jan_here | hello? | 13:10 |
---|---|---|
Keybuk | StyXman: so the problem is to create an efficient readahead list | 19:06 |
Keybuk | we've got an inefficient one today | 19:06 |
Keybuk | replacing that with a differently ineffeicient one doesn't really help things in the long-term | 19:06 |
Keybuk | ideally we'd like to know, on average, which executables request which blocks of which files | 19:06 |
Keybuk | and at what time they do it | 19:06 |
Keybuk | then we can page them into the cache in order, in advance of when they're needed | 19:06 |
Mirv | are you planning on directly tackling some application problems, similar to this extremely great work done by Behdad: http://mces.blogspot.com/2008/12/improving-login-time-part-3.html ? | 19:07 |
StyXman | isn't too much low level? I mean, normally booting programs just read a file, maybe seq'ly, maybe randomly, and all you need is to have it already in ram, or do I miss something? | 19:08 |
Keybuk | Mirv: eventually | 19:08 |
Keybuk | StyXman: the page cache works on blocks, not files | 19:08 |
Mirv | yeah, I know, there is just so many even more low-hanging fruits, while the upstreams might actually fix something too | 19:08 |
StyXman | Keybuk: again, why so lo level? why not just read the file and let the kernel take care of page cache? | 19:09 |
Keybuk | StyXman: that's what we do today | 19:09 |
StyXman | if you're low in ram, you can make slices | 19:09 |
StyXman | Keybuk: ok | 19:09 |
Mirv | hmm, I could put a mental note somewhere to check whether compiz git fixes the 5s delay (no disk activity) I've seen, it'd good for jaunty if 0.8.0 does not come out | 19:09 |
Keybuk | but we read too much as a result | 19:09 |
Keybuk | in practice, when you execute a binary, you don't read all of it into memory | 19:09 |
Keybuk | you just read certain bits | 19:09 |
StyXman | Keybuk: how is the read/size ratio? | 19:10 |
StyXman | I mean, how much do you read? | 19:10 |
Keybuk | StyXman: I don't have numbers on that yet | 19:11 |
StyXman | Keybuk: ok | 19:11 |
StyXman | Keybuk: do you know if what readahead does (being so low level) is a consequence of already trying that? how was the approach selected? | 19:12 |
Keybuk | as you can probably tell from the graph changes, my focus atm is on the kernel, udev and modprobe load times | 19:12 |
Keybuk | StyXman: readahead() is just a syscall | 19:12 |
Keybuk | man readahead gives you the docs on it | 19:12 |
StyXman | Keybuk: ah, ok | 19:14 |
Generated by irclog2html.py 2.7 by Marius Gedminas - find it at mg.pov.lt!