CS 252 - Homework #2

Homework 2 is due Monday, September 30 at 5 PM. Turn it in in the appropriate box in 283 Soda. Note that the building is locked (for those who do not have keycard access at 6:45 PM during the week, as well as on the weekends.

Homework assignments should be done in pairs. In case this was not made clear enough last time, this is not a suggestion, this is a requirement. Each pair is to do their own work, separate from the other pairs. Each pair turns in one solution. If there are still an odd number of people in the class, one group may be a group of three (or one person may work alone). Whoever asks first will be granted that privilege.

• 5.4

• 5.2 - Instead of having to find a machine and run the code yourself, you will be given a selection of graphs representing similar runs on a variety of machines. (These graphs can be found outside of Dave's door.) Pick three of these graphs, one from each group, and, for each, answer as many of the questions in parts (b) through (k) as it is possible to deduce from the graph.

• This problem involves investigating the performance of the memory hierarchy by running a cache simulator on traces from various benchmarks. (The traces contain the first 1,000,000 instructions from two of the benchmarks used by Dick Sites in his "Caches Don't Work" talk. Tomcatv is one of the SPEC benchmarks, and SQL95 is a trace from a database.)

  The traces can be found in the directory ~cs252/F96/hw2 which can be accessed from any Instructional account. If you do not yet have an Instructional account, you should get one now. Log into any Instructional machine (like po.eecs or cory.eecs) as the user newacct to get one.

  The traces have been gzip'd. Copy them to some local tmp area and use gunzip to expand them. (They occupy about 15 MB each in uncompressed form.)

  The cache simulator you are going to use is dineroIII. A DEC-MIPS executable can be found in ~cs252/F96/bin/mips. You can find out which Instructional machines are for which architecture by typing clients at while logged into any Instructional machine. Documentation for dinero can be found in ~cs252/F96/man. If you really want to run on some other machine, you can copy all of the source files from ~cs252/F96/dinero and run make on your local machine.

  If there are any problems with any of this (like permissions set wrong, or any mistakes in the above instructions) please let me know ASAP. Okay, enough formalities. Here's the question:

  Assume that you have a CPU with the following characteristics:

  • base (execution) CPI is 1.5
  • 200 MHz frequency
  • 16 KB Icache, 16 KB Dcache, 64 byte blocks, 1 cycle hit time
  • 100 ns miss penalty to go off chip
  • cache is write-thru, no-write allocate
  • assume a write buffer big enough to not stall on a write-miss
  • assume the CPU stalls a read-miss

  For each of the benchmarks (tomcatv and sql95), answer the following questions, with respect to the machine described above:

  • what is the AMAT?
  • what is the CPI?
  • what is the total execution time?

  The next section addresses the question of what happens to performance given technology trends. Assume that the following have changed:

  • CPU is 4 times faster
  • on-chip accesses are still 1 cycle
  • memory only gets 20% faster
  • caches have 4 times the capacity

  What are the answers to the same questions as above, given this new machine? For each benchmark, what is the overall speedup?

  Assume all accesses must be in whole numbers of CPU cycles. Round up any fractional memory access values. Include the outputs from running dinero in the assigment that you turn in.

Back to CS252 home page...