I was asked to answer the following question as a discussion prompt for my computer architecture course:
Do you think that you could write a program that could determine whether a computer is using instruction pipelining? Explain your answer.
I decided to try it out.
The Makefile defines how to compile the C and assembly programs as well as
some simple cleanup behavior.
The run_tests.sh file is used to orchestrate the tests. It runs some cleanup,
compiles all the programs, and runs a few tests in different orders to help
reduce the impact of OS process scheduling.
The timer.c file is designed to time the execution of a shell command. It
is only accurate on the scale of microseconds. I couldn't find a way to do
nanosecond timing on OS X with the amount of effort I wanted to put in to this.
So, we have microsecond timing. The results of the timer are dumped to .dat
files by the run_tests script.
program1.s and program2.s are assembly programs with and without pipelining
hazards respectively. They execute the same exact instructions, but in a
different order. They each loop over these instructions 25 million times. So,
they take a while to run. This is necessary.
It turns out, we need to loop many many times in the assembly code for two reasons:
- The operating system's prioritization of program execution causes a lot of deviation in execution time.
- We only have microsecond timing and the difference in execution time of each loop iteration between program 1 and program 2 is very small. We need to run many times to measure the differences on the scale of microseconds.
Finally, a Python script results.py reads the timing results and plots them.
Then you get a result looking something like this:
First, I ran this on Mac OS X. I have not tested it on any other system.
To assemble the assembly programs, you need nasm installed.
To compile the C code, you need gcc.
The Python script depends on matplotlib and
NumPy. If you already have
pip and Python installed, just run
pip install -r requirements.txt.
Once you've gotten all the dependencies installed (and made any necessary
cross-platform changes), just run ./run_tests.sh.
Loop 25,000,000 times
Loop 1,000,000 times
Loop 100,000 times
Loop 10,000 times
