CSCI 3323 (Principles of Operating Systems), Fall 2017:
Homework 5

Credit:

55 points.

Reading

Be sure you have read, or at least skimmed, Chapter 3.

Honor Code Statement

Please include with each part of the assignment the Honor Code pledge or just the word ``pledged'', plus one or more of the following about collaboration and help (as many as apply).¹Text in italics is explanatory or something for you to fill in. For written assignments, it should go right after your name and the assignment number; for programming assignments, it should go in comments at the start of your program(s).

• This assignment is entirely my own work. (Here, ``entirely my own work'' means that it's your own work except for anything you got from the assignment itself -- some programming assignments include ``starter code'', for example -- or from the course Web site. In particular, for programming assignments you can copy freely from anything on the ``sample programs page''.)
• I worked with names of other students on this assignment.
• I got help with this assignment from source of help -- ACM tutoring, another student in the course, the instructor, etc. (Here, ``help'' means significant help, beyond a little assistance with tools or compiler errors.)
• I got help from outside source -- a book other than the textbook (give title and author), a Web site (give its URL), etc.. (Here too, you only need to mention significant help -- you don't need to tell me that you looked up an error message on the Web, but if you found an algorithm or a code sketch, tell me about that.)
• I provided help to names of students on this assignment. (And here too, you only need to tell me about significant help.)

Problems

Answer the following questions. You may write out your answers by hand or using a word processor or other program, but please submit hard copy, either in class or in one of my mailboxes (outside my office or in the ASO).

1. (5 points) The operating system designers at Acme Computer Company have been asked to think of a way of reducing the amount of disk space needed for paging. One person proposes never saving pages that only contain program code, but simply paging them in directly from the file containing the executable. Will this work always, never, or sometimes? If ``sometimes'', when will it work and when will it not? (Hint: Search your recollections of CSCI 2321 -- or another source -- for a definition of ``self-modifying code''.)

2. (5 points) How long it takes to access all elements of a large data structure can depend on whether they're accessed in contiguous order (i.e., one after another in the order in which they're stored in memory), or in some other order. The classic example is a 2D array, in which performance of nested loops such as

   	for (int r = 0; r < ROWS; ++r)
   	  for (int c = 0; c < COLS; ++c)
   	    array[r][c] = foo(r,c);
   

   can change drastically for a large array if the order of the loops is reversed. Give two explanations for this phenomenon based on what you have learned from our discussion of memory management. (Hint: One possible explanation is based on a topic we discussed extensively but that on current systems is less likely than it was before huge amounts of RAM became common. The currently-more-likely explanation is one we touched on but did not discuss extensively.)

3. (10 points) Consider (imagine?) a very small computer system with only four page frames. Suppose you have implemented the aging algorithm for page replacement, using 4-bit counters and updating the counters after every clock tick, and suppose the $R$ bits for the four pages are as follows after the first four clock ticks.

   Time	$R$ bit (page 0)	$R$ bit (page 1)	$R$ bit (page 2)	$R$ bit (page 3)
   after tick 1	0	1	1	1
   after tick 2	1	0	1	1
   after tick 3	1	0	1	0
   after tick 4	1	1	0	1

   What are the values of the counters (in binary) for all pages after these four clock ticks? If a page needed to be removed at that point, which page would be chosen for removal?

4. (10 points) A computer at Acme Company used as a compute server (i.e., to run non-interactive jobs) is observed to be running slowly (turnaround times longer than expected). The system uses demand paging, and there is a separate disk used exclusively for paging. The sysadmins are puzzled by the poor performance, so they decide to monitor the system. It is discovered that the CPU is in use about 20% of the time, the paging disk is in use about 98% of the time, and other disks are in use about 5% of the time. They are particularly puzzled by the CPU utilization (percentage of time the CPU is in use), since they believe most of the jobs are compute-bound (i.e., much more computation than I/O). First give your best explanation of why CPU utilization is so low, and then for each of the following, say whether it would be likely to increase it and why.
   1. Installing a faster CPU.

   2. Installing a larger paging disk.

   3. Increasing the number of processes (``degree of multiprogramming'').

   4. Decreasing the number of processes (``degree of multiprogramming'').

   5. Installing more main memory.

   6. Installing a faster paging disk.

Programming Problems

Do the following programming problems. You will end up with at least one code file per problem. Submit your program source (and any other needed files) by sending mail to bmassing@cs.trinity.edu with each file as an attachment. Please use a subject line that mentions the course and the assignment (e.g., ``csci 3323 hw 5'' or ``O/S hw 5''). You can develop your programs on any system that provides the needed functionality, but I will test them on one of the department's Linux machines, so you should probably make sure they work in that environment before turning them in.

1. (10 points) Write a program or programs to demonstrate the phenomenon described in problem 2. Turn in your program(s) and output showing differences in execution time. (It's probably simplest to just put this output in a text file and send that together with your source code file(s).) Try to do this in a way that shows a non-trivial difference in execution time (so you will likely need to make the arrays or other data structures large). I strongly recommend that you write your programs in C or C++, or some other language where timing results are more predictable than they're apt to be in, for example, a JVM-based language such as Java or Scala (because ``just-in-time'' compilation makes attempts to collect meaningful performance data difficult). But anything that can be compiled and executed on one of the Linux lab machines is acceptable, as long as you tell me how to compile and execute what you turn in, if it's not C or C++. You don't have to develop and run your programs on one of the lab machines, but if you don't, (1) tell me what system you used instead, and (2) be sure your programs at least compile and run on one of the lab machines, even if they don't necessarily give the same timing results as on the system you used.

