1 |
Volker Armin Hemmann posted on Tue, 04 Jan 2011 08:27:28 +0100 as |
2 |
excerpted: |
3 |
|
4 |
> swap kills interactivity and is really, really bad on linux. No matter |
5 |
> how much you stripe it. Swap is a true horror. So setting swappiness to |
6 |
> 100 (which means: keep caches alive, no matter what and swap the hell |
7 |
> out of it) is a really bad idea. In my experience it is better to have a |
8 |
> very low swappiness and let the kernel get the occasional data from |
9 |
> disk, than to swap to the same disks. Strange eh? |
10 |
> |
11 |
> But better to wait 1.5s longer for konqueror to display a directory than |
12 |
> to have a jerky mouse and input lag because of swap. Remember: the |
13 |
> kernel ALWAYS swaps out the wrong stuff. |
14 |
> I have swappiness at 60 - and back in the 4GB days at 0. Because swap |
15 |
> sucks so much. |
16 |
> |
17 |
> You are not required to believe that. But just google. |
18 |
|
19 |
All I can go on is my experience, which agrees with you (and most of the |
20 |
advice on the net) when it's a single-core CPU on single-spindle storage, |
21 |
but I've found the experience rather different on multi-core machines |
22 |
driving quad-spindle striped swap on mirrored RAID, with enough memory |
23 |
swap usage is trivial in the ordinary case. Once swap usage hits half a |
24 |
gig or so, yes, it's noticed, but until then, I literally don't normally |
25 |
notice it unless/until I happen to see the usage reported on the system |
26 |
monitors. |
27 |
|
28 |
OTOH, losing portage tree cache or news (nntp) article cache and having to |
29 |
fetch the data from disk can be quite noticeable, as it DEFINITELY is the |
30 |
first time I access after a reboot (the infamous cold-cache case), the |
31 |
biggest reason I tend to leave the system running for weeks at a time, |
32 |
rebooting only to load a new test kernel or the like. |
33 |
|
34 |
(FWIW, I use app-admin/lib_users to track programs using stale/already- |
35 |
deleted libs after an update, and after quitting kde/X, restart daemons, |
36 |
etc, to clear the list if necessary, thus clearing that source of both |
37 |
security vulns and so-called anon-memory usage. I do use the |
38 |
portage-2.2.0_alphas with preserve-libs, but use FEATURES=-preserve-libs |
39 |
to avoid that source of bugs, so the old libs do normally get deleted.) |
40 |
|
41 |
-- |
42 |
Duncan - List replies preferred. No HTML msgs. |
43 |
"Every nonfree program has a lord, a master -- |
44 |
and if you use the program, he is your master." Richard Stallman |