Talk:CSC 456 Fall 2013/1b ra

From Expertiza_Wiki
Revision as of 04:09, 1 October 2013 by Admin (talk | contribs)
(diff) ← Older revision | Latest revision (diff) | Newer revision → (diff)
Jump to navigation Jump to search

Round 3 comments

You've done a pretty extensive job of covering these issues. This is good. I have a few suggestions on how to improve.

Can you find how large the sectors, lines, and tags were on the 360/85?

Can you find what "81% of ideal efficiency" means in terms of CPI?

I found some other good info from looking in the citations to Rothman's 2000 paper, "Sector cache design and performance." See http://scholar.google.com.prox.lib.ncsu.edu/scholar?cites=7200973186802245143&as_sdt=5,34&sciodt=0,34&hl=en There's even a paper on combining sectored & victim caches.

Citations should come after punctuation, except for semicolons, colons, and dashes (this is done correctly in the sectored-cache section, but not in the victim-cache section).

It's not clear to me that a system using a victim cache needs to be direct mapped. If it's, say, 2-way associative, the victim cache can store the 3rd most recently referenced line in a set.

In the selective victim-caching paper, you refer to 1997 twice.

I fixed some punctuation problems, but there are others, including lack of space before open parens, and "it's" for "its".

The paper on victim replication--is it implemented in a real architecture? In any case, I'd like to see you give the parameters of the victim cache in real architectures (how many lines, etc.).


Round 2 comments

I suggest you say, "a sector cache" or "a sectored cache"; that's more usual usage.

It would be useful to explain why set-associative caches were found to be more efficient.

In early sectored caches, how large were the lines and the sectors?

The organization is mostly good, but it's possible that it would be more efficient to combine the advantages/disadvantages section with Background.

In victim caches, I would suggest talking about the size of caches in the Background section, instead of in the Handling Misses section. It would also be good to tell how the size has changed over time, e.g., how large are victim caches today, and how does that compare with their size in the '90s?

Can you find anything about victim caches in Intel architectures?


Round 1 comments

Here are some suggestions on how to improve your page.

  • For sectored caches, it is also important to talk about recent designs. I tried a Google Scholar search limited to 2000–2013, and came up with many references, including several patents. I wouldn't spend a lot of time looking through patents, but it _is_ helpful to look at them, since they represent work that is more applied than research papers. Relatively few research papers ever result in patents, but any ideas from research papers that are adopted in real architectures are likely to have patents filed on them.
  • There is quite a bit of recent work on victim caches. The top hit from the past year is on optimizing power used by victim caches. This would be a good issue to explore.
  • You could go into a lot of detail about specific organizations for sectored or victim caches. In general, this is not a good idea. The first step should be to decide which designs are really important, e.g., which have been implemented in more than one architecture, and then describe those. There is way too much research to describe in a single page (and hardly anyone, other than researchers, would be interested in all of it anyway).