1 |
Volker Armin Hemmann wrote: |
2 |
> On Mittwoch, 13. Februar 2008, Duncan wrote: |
3 |
>>> removed lots of irrelevant 'my hardware is so cool' stuff'. |
4 |
> |
5 |
> You forget some (little) things. Not everything can be swapped out. |
6 |
|
7 |
How is this relevant? Some memory is locked or effectively locked. |
8 |
That limits buffers/cache in the disk-based fs case, and swap in the |
9 |
tmpfs case. I don't see that there is much of a difference. |
10 |
|
11 |
> Swap is |
12 |
> extremly slow AND it is much worse to swapout/swapin programm code that |
13 |
> should be run, instead of fetching some files from disk while the programm |
14 |
> runs. |
15 |
|
16 |
You keep asserting this. Why should fetching a page from swap be slower |
17 |
than fetching the equivalent space from a file? In fact, if mmap is |
18 |
used to read a file the same code probably is used for both. |
19 |
|
20 |
And the kernel is not likely to swap out recently-run program pages. It |
21 |
isn't going to keep a bunch of compiled binaries in memory for 10 |
22 |
minutes while it swaps in and out the compiler - it would swap in and |
23 |
out source files and binary files - just as with a disk-based |
24 |
filesystem. If for some reason part of the compiler is swapped out, it |
25 |
would be rewritten to disk every time it got swapped out - only the |
26 |
first time. I'm sure the kernel would notice that the pages were never |
27 |
written since the last swap and would just flush them from memory |
28 |
without another write. However, this should still only happen on idle |
29 |
pages, and swapping those out is probably better than any other use of |
30 |
memory (and those pages would get swapped out even with a disk-based |
31 |
filesystem in use). |
32 |
|
33 |
I'm of course interested in benchmarks. In theory swap should be faster |
34 |
than files. However, implementation does matter and maybe there is some |
35 |
flaw that makes things slower. My assertion is that a swap-based |
36 |
solution gives the kernel more freedom to optimize memory use than a |
37 |
disk-based solution. So, this should be a fundamentally-superior solution. |
38 |
-- |
39 |
gentoo-amd64@l.g.o mailing list |