Deuce

A two-node Beowulf Cluster running Redhat Linux

Two PCs:

	Master node: 
		Pentium II 300
		64M RAM
		40GB hard drive
		2 100bT Ethernet NICs
			3Com
			Linksys

	Slave node: 
		AMD K6-2 333
		48M RAM
		4GB hard drive
		SMC 100bT NIC

	Network:
		Intel 8-port 10bT ethernet hub
		Cat 5 ethernet cables

	Software:
		OS: Redhat Linux 9.0
			http://www.redhat.com
		Message passing: MPICH
			http://www-unix.mcs.anl.gov/mpi/mpich/





"Deuce" Beowulf cluster
Deuce is the third Beowulf cluster I've built. The first was Mini-wulf, which runs FreeBSD, and the second is Zeus, which is still being put together as I write this.

Purpose:
I built Deuce to gain some experience in setting up clusters based on Redhat Linux. Since Zeus will run this OS, setting up a minimal cluster with two old computers seemed like a good weekend project.

Build:
The hardware build for Deuce was pretty straight forward. I added the second network card in the Pentium II, configured both systems to boot from their CDROM drives, and cabled things together with the hub. I made install CDs from ISO images of Redhat 9.0 from a mirror server, booted each box in turn, and ran the install program in text mode (since neither system had a mouse attached).

I chose an internal schema of 10.0.0.x for the LAN, since I was already using 192.168.1.x on my home network, and didn't want to get the two confused. The Redhat installer wasn't specific as to which ethernet interface was which card, so I went ahead and configured the two and waited to cable the master node until I could determine that later. I used the 'custom' system type in the installer, and added all the development packages I thought would be needed (gcc, glibc, etc). The slave node I left pretty bare of compilers, since none would be needed on it. Both systems got network services they would need for things like ssh, rsh, and rlogin, as well as NFS and NTP. Since deuce lives on my home network, behind a NAT router, I disabled the firewalls on the master and slave. Normally one would run a firewall only on the master node, and modify it so the slave node(s) can attach to NFS, NTP, and rsh on the master.

After initial setup, I plugged the nodes into the hub and tried pinging the slave from the master. Trial and error revealed which ethernet interface was which, and I hooked up my home network to the 'outside' interface on the master. I next modified the /etc/hosts file (master slave) on both nodes, adding the 'master' and 'slave' IP number assignments. The /etc/ntp.conf on the master was pointed at an internet time server, and the slave /etc/ntp.conf was pointed to the master.

Home directories were shared by adding:
/home	10.0.0.0/255.255.255.0(rw)
to the /etc/exports file on the master. I moved the link in /etc/rc3.d from K20nfs to S20nfs to cause nfs to start on boot. After running '/etc/rc3.d/S20nfs restart' to start the nfs server, I modified the slave /etc/fstab file by adding:
master:/home	home	nfs	rw	0 0
and mounted the directory with 'mount /home'. I created a personal account on both machines, making sure to use the same uid, shell, and home directory path.

Now the fun part: getting rsh to work. I added 'master' and 'slave' to the /etc/hosts.equiv file on both nodes, which should allow rsh to run without individual .rhosts files in each home directory. No dice (it didn't work). I had to edit the /etc/xinetd.d/rsh and /etc/xinetd.d/rlogin files, setting 'disable' to 'no'. in /etc/pam.d/rlogin I moved the "auth sufficient pam_rhosts_auth.so" line to the top, which according to the Cluster quick start guide allows root to rlogin without a password (scary). I also verifyed that /etc/pam.d/rsh existed. I did a '/etc/rc3.d/S56xinetd restart' to make the changes take effect, and verified rsh functionality with 'rsh [node] date', where [node] was the other cluster node on each machine as I tested it (i.e. I did 'rsh slave date' on the master node).

I downloaded MPICH 1.2.5 and ran './configure --prefix=/usr/local/mpich-1.2.5'. 'make' and (as root) 'make install' built and installed the package. I modified the /usr/local/mpich-1.2.5/share/machines.LINUX file to list the master and slave nodes, and added /usr/local to /etc/exports. I mounted /usr/local on the slave node to allow access to mpich, and modified my path to include /usr/local/mpich-1.2.5/bin for access to the executables.

To test the system, I copied over the cpi.c program, compiled it with 'mpicc cpi.c -o cpi', and ran the program with one and two nodes via 'mpirun -np 1 ./cpi' and 'mpirun -np 2 ./cpi'. Here's the output:
[user@master ~/pi]$ mpirun -np 1 ./cpi
Process 0 of 1 on master
pi is approximately 3.1415926535897287, Error is 0.0000000000000644
wall clock time = 3.253982

[user@master ~/pi]$ mpirun -np 2 ./cpi
Process 0 of 2 on master
pi is approximately 3.1415926535899814, Error is 0.0000000000001883
wall clock time = 1.634312
Process 1 of 2 on slave

For further fun, I ran the primitive benchmark program I made by modifying the cpi program on this cluster. Here are the results: 
	Number of nodes		MFLOPS
	---------------		------
	       1		 39.7
	       2		 45.4


Conclusion:
This experiment taught me some of the specifics of setting up Redhat linux 9.0 cluster systems. These will prove valuable in setting up and trouble-shooting future clusters.

Update: May 19, 2003
Replaced 10/100 switch with 10bT hub. Although this cuts down the bandwidth quite a bit, it did solve a flapping (NIC speed flip-flopping) problem which was causing packet loss. I also ran the cpi benchmark on 1 - 20 processes to gauge the scalability of the cluster. This gave some rather odd results: running three processes gave the optimum crunching ability on the two CPU cluster. I'm not sure why this is.

Update: May 21, 2003
Ran the Pallas Benchmark on Deuce. Here are the results. It looks like Deuce is actually maxing out the 10bT hub in some instances, and getting close in others. The large step discontinuity that Mini-wulf displayed is not evident on Deuce.

Update: October 23, 2003
Deuce has gone the way of Mini-wulf. Having served its purpose (research on Redhat clustering), I pulled the second NIC from the master node, reformatted the disk, and retasked the machine. The compute node still runs, but with no compiler or other tools, it's not much use. I'll likely reformat it with Fedora, as a test of that new packaging of the old Redhat Linux distro.

So now I'm down to one cluster: Zeus. Building clusters out of old computers is a real kick. I highly recommend it to any geeks out there with too many computers on their hands. Enjoy!

Links:


Tools for building clusters:
Other cluster stuff: