Skip to content

Instantly share code, notes, and snippets.

@AndreaCrotti
Created December 4, 2009 10:45
Show Gist options
  • Save AndreaCrotti/248980 to your computer and use it in GitHub Desktop.
Save AndreaCrotti/248980 to your computer and use it in GitHub Desktop.
\documentclass[11pt]{article}
\usepackage[utf8]{inputenc}
\usepackage[T1]{fontenc}
\usepackage{graphicx}
\usepackage{longtable}
\usepackage{float}
\usepackage{wrapfig}
\usepackage{soul}
\usepackage{amssymb}
\usepackage{hyperref}
\title{EXERCISE 6}
\author{Andrea Crotti (\#299466), Can Liu (\#286105), Sebastian Roidl (\#281941)}
\date{04 Dicembre 2009}
\begin{document}
\maketitle
\obeylines
\section*{}
\texttt{DEADLINE:} \textit{2009-12-02 Mer 12:15}\newline
\section*{1. Selection and projection}
\label{sec-2}
\subsection*{1. Given the size of a relation (1000 pages), what are the I/O costs for an equality selection on a non-key attribute for the following cases? Assuming each B-tree leaf holds 20 record pointers while each page contains 10 records}
\label{sec-2.1}
In worst case equality selection must be done scanning the whole data and checking over the key we need to compare.
Using indexes or (un)clustered B+-trees can reduce the number of I/O operations.
\subsubsection*{(a) with a clustered b-tree of height 3 (matching records are located in one page)}
\label{sec-2.1.1}
This is the best scenario, a clustered B+Tree records all the data in a sorted order on disk.
$cost = h + 1 = 4$ I/Os, we only need traverse the tree from root to leaf and then read the data page.
\subsubsection*{(b) without any index, nor is the file sorted on the attribute occurring in selection}
\label{sec-2.1.2}
This is the worst possible case, we need to scan over all the records and check the equality of the attribute.
So the number of I/Os needed is the total number of pages of the relation.
$cost = 1000$ I/Os
\subsubsection*{(c) with an unclustered b-tree index of height 3, and there are 10 matching records}
\label{sec-2.1.3}
We scan through the tree and in plus we must read the pages which this time are not contiguous as before.
$cost = h + k = 13$ I/Os
\subsubsection*{(d) with an unclustered b-tree of height 3 and one tenth of the records match the selection}
\label{sec-2.1.4}
Same as before, having more records matching the selection makes the unclustered b+-tree much less convenient.
$cost = h + #tuples/10 = 3 + 1000 = 10003$ I/Os
If the matching rate is higher than it would become even faster to do a total lookup on the pages.
\subsection*{2. We assume the relation is of N pages.}
\label{sec-2.2}
\subsubsection*{(a) What is the complexity of sorting-based projection?}
\label{sec-2.2.1}
When doing a \emph{projection} we must
\begin{itemize}
\item remove unnecessary attributes
\item delete all the possible duplicates that we generated
\end{itemize}
If we use a sorting-based algorithm then all the duplicates can only be adjacent, and it becomes very easy to detect and delete them.
The complexity is then $O(N \log{N})$, where for N we denote the number of pages.
\subsubsection*{(b) If the records are distributed uniformly, what is the complexity for hash-based projection?}
\label{sec-2.2.2}
For hash-based projection the best-case is $O(N)$.
To be more precise we have to compose the cost of partitioning and comparing, so we get
$partCost + inBucketCompCost = (N + N) + N = 3N$
To have good performances with hash-based projections the buckets must fit in memory.
\subsubsection*{(c) Why is sorting-based projection the standard solution in practice (or why is it superior to hash-based projection)?}
\label{sec-2.2.3}
In general a sorting-based projection is used because:
\begin{itemize}
\item sorting is a basic routine in database implementations and the code is fairly optimized
\item hashing can fail if the bucket is too large
\item hashing-based projection can become very slow if the data are not distributed uniformly
\end{itemize}
\section*{2. Join}
\label{sec-3}
Consider $Order \Join_{Order.cid=Customer.id} Customer$.
Cost metric is the number of page I/O.
\begin{itemize}
\item Order contains 10,000 tuples with 10 tuples per page.
\item Customer contains 2,000 tuples and 10 tuples per page.
\item available 52 buffer pages
\item each Customer tuple matches 5 tuples in Order
\item it takes 3 I/O operations to access a leaf of $B^{+}- tree$ for Order, 2 I/Os for Customer.
\end{itemize}
Formalizing the following data are given:
\begin{itemize}
\item M = 1000 (number of pages in \emph{Order})
\item N = 200 (number of pages in \emph{Customer})
\item Po = 10 (number of tuples/page in \emph{Order})
\item Pc = 10 (number of tuples/page in \emph{Customer})
\item B = 52 (number of buffer pages)
\item Ord\_for\_Cust = 5 (tuples of Order matching any Customer)
\end{itemize}
\subsection*{1. What is the cost of joining Order and Customer using a page-oriented simple nested loops join?}
\label{sec-3.1}
The Order relation is the outer relation, while the customer relation is inner.
Since in this case the join is page-oriented, the cost is:
$M + (Po * M * N) = 1000 + (10 * 1000 * 200) \approx 2 * (10^6)$ I/Os
\subsection*{2. What is the cost of joining Order and Customer using a block nested loops join?}
\label{sec-3.2}
Having
One buffer page is for scanning \emph{Customer} (inner relation), so we use (52-2) buffer pages for holding the blocks of \emph{Order} (outer relation).
The cost is:
$M + \lceil M / (B - 2)\rceil * N = 1000 + \lceil 1000 / 50 \rceil * 200 = 5000$ I/Os
\subsection*{3. What is the cost of joining Order and Customer using a sort-merge join?}
\label{sec-3.3}
First we need to sort the two relations. So we get cost: $10000 * log10000 + 2000 * log2000$
Order is scanned once, each Customer group is scanned once per matching Order tuple.
So we get the final cost: $M * logM + N * logN + M + N = 1000 * log1000 + 200 * log200 + 1200$ I/Os
\subsection*{4. What is the cost of joining Order and Customer using a hash join?}
\label{sec-3.4}
We first partition both relations, which means we read and write both relations, so we have cost of 2 * (M + N) = 2 * 1200 I/Os.
Then we scan matching partitions, read both relations, so we have cost of 1000 + 200 I/Os.
$Cost = Cost(Partitioning) + Cost(Matching) = 2400 + 1200 = 3600$ I/Os
\subsection*{5. If unclustered B+tree index existed on Order.cid or Customer.id,would either provide a cheaper alternative (index nexted loop join) for performing the join than a block nested loop join?}
\label{sec-3.5}
We can use the index on the join column of one relation and make it the inner loop (that's where you lose more time).
So we get:
$Cost = M + (M * Po * Ord\_for\_Cust) = 1000 + (1000 * 10 * 5) = 51000$
In this case we don't get any advantage using a B+tree.
\subsection*{6. Reconsider the question above when the B+−tree index is clustered.}
\label{sec-3.6}
Here the only difference is that we can access to the matching Customer's tuples in just 1 I/O because of the clustering.
So we get:
$Cost = M + (M * Po * 1) = 1000 + 10000 = 11000$ I/Os
\end{document}
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment