Scyld and fstab for Diskless slaves

Chip Coldwell coldwell at
Tue Dec 19 07:57:00 PST 2000

Interesting, so what is the mechanism for transmitting the compressed
root filesystem image to the node, if not tftp?  I suppose you could
configure the kernel to mount an NFS filesystem, retrieve a compressed
filesystem image from it, expand it into a ramdisk, and then remount
the result as /, but that's pretty kludgey.

BTW, if I'm getting close to any of Scyld Computing's IP it's not
because I want to avoid buying something from you ... it's just that I
want to make sure that the cluster I spec will work with your
software.  In particular, I'm planning to spec nodes that have no
local persistant storae whatsoever, i.e. no hard disk, no floppy and
no cd-rom (the less there is, the less there is to go wrong).  I know
this configuration will work with my NFS-root diskless configuration
because I've done it, but I like your ramdisk system better when the
network is the bottleneck (and mine better when memory is ...).

Thanks for the prompt reply,


Charles M. "Chip" Coldwell
"Turn on, log in, tune out"

On Tue, 19 Dec 2000, Donald Becker wrote:

> On Tue, 19 Dec 2000, Chip Coldwell wrote:
> > I know how to compile a kernel to NFS mount a root filesystem, and I
> > have set up diskless clusters using this method; I'm very curious how
> > the ramdisk root filesystem works.  Does the node tftp down a
> > compressed filesystem image and load it into the ramdisk analogously
> > to the root floppies of yore, except coming over tftp instead of from
> > a floppy?
> Yes, that's roughly the idea.
> We avoid TFTP where possible because of our experience with performance
> problems and unreliability with TFTP on larger cluster.  The network is
> pretty busy when all of the compute nodes are booted at once.  A simple
> non-windowed protocol with short timeouts and without congestion control can
> lose badly.  There is a reason that TFTP isn't used over the Internet :->.
> Multicast-TFTP potentially solves the problem, and you might see that
> introduced for PXE boots on large clusters.
> >  Does this require custom kernel modifications, or is it a
> > functionality available in stock kernels?
> > On Fri, 15 Dec 2000, Daniel Ridge wrote:
> > > Also, while we set up a beefy ramdisk during boot -- this space is
> > > reclaimed when the node finishes booting if you aren't doing a ramdisk
> > > root.
> ...
> > > Our upcoming contains a lot of effort directed towards reducing the size
> > > of ramdisk. This is to make the system work on smaller memory machines
> ...
> > > Because we can now do this, I expect that the next version of our product
> > > will ship with ramdisk root as the default.
> Donald Becker				becker at
> Scyld Computing Corporation
> 410 Severn Ave. Suite 210		Second Generation Beowulf Clusters
> Annapolis MD 21403			410-990-9993

More information about the Beowulf mailing list