k7s5a mobo based cluster

Dr. Rochus Schmid rochus.schmid at ch.tum.de
Fri Apr 12 12:33:18 EDT 2002


dear wulfers,

i recently assembled a tiny and cheap (i know :-) cluster using the ecs
k7s5a mobo (SIS 735 chipset).
the board is very cheap (~75 €) and comes with FE (SIS900) onboard.
just want to tell about my experience (other hardware, netbooot  +
stream / netpipe results).

i would be happy to know of others using this board. this mail also
contains some questions ... maybe someone can answer or help me here?
if this doesn't interst you please skip - apologies for the bandwidth.

#############
HARDWARE
(currently 4 nodes .. hope to get 4 more :-)
   mobo: k7s5a
   cpu:  athlonxp 1,4 ghz (1600+)
   ram: 256 MB DDR (266MHz, CL2)
   graphics: various pci/agp graphics cards i could find
   floppy
  small tower with 250W PS

nodes are diskless, master has an additional 40GB IDE disk.
switch: D-Link DES 1008D 8 port switch.

##############
POWER /GRAPHICS

the cluster has continuously been up for about 3 weeks now with quite
some load for most of the time. as far as i can tell, the 250W seems to
be ok for the board and the 1,4 ghz athlonxp.

the ami-bios does not allow booting without a graphics adapter. someone
on the net (using a lot of the boards for a SETI at home "farm" told me
that he did not get around it even with teaking tools for the ami-bios.
i am happy to have a console for maintenance but one has to find a
cheapo graphics card  ... anyone out there managed to avoid this?

##############
NETBOOT / OS

i run a RH7.2 on it with a 2.4.17 kernel with NFS-root.
the bios supports the RPL protocol for netbooting. i tried the rpld for
linux and the board seems to communicate with the rpl-server and
download something, but i didnt get it to boot. the rpld-developers sent
me some patch to "switch off a DMA channel" of the onboard NIC but i
have to admit that i didnt really understand what to do, nor did i try
it. i currently boot from a syslinux floppy and use NFS-root.
did soemone manage to netboot linux with this hardware?

###############
STREAM

because of the comments on the gcc versions 2.96 versus 2.95 issue
mentioned on the ATLAS webpages i reinstalled the gcc 2.95 and found
differences also for stream results and therefore i will post both here
(both compiled with -O2 for comparison)
Array size = 2000000, Offset = 0

gcc version 2.96 20000731 (Red Hat Linux 7.1 2.96-98)
-O2
Function      Rate (MB/s)   RMS time     Min time     Max time
Copy:         666.0980       0.0489       0.0480       0.0559
Scale:        585.3939       0.0547       0.0547       0.0549
Add:          726.1178       0.0662       0.0661       0.0663
Triad:        679.6655       0.0707       0.0706       0.0707

gcc version 2.95.3 20010315 (release)
-O2
Function      Rate (MB/s)   RMS time     Min time     Max time
Copy:         727.6031       0.0440       0.0440       0.0443
Scale:        627.5864       0.0526       0.0510       0.0649
Add:          798.1775       0.0602       0.0601       0.0603
Triad:        727.6691       0.0660       0.0660       0.0661

i guess it is obvious to reinstall gcc-2.95 when using RH7.1 or RH7.2.
these results are not as good as reported recently for the nforce
chipset.
i tried to set the bios settings for the ddr-ram to optimal, but i didnt
test/experiment.

##########
NetPipe-2.4

The following results are NOT from a crossconnect cable but measured
throug the D-Link switch!!

kernel 2.4.17 / NIC-driver SIS900
for MPI: LAM-MPI 6.5.1

NPtcp:
    latency: 33 us
    bandwidth: ~89.7 MBit/s

NPmpi:
    latency: 41 us
    bandwidth: ~82 MBit/s (maximux at about 85 MBit/s)

the latency of around 40 microsec seems to be very low as far as i can
tell from the information on the net (i am absolutly a beginner in this
field). is there anything one can seriously do wrong? i tried it a
couple of times. between different nodes always with basically the same
result.

###################
i hope i did not anoy the pros on this list too much, and this is
helpfull for comparison.
again: please contact me off list if you also use this type of hardware.

thanks and best greetings from munich,

    rochus



--

Dr. Rochus Schmid
Technische Universität München
Lehrstuhl f. Anorganische Chemie
Lichtenbergstrasse 4, 85747 Garching

Tel.    ++49 89 2891 3174
Fax.    ++49 89 2891 3473
Email   rochus.schmid at ch.tum.de


_______________________________________________
Beowulf mailing list, Beowulf at beowulf.org
To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf



More information about the Beowulf mailing list