Closed Bug 68802 Opened 25 years ago Closed 25 years ago

win32 function orderer should accumulate call information

Categories

(SeaMonkey :: Build Config, defect)

x86
Windows 2000
defect
Not set
normal

Tracking

(Not tracked)

VERIFIED DUPLICATE of bug 65845
mozilla0.9.2

People

(Reporter: waterson, Assigned: rogc)

Details

(Keywords: memory-footprint)

jband says, As I originally imagined this, we would be accumulating this information on various runs over time. This requires the ability to read the order files, merge in info from the current run, and then write out the new file. You are planning to support this, right? Right. dprice, make it so!
milestone = .9
Target Milestone: --- → mozilla0.9
-> 0.9.1
Target Milestone: mozilla0.9 → mozilla0.9.1
I'm looking into this now... -Roger
Assignee: dprice → rogc
Here a couple of tidbits I've learned so far: - In order for the /ORDER linker flag to work, everything needs to be compiled /Gy. I don't think we're currently doing this. - It also looks like there's some major suckage in that reordering doesn't work on static functions. I have the feeling this is going to be a big problem, and the obvious solution (don't use static functions) is almost certainly going to be impractical. Also, ordering functions based on frequency of execution is not going to yield desired results. There has been research on this, and I remember going through some of it on a project I worked on a long time ago using cord. I'll see what I can dig up; so far I've found some references that have me thinking along these lines: Build a graph where each node represents a function, and whenever a function is called an arc is added between the node for that function and the node for the function that was last called. After execution, form an ordering by starting with the two nodes that have the most arcs between them. The two corresponding functions go first. Then collapse those two nodes and repeat. This roughly corresponds to what I remember from working with pixie and cord on IRIX, but I'll do more digging to make sure. Also, I think we'll want to do some special-casing for startup; we'll want to put all the functions that only get called once during startup together so they can all get paged in together at startup and then get paged out together later. -Roger
rogc: From your comments I wonder if you've read bug 65845 and discussed these issues with dprice or waterson? Much has been discussed before. *This* bug is really just about the added feature of allowing those whoe chhose to participate to accumulate the info accross multiple runs. I'm thinking now that this is acutally of dubious value. I've been pushing the idea that the best strategy for this whole process is not to try to get a best ordering of called functions, but rather to segregate the non-called functions. It seems to me that to do so would be the bigest win for reducing the count of physical pages swapped in. I could be wrong. If you are going to work on this general issue, *I'd* suggest you work on the questions at the bottom of that other bug: How to best measure the effectiveness of the ordering scheme we already have. And, why it does not seem to work as well as expected and how to make it work better.
jband: Thanks. I'm now much more informed. I'll work with dprice to figure out who is going to do what here. :-) I completely agree with you about measurement being the first step! Segregating non-called functions from called functions certainly seems like it would be better than doing nothing at all. Based on past experience, however, I think that going further could have even more benefit. Since doing the called/non-called segregation is less work than the graph stuff I was writing about, it makes perfect sense to me to measure it both ways to see if the more complicated approach really has any benefit.
Moving milestone out. I'm still in the research phase.
Target Milestone: mozilla0.9.1 → mozilla0.9.2
The stuff I'm doing for 65845 covers this one too. *** This bug has been marked as a duplicate of 65845 ***
Status: NEW → RESOLVED
Closed: 25 years ago
Resolution: --- → DUPLICATE
verified.
Status: RESOLVED → VERIFIED
Product: Browser → Seamonkey
You need to log in before you can comment on or make changes to this bug.