User: Password:
Subscribe / Log in / New account

Swap: an idea whose time has come, and gone.

Swap: an idea whose time has come, and gone.

Posted Mar 23, 2007 8:38 UTC (Fri) by njs (guest, #40338)
In reply to: Swap: an idea whose time has come, and gone. by zooko
Parent article: Toward improved page replacement

We've hard this argument before, but...

It may be that in your situation, with current kernel memory management and workloads and hardware etc., running without swap improves your experience. It's a big honking jump to go from that to saying that swap is an idea whose "time has gone". If you want to make an argument that universal, you can't really just cite your particular set of factors and entirely ignore the theoretical reasons why swap is such a good idea...

Those being, of course, that swap lets the kernel toss out any pages of program memory that are rarely accessed in favor of any pages of data that are frequently accessed. Or more concretely, from a glance at top right now, it looks like my swap contains ~100M of firefox, ~1325M of evince (!! probably large rendered image caches?), 350M of matlab, 190M of openoffice, 40M of abiword... notice that the workload so far here is just "reading papers on data analysis" :-). Oh, and 350M of amarok... looks like a memory leak. (These numbers are probably not entirely accurate, since I'm just subtracting RES from VIRT, but you get the idea.)

So swap means that I have well over a gig of extra memory to cache things -- that's enough for, say, the entire source+build tree for a project I compile -- even if I didn't have that crazy evince stuff there. And it also means that I can happily run a dozen instances of evince to look at different high-res images, and not run out of memory in the process :-).

So, another data point for readers.

(Hey, I know -- next let's talk about swap-to-ramdisk!)

(Log in to post comments)

Copyright © 2017, Eklektix, Inc.
Comments and public postings are copyrighted by their creators.
Linux is a registered trademark of Linus Torvalds