Talk:CSC 456 Fall 2013/1c wa

Round 3 comments

Please capitalize the L in L1, L2, etc.

Can you cite a source for your theory about cache associativity? It does seem reasonable, although I would expect working-set size to play a role.

There's still a "Systems to consider in table", so it appears that this section is not done.

In your table, there's a long gap between 1968 and 1989. Influential machines of this time would include the IBM S/370 (in a multitude of models), the DEC 10 & DEC 20, and the DEC PDP-11 (also lots of models). As well as the DEC VAX/780, which debuted in 1978 and came in many models. Amdahl was a major competitor to IBM during this period; you might look for "Amdahl cache parameters." It might be worth noting that machines designed by Seymour Cray (CDC 6400/6500/6600/6700 and Cray 1, Cray 2, etc.) between 1964 and about 1990 did not use caches at all, opting instead for large numbers of programmable registers, which could be managed by the compiler. (Some of these are in your next table.)

What was the 1st Intel chip to include a cache? Was it the 486, or an earlier model?

You really should include narration talking about what the tables show.

Round 2 comments

You have made a good start to talking abut the reasons for increase and decrease of associativity. Can you explain further why it was more expensive to search larger set-associative caches? Consider this: If a cache were as large as main memory, there would be no conflict misses, and thus no reason for associativity. Consider what happens when the cache is really small. Does that explain why larger size is often associated with less associativity?

But, trends have changed over time. So evidently it is not as simple as the analysis in the previous paragraph would suggest. Why might it not be? Can you find literature that explains this?

Can you expand your table to cover L3 caches? Or, would you want to cover them in another table?

Round 1 comments

Here are some suggestions for improving your page.

• My first suggestion is to read the Wikipedia article on CPU caches (http://en.wikipedia.org/wiki/CPU_cache) and look up the references at the end. You are not expected to duplicate this article, but it does offer several insights into how caches have changed over the years.

• To look for trends, it is helpful to search in dbs based on years. Look in specific dbs for how to do this. For example, in Google scholar, you can do it with "Advanced search", which is an option from the dropdown at the very right edge of the page.

• Using this approach, I found a paper from 1992: http://dl.acm.org/citation.cfm?id=145814. It talks about _page coloring_ as a technique to allow lower-latency searching of caches.

• A plain Google search for '"cache associativity" trends' seems to return several useful results.