At first you have to setup a DHCP-server which answers the DHCP-request
for an ip-address when a disk-less client boots. This DHCP-Server (i call it
master in this HOWTO) acts additional as an NFS-server which exports the
whole client-file-systems so the disk-less- cluster-nodes (i call them
slaves in this HOWTO) can grab this FS (file-system) for booting as soon as
it has its ip. Just run a "normal"-MOSIX setup on the master-node. Be
sure you included NFS-server-support in your kernel-configuration. There
are two kinds (or maybe a lot more) types of NFS:
It does not matter which one you use but my experiences shows to use
kernel-nfs in "older" kernels (like 2.2.18) and daemon-nfs in "newer"
ones. The NFS in newer kernels sometimes does not work properly. If your
master-node is running with the new MOSIX-kernel start with one file-system
as slave-node. Here the steps to create it: Calculate at least 300-500 MB
for each slave. Create an extra directory for the whole cluster-file-system
and make a symbolic-link to /tftpboot. (The /tftpboot-directory or link is
required because the slaves searches for a directory named
/tftpboot/ip-address-of-slave for booting. You can change this only by
editing the kernel-sources) Then create a directory named like the ip of
the first slave you want to configure, e.g. mkdir /tftpboot/192.168.45.45
Depending on the space you have on the cluster-filesystem now copy the
whole filesystem from the master-node to the directory of the first slave.
If you have less space just copy:
/bin
/usr/bin
/usr/sbin
/etc
/var |
You can configure that the slave gets the rest per NFS later.
Be sure to create empty directories for the mount-points.
The filesystem-structure in /tftpboot/192.168.45.45/ has to be similar to
/ on the master.
/tftpboot/192.168.45.45/etc/HOSTNAME //insert the hostname of the slave
/tftpboot/192.168.45.45/etc/hosts //insert the hostname+ip of the slave |
Depending on your distribution you have to change the ip-configuration of
the slave :
/tftpboot/192.168.45.45/etc/rc.config
/tftpboot/192.168.45.45/etc/sysconfig/network
/tftpboot/192.168.45.45/etc/sysconfig/network-scripts/ifcfg-eth0 |
Change the ip-configuration for the slave as you like.
Edit the file
/tftpboot/192.168.45.45/etc/fstab //the FS the slave will get per NFScoresponding to
/etc/exports //the FS the master will export to the slaves |
e.g. for a slave fstab:
master:/tftpboot/192.168.88.222 / nfs hard,intr,rw 0 1
none /proc nfs defaults 0 0
master:/root /root nfs soft,intr,rw 0 2
master:/opt /opt nfs soft,intr,ro 0 2
master:/usr/local /usr/local nfs soft,intr,ro 0 2
master:/data/ /data nfs soft,intr,rw 0 2
master:/usr/X11R6 /usr/X11R6 nfs soft,intr,ro 0 2
master:/usr/share /usr/share nfs soft,intr,ro 0 2
master:/usr/lib /usr/lib nfs soft,intr,ro 0 2
master:/usr/include /usr/include nfs soft,intr,ro 0 2
master:/cdrom /cdrom nfs soft,intr,ro 0 2
master:/var/log /var/log nfs soft,intr,rw 0 2 |
e.g. for a master exports:
/tftpboot/192.168.45.45 *(rw,no_all_squash,no_root_squash)
/usr/local *(rw,no_all_squash,no_root_squash)
/root *(rw,no_all_squash,no_root_squash)
/opt *(ro)
/data *(rw,no_all_squash,no_root_squash)
/usr/X11R6 *(ro)
/usr/share *(ro)
/usr/lib *(ro)
/usr/include *(ro)
/var/log *(rw,no_all_squash,no_root_squash)
/usr/src *(rw,no_all_squash,no_root_squash) |
If you mount /var/log (rw) from the NFS-server you have on central
log-file! (it worked very well for me. just "tail -f /var/log/messages" on
the master and you always know what is going on)
The cluster-filesystem for your first slave will be ready now. Configure
the slave-kernel now. If you have the same hardware on your cluster you
can reuse the configuration of the master-node. Change the configuration
for the slave like the following:
CONFIG_IP_PNP_DHCP=y
and
CONFIG_ROOT_NFS=y |
Use as less modules as possible (maybe no modules at all) because the
configuration is a bit tricky.
