CS 382M: Advanced Computer Architecture

Homework 3

Due: MONDAY Mar 8, 1999 2PM No late homeworks will be accepted

Work in groups of two. Each group should turn in one homework.

• Memory System Analysis.

Do exercise 5.2 from the text. The measurement program is available here. Turn in a graph of your measurements and for each question, explain your answer and highlight the relevant portion of the graph.

• Cache Design.

Do exercise 5.7 from the text.

• Dinero Experiments.

This question involves a uniprocess cache simulator called dinero. An executable version of this simulator is avaliable in /public/solaris/bin on Computer Science SUN machines. You will run the simulator on ATUM traces of several programs available in /u/dahlin/public_html/Classes/GradArch/homework/hw3/traces/atum/. As an example, to run dinero on trace file mul8.003.din.Z, use:

zcat /u/dahlin/public_html/Classes/GradArch/homework/hw3/traces/atum/mul8.003.din.Z | /public/solaris/bin/dinero -options > outfile

where -options is replaces by the specific options you need and outfile is the name of you ouput file. Since the trace files are very large, you do not need to uncompress or copy them. For more information, see dinero's man page.

Run dinero on trace files pasc.000.din.Z, lisp.000.din.Z, and fsxzz.000.din.Z to answer the following qestions. Use the dinero default for any cache options not mentioned in the question. Turn in your results and your conclusions.

• Replacement policy.

Examine a 8K-byte, 4-way associative cache with 16-byte blocks. What are the average miss ratios for LRU, FIFO, and random replacement policies? What conclusions can you draw from this experiment?

• Unified versus separate instruction and data caches.

Compare the cache miss ratios of the following two systems:

• a system with a 16K-byte unified cache
• a system with a 8K-byte instruction-only cache and a 8K-byte data-only cache.

Assue the caches are 2-way set associative and the block size is 16 bytes. What conclusions can you draw from this experiment?

• Miss ratio variation.

Examine a direct-mapped 16K-byte unified cache with 16-byte blocks. What are the miss ratios for the first 50,000 references (reference 1-50,000), the second 50,000 (50,001-100,000), the third 50,000 (100,001-150,000), and the fourth 50,000 (150,001-200,000)? What is the average miss ratio for references 1-200,000? What conclusions can you draw from this experiment? Is running portions of traces a good indication of the overall behavior of a program?

• Design your own experiment.

Design an experiment to study some aspect of caches not covered by the above experiments. Describe the experiment, the hypothesis (including the purpose of the experiment, what the results would look like if your hypothesis is false, and what the results would look like if your hypothesis is true), the actual results of running the experiments, and you conclusions. Give an resonable explanation if your results are different from your initial hypothesis.

• Paper critique

Read the two papers assigned for this class.

For each of the papers, turn in a short (1/2 page max critique of the paper answering the following three questions.

a. What are the three most important concepts in the paper?

b. What is the most glaring flaw in the paper?

c. What is the main conclusion you take away from this paper regarding the future of microprocessor architectures?

The main goal of this homework question is to encourage you to read the papers in advance of the class discussion so that you come to class prepared to be an active participant in the discussions. Our grading of the above questions will reflect that goal. Note that it is not appropriate to turn in an answer to this question if you have not made an honest effort to read and understand the papers.