serial pbulk: Re: suggestion: Host fly-off between pkgin and nih and subsequent official integration

[John Marino <netbsd%marino.st@localhost>]

On 4/17/2013 10:35, Thomas Klausner wrote:
> Why not run multiple pbulk clients on the machine? That should help
> parallelize.

It's definitely not a fair comparison.
Poudriere has been modified to copy a full world in a tmpfs, and then 
each slave nullmounts to it (creating a readonly copy, but then writable 
areas are tmpfs mounted to those).

So with one jail, slaves are created, destroyed, refreshed pretty much 
instantly in memory + swapcache versus the pbulk approach of unrolling a 
tarball.  So it's not just a matter of adding slaves, but also improving 
resource use.

Basically I haven't had time to look at what needs improving, now if 
distbb is a better approach, or porting poudriere would even more 
attractive.  Don't know.

I give you the configuration for pkgsrc is nowhere close to optimal and 
I can squeeze a lot more out of the machine.