Company computer utilization?
becker at scyld.com
Tue Jul 16 15:48:24 PDT 2002
On Tue, 16 Jul 2002, Mark Hartner wrote:
> > this, except I need or at least prefer them to be linux. Therefore, we need
> > a way to auto reboot the boxes at some time in the evening to linux and then
> > back to windows in the morning.
As several postings have pointed out, there is software that handles the
reboot scheduling and alternating between systems.
> One of our IT folks is already working on
> You may want to take a look at GRUB. Newer versions of GRUB can do a
> network boot (for some subset of network cards),
Not quite. GRUB has support for the 'netboot/etherboot' system, so the
you must first have 'netboot' support for the network adapter you are using.
[[ I don't see a compelling reason to make GRUB and netboot depend on
each other, but... ]]
> so you could just put your linux kernel and initrd our on a tftp
> server and boot into linux without having to install anything local to
> the PC's.
Errrm, exactly the same as having a network boot ROM.
> I have done this before using a floppy.
Yes, using a floppy is the way to avoid needing a disk partition or
Our boot floppy accomplishes the same thing. We use Two Kernel Monte
("TKM"), developed here at Scyld, to switch from the kernel on the
floppy to the final kernel.
Here are the tradeoffs:
Using 'GRUB' and netboot
- avoids the need for TKM,
- limited choice of network device drivers and
- you will encounter the TFTP reliability problems with >20 machines
booting at once (a problem with clusters, but not typical TFTP
With Scyld Beoboot and TKM
- You may need an updated TKM for BIOS and OS changes.
- All Linux network device drivers are available
(If it's not supported, you won't be able to use it in the cluster anyway)
- Choice of protocols to download the stage 2 image.
> > anyone has any ideas/suggestions, etc. I'd appreciate it. It has to be
> > automatic because most of the users would probably not do anything as simple
> > as reboot their computer when they go home.
> Then they probably would not do something as simple as "save their
> work" before they go home either.
That's a serious problem for off-hours clustering. A headless compute
node is significantly less expensive than one day's lost work.
Donald Becker becker at scyld.com
Scyld Computing Corporation http://www.scyld.com
410 Severn Ave. Suite 210 Second Generation Beowulf Clusters
Annapolis MD 21403 410-990-9993
More information about the Beowulf