High memory pressure makes system nearly unresponsive. 326

On 3 Feb 2006 20:19:58 GMT, Unruh

Well, if I run the pngtopnm pamdice pipe on a large PNG under Slackware, the swap usage grows to about 600MB, and UI interactivity suffers horribly. If I run the same pipeline against the same file in Windows under Cygwin, the swap usage grows to about 3.5GB, but the system remains responsive enough to load Opera and gvim in reasonable time, and actually be able to use them without cursing.

I don't think it could have much to do with swapping, anyhow. As I said in my original post:

^^^^^^^^^^^^^^^^^^^^^^^^^^ The "it" in the phrase "tried it" is the "catdev-zero" thing.

I covered that in my OP, as well:

Just for another example, while dicing a large ppm (it's about 3.5GB uncompressed, and stored as a 650MB gzip file):

nice -n 19 pamdice -w 512 -h 512 -outstem tx

I'm writing this in vim as that works. Somtimes the letters take several seconds to appear, and ":w" to write the file takes a couple of seconds to finish as well. I'm not swapping. Swap is off:

total used free shared buffers cached Mem: 1033516 1019536 13980 0 22544 819024 --+ buffers-cache: 177968 855548 Swap: 0 0 0

The netpbm utils are *very* conservative about memory usage, as they work one line at a time. If it weren't for buffers-cache, I'd have 850MB free RAM. Disk throughput on hdb is never more than 16MB-s, according to gkrellm, (and 0MB-s on hda, where the swap parbreastion lives) while hdb is capable of over 40MB-s when the wind is right. CPU usage spikes to 100%, but the processes are nice'd to 19. The only thing I can think that might be happening here is that the RAM that's in cache can't-won't be released to allow me to use vim without stuttering and hangups.

This just can't be normal behavior.

Jonathan

-- Don't just hit reply. Email address is broken. Thank your friendly neighborhood spammer. Email replies to: