> Far be it from me to criticize ideas made by those who know
> much more than me, but personally I don't see how transfering stuff to swap
> immediately instead of grabbing what ram we can from the disk
> cache first could be considered faster or preferable. As far as
> I can see, it would be much better performance-wise to delay swapping as
> long as there's ram to be had from the disk cache. Isn't this the main
> reason the disk cache is dynamic in size?
When you just grab random disk cache memory, that piece
of data might be needed again, and has to be reread from
the (spread-out) filesystem.
Swapping, OTOH, is done in a clustered way (less seek-time)
on a small part of the disk (still less seek-time).
I would be interested though, in the partition layout of
the people who saw worse/better performance:
begin end
worse: data sys data swap
better: data swap sys data
???
Maybe something like this could explain the difference.
Personally, when I switched from swap at the far end of
the disk to swap in the middle (somewhere autumn '95) I
saw a _HUGE_ improvement.
Now, I tend to tune my stuff for swap-in-the-middle situations,
since I only have swap-in-the-middle as a test situation (and
who wants to run with slow swap anyway?).
If this explains the difference, I'll fix it by making
page-aging tunable for the following categories:
- user pages
- cache pages
- buffer pages
- shared pages (?)
Rik.
+-------------------------------------------+--------------------------+
| Linux: - LinuxHQ MM-patches page | Scouting webmaster |
| - kswapd ask-him & complain-to guy | Vries cubscout leader |
| http://www.fys.ruu.nl/~riel/ | <H.H.vanRiel@fys.ruu.nl> |
+-------------------------------------------+--------------------------+
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.rutgers.edu