There are two goals for this coursework:
-
Make sure you understand the working of a cache, including the specifics of read misses versus write misses, and the managment of valid and dirty flags.
-
Develop skills in working with input and output streams for programs, and methods for checking that files match a particular format.
Write a program which models the behaviour of an arbitrary single-level cache configuration, including both a cache, and the memory behind the cache. The program will receive the configuration as command line arguments, a stream of input memory transactions, and produce a stream of response transactions.
The program should start with a memory full of zeros, and an empty cache. Input transactions should be processed until end-of-file on stdin, which will result in the program exiting.
Input parameters (in order of appearance on the command line) are:
-
Address bits
-
Bytes/Word
-
Words/Block
-
Blocks/Set
-
Sets/Cache
-
Hit time (cycles/Access)
-
Memory read time in cycles/Block
-
Memory write time in cycles/Block
All parameters are positive integers, encoded as decimal. Parameters 2-5 will be integer powers of 2 (e.g. 1, 2, 4, ...) Total memory space size is at most 8MB.
The cache should operate a LRU replacement policy, and a write-back write policy.
The input stream on stdin will consist of three types of
transaction, with the following patterns:
-
Read request: "read-req" address
-
Write request: "write-req" address data
-
Flush request: "flush-req"
-
Debug request: "debug-req"
-
Comments: "#" followed by any characters.
A flush request ensures that memory is consistent with current cache contents (recalling that the cache is a write-back cache). The end of the input stream will be indicated with end-of-file.
A debug request results in user-defined debug output, and can be used (for example) to print the current state of the cache.
Comments do not produce any response, and should be discarded.
The maximum length of any line in the input string is 1020 characters.
In response, your program will produce a stream of
responses on stdout, describing the result of the transaction:
-
Read response (a): "read-ack" data set-index "hit"|"miss" time
-
Read response (b): "read-ack" set-index "hit"|"miss" time data
-
Write response: "write-ack" set-index "hit"|"miss" time
-
Flush response: "flush-ack" time
-
Debug response: "debug-ack-begin"\n\n"debug-ack-end"
-
Comments: "#" followed by any characters.
(The two formats for read-ack are to reflect a historical
lack of preciseness in the spec. A program should choose
one output form to use, and it is up to consumers to
work out which one is in effect. So this is an example
of de-facto specification, where the existance of mutually
incompatible standards in the wild pushes complexity
onto the user.)
set-index is the index of the set in which the cache
places the address.
The time associated with each response is the total
time (in cycles) needed by the memory system to service each request.
For read and write requests, all cache processing is included
within the hit time, and the only extra time needed is the
time taken to perform any reads and/or writes to memory. Flushes
do not include the hit time, and only consume the time taken to
write dirty pages to memory. In each case, the time should be the
minimum possible, so solutions which by-pass the cache for every
read and write are not correct.
A debug response will contain zero or more lines, sandwiched between a line containing "debug-ack-begin" and "debug-ack-end". The response can be empty, i.e.:
debug-ack-begin
debug-ack-end
or you may wish to include logic that prints the current state of the cache for debugging purposes. Either way, the command must be supported.
Comments can be produced at any time (or not at all), and will be discarded.
All addresses are byte addresses, and will be word aligned. Both addresses and set indices are decimal integers
Data is a string of hex digits with upper-case letters, which must be of the same width as a cache word (i.e. two hex digits per byte). The ordering is from highest address to lowest. For example, the 64-bit data string
FEDCBA9876543210
contains 0xFE at byte offset 7, and 0x10 at byte offset zero. Note that there is no notion of endian-ness here, as we are not inside the CPU.
Your submission should be a zip file containing:
-
A C++ source file called
mem_sim.cppwhich implements the simulator. -
(Optional) Any ancilliary C++ sources called
mem_sim_*.cppwhich should be compiled into the program. -
A file called
readme.txtwhich briefly (i.e. no more than a few lines) explains how you tested and debugged your simulator. -
Any supporting files used during testing, such as test input and output, and/or scripts used. These are not part of the assessed work, they are there to explain how you arrived at the solution (e.g. if external examiners are interested, or marks need to be justified).
