[pve-devel] mass deployment for testing ipcc_send_rec failed

[Stefan Priebe]

Am 28.05.2014 21:58, schrieb Daniel Hunsaker:
>  > > seems like that - yes. But do you know why this happens? it happens
> already
>  > > starting 15 vms in parallel while everything is idle.
>  >
>  > That is quite normal behavior.
>  >
> There's a lot that goes on when starting a VM, and it consumes resources
> like crazy.  Each VM (once the KVM process starts up) has to create all
> of its virtual hardware, initialize it, and start the boot process.
> That involves a lot of disk activity, since the virtual devices need to
> be loaded individually, and the boot data is also not only stored on
> disk, but frequently inside a virtual disk file.  There's a lot of other
> stuff happening, too, but these are the main resource hogs.
>
> The process tends to be pretty quick most of the time, so it may not
> register on load averages (and I/O wait doesn't always show in load
> averages anyway), but it gets pretty intense.  So it's perfectly normal
> behavior that makes a lot of sense.  Containers will be somewhat less
> demanding, due to their nature, so you might be able to start more of
> them at once, but there will still be limits even there.  This is why
> the on-boot VM startup process works the way it does.  I forget where
> that logic is stored, but that approach would probably be your best bet,
> at least until the changes Dietmar mentioned earlier can be merged and
> released.
>

Disks are on external ceph system - so compute node is not under load. 
15 VMs for 48 cores are not much.

Stefan