   Possibly-helpful hints:

   • An easy way to measure how long program mypgm takes on a Linux system is to run it by typing time mypgm. Another way is to run it with /usr/bin/time mypgm. (This gives more/different information -- try it.) If you'd rather put something in the program itself to collect and print timing information, for C/C++ programs you could use the function in timer.h to obtain starting and ending times for the section of the code you want to time.
   • Your program doesn't have to use a 2D array (you might be able to think of some other data structure that produces the same result). If you do use a 2D array, though, keep in mind the following:
     • To the best of my knowledge, most C and C++ implementations allocate local variables on ``the stack'', which may be limited in size. Dynamically allocated variables (i.e., those allocated with malloc or new) aren't subject to this limit.
     • Dynamic allocation of 2D arrays in C is full of pitfalls. It may be easier to just allocate a 1D array and fake accessing it as a 2D array (e.g., the element in x[i][j], if x is a 2D array, is at offset i*ncols+j).

2. (15 points) The starting point for this problem is a C++ program that simulates execution of a page replacement algorithm. Currently the program simulates only the FIFO algorithm. Your mission is to make it simulate one or more of the other algorithms mentioned in the text (and listed in the main program). You will get full credit for simulating one algorithm, extra points for simulating additional algorithms. The starter code -- well, there's a lot of it, but my hope is that I've structured it and commented it in such a way that you will not find your job too daunting.

   The program gets input from a combination of command-line arguments and an input file, described below, and produces some statistics including how many page faults each algorithm generates. To compile the program, use any C++ compiler (I've only tested with g++ on our machines, but I don't think I've done anything that wouldn't work with other compilers); you may need the -std=c++11 flag. Command-line arguments:

   • Required:
     • name of input file (format below)
     • number of page frames
   • Optional:
     • ``-interval N'' to specify interval for ``clock ticks``, for algorithms that need this (almost but not quite all of them)
     • ``-tau N'' to specify time interval for working set algorithms
     • ``-debug'' to have program print extra information about what it's doing, potentially useful when debugging, or just to see details of its operation
   Input file format:
   • number of pages
   • one or more lines of the form ``t R n'' or ``t W n'', where t is the access time (in increasing order but not necessarily without gaps), R/W indicates whether this is a read or write reference, and n is the page number being referenced
   Output should be the following information, for each page replacement algorithm implemented:
   • name of algorithm
   • total number of page references
   • number of page references that changed the page ('W')
   • number of page faults
   • number of times a page had to be written out
   Make the following assumptions:
   • Initially memory is empty.
   • All memory references are valid -- if the page is not in memory, it can be read in from disk. (You don't have to simulate the actual reading-in.)
   Sample input and output:
   • Command-line parameters pagingsimulator.in 4 -interval 5 -tau 10
   • Input file
   • Output (This is output for my sample solution, which implements all the algorithms. As provided, the program only simulates FIFO and so only produces output for that algorithm.)
   Algorithm-specific notes:
   • Some of the algorithms (e.g., NRU) say they choose a page ``at random'' from some group of pages. Just pick the first one in the group.
   • For LRU I want you to do the version that involves additions to hardware (instruction counts in the PTE). Rather than trying to simulate actual instruction counts, just use time of last reference.
   • For Aging, use 16-bit counters.
   • WSClock is described in terms of ``scheduling I/O'' to write out pages. Doing this in a realistic way is beyond the scope of this problem, so I want you to just fake it by (simulating) writing out the page immediately (but note that the algorithm skips these pages on its first trip around the circle of pages, only considering them if it doesn't find a page that didn't need to be written out).
   If there are other details you find unclear from the textbook's description, please feel free to ask! When I started writing my sample solution I found that there were some details that were not spelled out in the textbook as clearly as one might like.

   To get started, get a copy of this ZIP file containing the starter code, unzip it (command unzip on our machines), and try compiling the main program (main.cpp) and running it (you might try it with -debug too). Then start looking at code, which is structured as follows.

   • Files you should not need to change:
     • Main program main.cpp This program makes an object for each algorithm (classes below) and calls its simulate() method. These methods produce Results objects with field isValid that should be true for the algorithms actually implemented; the main program prints all such objects where this value is true.
     • Class for page references (inputs) page-reference.hpp (you may not need to look at this)
     • Base class for page replacement algorithsm pra-Base.hpp (you may not need to look at this)
     • Class for page table entries pte.hpp
     • Class for results of simulation results.hpp
     • Class representing the FIFO algorithm pra-FIFO.cpp
   • There are also ``stub'' versions for classes representing the other algorithms:
     • pra-Aging.cpp
     • pra-Clock.cpp
     • pra-LRU.cpp
     • pra-NFU.cpp
     • pra-NRU.cpp
     • pra-Optimal.cpp
     • pra-SecondChance.cpp
     • pra-WSClock.cpp
     • pra-WorkingSet.cpp
     You should pick one or more of these and complete it to implement the desired algorithm. Comments including the word FIXME show where you need to make changes/additions. (You probably don't need to make other changes.) The code for FIFO is meant to serve as a model, and comments with the word HINT are meant as hints about what parts you can probably copy as is and what parts you may need to adapt.

Footnotes

... apply).¹

Credit where credit is due: I based the wording of this list on a posting to a SIGCSE mailing list. SIGCSE is the ACM's Special Interest Group on CS Education.