The submission should be submitted via blackboard.
The marks weighting is broken down as follows:
-
20% Specification: Are the files called the right thing, do they compile?
-
20% Input/output: Does the simulator parse all valid input files without crashing, and is the output always in the valid format (regardless of whether it is correct).
-
60% Correctness: Does the simulator produce the right output? This will be assessed on a number of inputs, from simple direct mapped (including the example input) to more complicated.
As before, there is a formative deadline, where I try your code on a sub-set of tests, and try to point out if it is not following the spec in an obvious way, before the final hard deadline.
Formative deadline: Thursday 4th 23:59.
Hard deadline: Thursday 11th 23:59.
The files direct-mapped.input and direct-mapped.output give
example input and output files for a cache with:
- Address bits = 8
- Bytes/Word = 2
- Words/Block = 2
- Blocks/Set = 1
- Sets/Cache = 2
- Cycles/Hit = 1
- Cycles/ReadBlock = 2
- Cycles/WriteBlock = 2
To execute your program with this input, you would do:
cat direct-mapped.input | ./mem_sim 8 2 2 1 2 1 2 2
in a unix-style command line, or:
type direct-mapped.input | mem_sim 8 2 2 1 2 1 2 2
in a windows command prompt. In both cases the output
should be printed to stdout, and could be captured to
the file direct-mapped.got by redirecting:
cat direct-mapped.input | ./mem_sim 8 2 2 1 2 1 2 2 > direct-mapped.got
or:
type direct-mapped.input | mem_sim 8 2 2 1 2 1 2 2 > direct-mapped.got
Note that the method you use to get input and output to your program depends on the OS you choose. As long as your program is C or C++, takes its parameters as command line arguments, its input from stdin, and sends output to stdout, then it will work on any platform/OS (which is all the spec cares about).
If there is only one word per block, then write misses can be optimised to be slightly faster than write misses on multi-word blocks.
If you associate each valid (i.e. non empty) block in the cache with the read or write request it was last (most recently) accessed in, you will find that each block must be associated with a unique request. If you assign each read/write request an increasing index, then the block with the smallest index is the least recently used.
Alternatively, if you consider the blocks in a set as a list, then another approach is to keep moving the block being accessed to the front of the list, while maintaining the relative order of all other blocks. So if the current order is [B0,B3,B2,B1], then B0 was MRU, and B1 was LRU. If block B2 is now accessed, the order changes to [B2,B0,B3,B1]. Note that you don't have to move the blocks themselves around, you just need a list of block indices.
Creating input by hand works quite well for small sizes, e.g. using a spreadsheet to manually track what is in each block. You don't need to test every combination of input parameters, only to exercise a few interesting combinations. For example direct-mapped, 4-way, and fully associative combined with a few different word and block sizes should suffice.
While it's a bit boring, it shouldn't take more than 5 minutes to generate both input and output, assuming you have a good understanding of cache operation. If you don't, then that probably shows you need to spend more time with the book (which is rather the point of the exercise).
You can read from stdin using fgets, scanf or std::cin.
Similarly, printf and std::cout will write to stdout.
I have specified a maximum input length size to make your life
easier if you want to work with fixed-length buffers using
fgets. If you use a 1024 byte input buffer, it will always be
large enough.
Try not to use use gets, as you will make the baby
Jesus cry:
http://stackoverflow.com/questions/3302255/c-scanf-vs-gets-vs-fgets
However, I won't mark you down or anything.
This is a very restricted form of parsing, as you know there are only four possible input types, each one is solely contained within one line, and the structure of each line is precisely fixed. There are simple ways of doing this form of parsing, in either C style or C++ style.
In C++ style, just use the >> operator to bring
in the different parts. First read a string, and
it will be one of read-req, write-req etc. That
then tells you whether you need to read an address
(i.e. an integer), and possibly also a string of
hex characters. So something like:
while(1){
std::string cmd;
std::cin>>cmd;
// Handle end of file here
if(cmd=="read-req"){
unsigned addr;
std::cin>>addr;
// Now handle a read at address "addr"
}else if(cmd=="write-req"){
...
