main-memory data structure that will hold the ``current''
tuple of each sorted sublist, and use the Open() and
GetNext() functions
of an iterator from each sublist to initialize the structure to have the
first tuple from each sublist.
Finally, close R1 and R2.
- GetNext():
-
Find the least tuple t among all the first elements of the sorted
sublists.
If t occurs on a list from R1 and a list from R2,
then output t, remove the copies of t, use
GetNext from the two sublists involved to replentish the
main-memory structure, and return.
If t appears on only one of the sublists, do not output t.
Ratherm remove it from its list, replentish that list, and repeat these
steps with the new least t.
If no t exists, because all the lists are exhausted, set
Found = False and return.
- Close():
-
Close all the sorted sublists.
Exercise 15.4.3(b)
The formula of Fig. 15.11 gives 5*(10000+10000) = 100,000.
Note that the memory available, which is 1000 blocks, is larger than the
minimum required, which is sqrt(10000), or 100 blocks.
Thus, the method is feasible.
Exercise 15.4.5(a)
We effectively have to perform two nested-loop joins of 500 and 250
blocks, respectively, using 101 blocks of memory.
Such a join takes 250 + 500*250/100 = 1500 disk I/O's, so two of them
takes 3000.
To this number, we must add the cost of sorting the two relations, which
takes four disk I/O's per block of the relations, or another 6000.
The total disk I/O cost is thus 9000.
Exercise 15.4.7(a)
The square root of the relation size, or 100 blocks, suffices.
Return to Top
Solutions for Section 15.5
Exercise 15.5.1(a)
To compute delta(R) using a ``hybrid hash'' approach, pick a
number of buckets small enough that one entire bucket plus one block for
each of the other buckets will just fit in memory.
The number of buckets would be slightly larger than B(R)/M.
On the first pass, keep all the distinct tuples of the first bucket in
main memory, and output them the first time they are seen.
On the second pass, do a one-pass distinct operation on each of the
other buckets.
The total number of disk I/O's will be one for M/B(R) of the data
and three for the remaining (B(R)-M)/B(R) of the data.
Since the data requires B(R) blocks, the total number of disk
I/O's is M + 3(B(R) - M) = 3B(R) - 2M, compared with 3B(R)
for the standard approach.
Exercise 15.5.4
This is a trick question.
Each group, no matter how many members, requires only one tuple in main
memory.
Thus, no modifications are needed, and in fact, memory use will be
pleasantly small.
Return to Top
Solutions for Section 15.6
Exercise 15.6.1(a)
Our option is to first copy all the tuples of R to the output.
Then, for each tuple t
of S, retrieve the tuples of R for
which R.a matches the a-value of t.
Check whether t appears among them, and output t if not.
This method can be efficient if S is small and V(R,a) is
large.
Exercise 15.6.2(a)
The values retrieved will fit on about 10000/k blocks, so that is
the approximate cost of the selection.
Exercise 15.6.5
With the index, we have only to retrieve the blocks
containing MovieStar records for the stars of ``King Kong.''
We don't know how many stars there are, but it is safe to assume that
their records will occupy many more blocks than there are stars in the
two ``King Kong'' movies. Thus, using the index allows us to take the
join while accessing only a small part of the MovieStar
relation and is therefore to be preferred to a sort- or hash-join, each
of which will access the entire MovieStar relation at least
once.
Return to Top
Solutions for Section 15.7
Exercise 15.7.1(a)
One of the relations must fit in memory, which means that
min(B(R),B(S)) >= M/2.
Exercise 15.7.1(b)
We can only assume that there will be M/2 blocks available during
the first pass, so we cannot create more than M/2-1 buckets.
Each pair of matching buckets must be joinable in a one-pass join,
which means one of each pair of corresponding buckets must be no larger
than M/2, as we saw in part (a).
Since the average bucket size will be 1/(M/2 - 1) of the size of
its relation, we require
min(B(R),B(S))/(M/2 - 1) >= M/2, or approximately
min(B(R),B(S)) >= M^2/4.
This figure should be no surprise; it simply says we must make the most
conservative assumption at all times about the size of memory.
Return to Top
Solutions for Section 15.8
Exercise 15.8.1(a)
The way we have defined the recursive algorithm, it always divides
relations into M-1 pieces (100 pieces in this exercise)
if it has to divide at all.
In practice, we may have options to use fewer pieces at one or more
levels or the recursion, and in this exercise there are many such
options.
However, we'll describe only the one that matches the policy given in
the text.
In what follows, all sorting and merging is based on the values of
tuples in the join attributes only, since we are describing a join
operation.
Level 1: R and S are divided into 50 sublists each.
The sizes of the lists are 400 and 1000 blocks, for R and S,
respectively.
Note that no disk I/O's are performed yet; the partition is conceptual.
Level 2: Each of the sublists from Level 1 are divided into 100
subsublists; the sizes of these lists are 4 blocks for R, and 10 blocks
for S.
Again, no disk I/O's are performed for this conceptual partition.
Level 3: We sort each of the subsublists from Level 2.
Each block is read once from disk, and an equal number of blocks are
written to disk as the sorted subsublists.
Return to Level 2: We merge each of the 100 subsublists that comprise
one sublist, using the standard multiway merging algorithm.
Every block is read and written once during this process.
Return to Level 1: We merge each of the 50 sorted sublists from R and
the 50 sorted sublists from S, thus joining R and S.
Each block is read once during this part.
Return to Top
Solutions for Section 15.9
Exercise 15.9.1(a)
If the division of tuples from R is as even as possible, some
processors will have to perform 13 disk I/O's, but none will have to
perform more.
The 13 disk I/O's take 1.3 seconds, so the speedup is 10/1.3 or about
7.7.
Return to Top