1 |
Roy Marples <uberlord@g.o> posted |
2 |
200611071058.54068.uberlord@g.o, excerpted below, on Tue, 07 Nov |
3 |
2006 10:58:53 +0000: |
4 |
|
5 |
> Actually, before I do that, let me attack this from another angle. |
6 |
> What do you gain from keeping it mounted as a ramdisk? |
7 |
> If the answer is performance, well you loose performance at start time as |
8 |
> you've lost the deptree. |
9 |
> |
10 |
> So why would you want to keep it? |
11 |
|
12 |
Well, as you point out it's mixed from a performance angle. One reason |
13 |
I'm beginning to appreciate a bit more now that I have a tmpfs mounted |
14 |
/tmp as well, however, is the clean start at every boot thing. |
15 |
Eliminating even the possibility of trouble from stale or corrupted |
16 |
carryover has its own >0 benefits. |
17 |
|
18 |
Back to performance, however, the deptree rebuild isn't as much of a |
19 |
pull-down as one might expect, for a number of reasons. Keep in mind that |
20 |
perhaps the biggest boot-time bottleneck is disk read bandwidth anyway. |
21 |
Preloading all those scripts effectively at once therefore has three |
22 |
overlapping effects. 1) There's the direct disk read bandwidth, but that |
23 |
would come in somewhere anyway when the scripts are actually executed. 2) |
24 |
Having them all loaded at once gives the filesystem elevator scheduler a |
25 |
bigger chance to do its thing, ordering the reads appropriately for better |
26 |
efficency. 3) Once it comes time to actually run the scripts, they are |
27 |
already loaded into cache, so no waiting to read them in then, simply |
28 |
virtually instant execution. |
29 |
|
30 |
Thus, while the performance benefit won't be great, it's likely to be |
31 |
there, and there's the "system hygiene" (I believe that's a decently |
32 |
accurate label) benefit as well. |
33 |
|
34 |
So, keeping it at minimum an option is IMO a good thing. If as proposed |
35 |
we make it standard, there are other benefits including serious code |
36 |
simplification as well. There could well be negatives, especially on |
37 |
certain archs I'm not familiar with, but for x86, amd64, and probably the |
38 |
ppc folks, at least, it sounds very reasonable from my perspective. From |
39 |
what I've read, the particular challenges faced on MIPS may change the |
40 |
time effects greatly there (basing this on remarks I recall from the |
41 |
split-KDE debate, but I haven't the foggiest whether that applies here |
42 |
or not), so I'd be interested in seeing someone address it from their |
43 |
perspective. |
44 |
|
45 |
-- |
46 |
Duncan - List replies preferred. No HTML msgs. |
47 |
"Every nonfree program has a lord, a master -- |
48 |
and if you use the program, he is your master." Richard Stallman |
49 |
|
50 |
-- |
51 |
gentoo-dev@g.o mailing list |