Hello On Mon, Sep 18, 2006 at 08:57:57PM +0200, Andreas Beckmann wrote: > Package: user-mode-linux > Version: 2.6.17-1um-2 > Severity: normal > > Hi, > > running several uml guests can consume the whole /dev/shm quickly: e.g. > host mem: 1GB > /dev/shm: 507 MB (default tmpfs size) > 2x linux mem=256m: 512MB ==> greater than /dev/shm > One uml guest crashed, the other could be shut down, but this took > nearly forever.
Yes, that's something that hit me too. I'm running on a 256M host and firing up up to 3 UML guests with mem=128M usually. I did notice that with 2.6.17 the situation got a definitely worse. Something that helps is obviously increasing /dev/shm size (I doubled it and I'm now an happy citizen). > Obviously the 'physical memory' used by the uml guests is taken from > /dev/shm and won't be freed again. not nice and definitely not that obvious. This could be considered a bug, but I think upstream is well aware of it. Anyway, could you check if you still have that UML instance running? I just reproduced something similar here and 'killall -9 linux' frees all memory: $ df -h Filesystem Size Used Avail Use% Mounted on /dev/hda1 2.8G 1.4G 1.5G 49% / tmpfs 128M 128M 0 100% /dev/shm $ killall -9 linux $ df -h Filesystem Size Used Avail Use% Mounted on /dev/hda1 2.8G 1.4G 1.5G 49% / tmpfs 128M 4.0K 128M 1% /dev/shm > linux --help says "mem= ... is not releated to the amount of memory in > the host. ... can be more ... will be swapped out." > Even using half of the host memory seems to be impossible under this > situation. > Is there a way to reduce /dev/shm usage? > > The host has enough swap (4GB) and tmpfs space available at /tmp (3GB). > More could be made available, if neccessary. You need to increase /dev/shm size, default is half of the available ram but you can edit /etc/default/tmpfs and put e.g.: SHM_SIZE=536870912 for 512M (it's in bytes). Or try mount -t tmpfs tmpfs /dev/shm -o remount,size=536870912 before, to test it. -- mattia :wq! -- To UNSUBSCRIBE, email to [EMAIL PROTECTED] with a subject of "unsubscribe". Trouble? Contact [EMAIL PROTECTED]