}
You can mimic the C++ style, using fscanf to
read each part of the input, first using it to
read a string and comparing it with read-req etc.,
then conditionally reading more arguments.
However, a more C-like way is to use fgets to read the
next input line into a character array, then use
sscanf to extract the different parts. If you
imagine the person producing the requests using
printf strings, then you can use sscanf with the
same string to parse it.
while(1){
char buffer[1024]={0};
fgets(buffer, sizeof(buffer), stdin);
// Handle end of file
unsigned address;
if(sscanf(buffer, "read-req %u", &address)){
// Handle a read request at "address"
}else{
... // other input patterns
}
For this exercise you don't need to worry about invalid input. Ideally you would fail gracefully with an error message highlighting the problem, but it is entirely acceptable to crash, start producing gibberish, whatever. The only thing you are not allowed to do is anything deliberately malicious.
There's a unix tool for that called diff, which will print
differences between two files.
There are a few tools which can do this, ranging from one liners in scripting languages like python and perl, to tool such as awk and sed: http://stackoverflow.com/a/5413132
One example is the sed command:
sed -r -e "/#.*/d" -
Which will take input on stdin, and write it on stdout without any comments.
A useful thing to do is to line up the input, your reference output,
and the actual output from your program. Unsurprisingly, there
are many ways of doing that. One tool which will do it is pr,
so if you have (un-commented) files test.input, test.output, and
test.got, you could do:
pr -m -t test.input test.output test.got
to get them printed out side by side.
Well, you could put your testing commands in a script, which runs your program, does any comment stripping, prints the columns side by side, and also diffs the output. Something like:
#!/bin/bash
INPUT=test.input;
OUTPUT=test.output;
GOT=test.got;
# Run your program, write the output to file name in GOT
cat ${INPUT} | ./mem_sim 8 2 2 1 2 1 2 2 > ${GOT}
# Look at the differences (this assumes no comments)
diff ${OUTPUT} ${GOT}
You can stick pretty much any commands in a script, and then execute them in one go (including compilation of the program if you want). The main steps are:
-
Create a text file (e.g.
shell.sh) -
Make sure the text file starts with
#!/bin/bash(or the name of some other shell), then put other comands on a following line. -
Make sure the text file is marked as executable, using
chmod u+x shell.sh. -
Run it, remembering to use a leading
./if you are in the same directory (e.g../script.sh).
Well, don't. Just type it in.
Hopefully no-one on EIE would actually ask that.
Well, you could install cygwin (or mingw if you are more hard-core), and all the tools will then be available.
Alternatively you could implement all of it yourself, writing a program that strips comments, prints them side by side, and highlights differences. It requires about 10 lines of C.
Nothing. You can manually run all the commands each time, or build a GUI in Visual Basic to do testing, or not do any testing at all, or whatever you feel like.
However, scripting is related to general computing, and these marks go into your Computer Lab module marks, so the idea is that you develop general computing skills while doing an exercise which develops your understanding of architecture. Working with files in a defined format is explicitly part of the exercise, doing so efficiently is only an implicit part.
Not really a question, but I'll answer it.
While it may look complicated (or it may not), this is actually much simpler than the MIPS exercise. There are a few fairly short steps:
-
Create a simple parser which can read the example input, split it into the arguments, and print it to the output. This takes about 10 lines of code.
-
Develop test inputs/outputs (do this second, to ensure they are in the right format). This should also check you know exactly how the cache should operate.
-
Define the cache and memory data-structures. This just follows the heirarchy of hardware in a cache, and takes very little time.
-
Map the input commands from your parser to operations on the cache.
-
Test and debug.
Of these steps, the most complicated are arguably 2, making sure you understand how a write-back LRU cache works, and 5, which is making sure your code works.
I originally didn't specify that byte addresses words would be word aligned, which was a failing on my part. In my mind it was obvious that they would be aligned, but that is because I had a particular mental model based on real buses.
A natural question people ask is "why are there address bits for the bytes, when you can only access words"?, which is quite reasonable. The main reason is that the address bus is not only for memory - as I showed in the final lecture, you can connect many things to one address bus, and some things like memory mapped IO may require individual byte transactions. But for the cache, you probably only want to work in words.