Problem Booting the Slaves

Christopher A. Busillo cbusillo at yahoo.com
Sat Dec 16 18:54:36 PST 2000


The 64meg minimum is not entirely true.  I am
currently running a test beowulf with between 24 and
48megs of ram in the nodes.  Of course the mem usage
is 70-90% under no load.  But then again, the cluster
is a small scale, 10 node model, sitting at the foot
of my bed.  I'm pretty sure the limit is between 20
and 24 megs, though.  Page faults occur with 16 megs. 
I assume the node runs out of memory, since the page
file is not yet mounted.

----------------------------
Christopher Busillo
Excalibur IT Solutions
cbusillo at excalibur-scs.com
----------------------------



--- Donald Becker <becker at scyld.com> wrote:
> > On Fri, 15 Dec 2000, David Leunen wrote:
> > > monte: command line: panic=30
> ramdisk_size=131072 apm=power-off
> ...
> > > after that, there are all the register and stack
> status, and it is
> > > frozen. The slaves are P200 16Ram Diskless. I
> tried with a double PII
> > > 256Ram Diskless and it turn 'unavailable' after
> reboot. What is wrong
> > > with the other ones?
> 
> On Fri, 15 Dec 2000, Dave Leimbach wrote:
> 
> > There is a parameter in that crash output about
> RAMDISK size.  If your
> > machine only has 16MB I think you may be out of
> luck.  You need to get the
> ...
> > The one that comes up unavailable has 256MB of RAM
> which is plenty of RAM
> > for the ramdisks.  Scyld has an option in the
> beowulf distribution to run
> > diskless.  I assume that this requires a large
> ramdisk that you can't get
> > on the 16MB node.
> 
> That is correct.  We recommend 128MB for the compute
> nodes.
> It's possible to run with only 64MB, but not less.
> 
> This minimum memory is the same for disk-based and
> diskless nodes.
> The compute nodes always start as diskless nodes
> running from "ramdisk",
> even if they get their initial boot kernel from the
> hard disk.
> 
> The nodes immediately start accepting commands from
> the master, before
> mounting any filesystems.  Running in "diskless"
> mode allows the master to
> control verification of file system integrity
> ("fsck"), create new file
> systems, and recover checkpoint files without the
> risk of a node boot script
> corrupting the system.
> 
> The bulk of the memory use is cached libraries. 
> 'Bproc' automatically
> transports missing/uncached libraries with the
> application executable, thus
> not every library on the system must be cached.  But
> the libraries used by
> the housekeeping processes ("slave daemons") on the
> compute node must always
> exist locally.
> 
> 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
> 
> 
> _______________________________________________
> Beowulf mailing list
> Beowulf at beowulf.org
> http://www.beowulf.org/mailman/listinfo/beowulf


__________________________________________________
Do You Yahoo!?
Yahoo! Shopping - Thousands of Stores. Millions of Products.
http://shopping.yahoo.com/




More information about the Beowulf mailing list