Re: [Nbd] nbd for large scale diskless linux -> tests & problem

[Wouter Verhelst <wouter@...3...>]

On Sat, Oct 08, 2005 at 10:45:24PM +0200, dsuchod wrote:
> Hi!! I did some more tests:
> 
> Booting of 15 machines simultaniously (each connected via separate 100Mbit 
> switch port, server on gigabit uplink of that switch) got me the 
> following:
> 
> Booting rather fast, 231MByte down 6MByte up for each client compared to
> nfs with 312MByte down and 13Mbyte up. Then on six clients: cat /dev/nb0>
> /dev/null (little bit pervert :-))
[...]
> So no much cpu consumption, but some of them waiting for IO. During bootup 
> the servers gigabit link got up to 360MBits, little bit less with the 
> "cat" ... With all the "cats" after a while around 230MBits ... A client
> is still operable (graphical login into KDE) while cat is running. I 
> have to purchase faster disks/and server memory now ...

Heh.

All very interesting, really. Would you object to me posting this story
on the nbd homepage?

> I think nbd could be a real killer for nfs :-)
> 
> But I had a problem within my ramdisk setting up the nbd on some clients 
> (dell 2,8Ghz P4, 1GB Ram, broadcom5700 gbit ethernet). I had to add a
> "sleep 10" after powering up nbd-client server 5000 /dev/nb0 ...
> Otherwise the system would crash because the device was not ready for
> ext2 mount. It seems that the autoconfiguratin of the ethernet link
> interupts the traffic and the nbd-client can not cope with that :-(

Hmm?

Is this an educated guess, or did you run some diagnostics that pointed
you in this direction?

If the nbd-client gives control back to the parent process too fast,
then that would be a bug in either nbd-client or the kernel. But I'd
need some more info to be sure...

-- 
The amount of time between slipping on the peel and landing on the
pavement is precisely one bananosecond