next up previous
Next: Nominal Hardware Costs Up: No Title Previous: No Title

Format of your submission

Your submission for this part of the assignment should consist of a report, between 4 and 10 pages in length, which tabulates your chosen parameters. You are advised to include, in tabular form, any data you may have generated and/or assumed in order to determine these choices. Note that the LATEX source for this document is available, for cutting and pasting of tables if required, in the file

    ~cs3/comparch/Coursework/doc/partII.tex

Marks will be awarded for submissions which:

1.
State clearly the design choices that have been made;
2.
Clearly indicate their cost, and show that it falls within the range of allowable costs as given in table 1;
3.
Demonstrate that the effective memory access time is as small as possible, given the constraints imposed by table 1.
4.
Show that both cost and effective access time have been optimised.

To satisfy the first point you simply need to include a table of all the cache and memory parameters, indicating those which you have determined yourself. You should also state clearly any further assumptions you have had to make in order to complete the task.

The second point will require you to compute the cost of your design, using the cost relationships outlined in the next section. If you over-spend or under-spend on the permitted budget, you must explain why.

To satisfy the third point requires that you compute the effective memory access time of your design, and show that any deviation from your design choices will either:-

1.
Lead to excessive cost, or
2.
Result in an increased effective memory access time.
Typically, this will require a tabulation of the sensitivity of your design to changes in each of your specified parameters. For example, suppose there exist parameters p1, p2 and p3, with chosen values v1, v2 and v3. We can perform a sensitivity analysis on parameter pi by computing the performance of the system when all parameters other than pi take on their nominal values, whilst pi takes on values $v_i+\delta$ and $v_i-\delta$. If the resulting performance in both of these cases is worse than the performance with pi = vi then it can normally be deduced that vi is optimal for pi. Of course, this is not always true, but the alternative is an exhaustive search of the design space and that is clearly not a viable design methodology.

In the case of this assignment, you should check the sensitivity to the degree of associativity, the cache capacity, cache policies and the block size. You may find that the total number of possible configurations, given the available degrees of freedom, is not too large, and that a simulation of each of them is possible. Alternatively, you may choose to be guided by the course text book, and use some of the information tabulated therein to determine your initial parameter values[*]. Even if you do this a set of sensitivity analyses will still be required.

The fourth point requires that having shown you have the best effective access time (given the constraints) that such an access time could not be obtained at lower cost.


next up previous
Next: Nominal Hardware Costs Up: No Title Previous: No Title
Nigel Topham
6/25/1998