Now (it is well described in the Beowulf-HOWTO) you have to create a
nfsroot-device.
It is only used for patching the slave-kernel to boot from NFS.
mknod /dev/nfsroot b 0 255
rdev bzImage /dev/nfsroot |
Here "bzImage" has to be your diskless-slave-kernel you find it in
/usr/src/linux-version/arch/i386/boot after successful compilation.
Then you have to change the root-device for that kernel
and copy the kernel to a floppy-disk
dd if=bzImage of=/dev/fd0 |
Now you are nearly ready! You just have to configure DHCP on the master.
You need the MAC-address (hardware address) of the network card of your
first slave.
The easiest way to get this address is to boot the client with the already
created boot-floppy (it will fail but it will tell you its MAC-address). If
the kernel was configured alright for the slave the system should come up
from the floppy, booting the diskless-kernel, detecting its network-card
and sending an DHCP- and ARP request. It will tell you its hardware address
during that moment! It looks like : 68:00:10:37:09:83.
Edit the file /etc/dhcp.conf like the following sample:
option subnet-mask 255.255.255.0;
default-lease-time 6000;
max-lease-time 72000;
subnet 192.168.45.0 netmask 255.255.255.0 {
range 192.168.45.253 192.168.45.254;
option broadcast-address 192.168.45.255;
option routers 192.168.45.1;
}
host firstslave
{
hardware ethernet 68:00:10:37:09:83;
fixed-address firstslave;
server-name "master";
} |
Now you can start DHCP and NFS with their init scripts:
/etc/init.d/nfsserver start
/etc/init.d/dhcp start |
You got it!! It is (nearly) ready now!
Boot your first-slave with the boot-floppy (again). It should work now.
Shortly after recognizing its network-card the slave gets its ip-address
from the DHCP-server and its root-filesystem (and the rest) per NFS.
You should notice that modules included in the slave-kernel-config must
exist on the
master too, because the slaves are mounting the /lib-directory from the
master. So they
use the same modules (if any).
It will be easier to update or install additional libraries or
applications if you mount
as much as possible from the master. On the other hand if all slaves have
their own complete
filesystem in /tftpboot your cluster may be a bit faster because of not so
many read/write
hits on the NFS-server.
You have to add a .rhost file in /root (for user root) on each
cluster-member which should look like this:
node1 root
node2 root
node3 root
.... |
You also have to enable remote-login per rsh in the /etc/inetd.conf. You
should have these two lines in it
if your linux-distribution uses inetd:
shell stream tcp nowait root /bin/mosrun mosrun -l -z /usr/sbin/tcpd in.rshd -L
login stream tcp nowait root /bin/mosrun mosrun -l -z /usr/sbin/tcpd in.rlogind |
And for xinetd:
service shell
{
socket_type = stream
protocol = tcp
wait = no
user = root
server = /usr/sbin/in.rshd
server_args = -L
}
service login
{
socket_type = stream
protocol = tcp
wait = no
user = root
server = /usr/sbin/in.rlogind
server_args = -n
} |
You have to restart inetd afterwards so that it reads the new
configuration.
/etc/init.d/inetd restart |
or
There may be another switch in your distribution-configuration-utility
where you can configure the security of the system. Change it to "enable
remote root login". Do not use this in insecure environments!!! Use SSH
instead of RSH! You can use MOSIXVIEW with RSH or SSH.
Configuring SSH for remote login without password is a bit tricky. Take a
look at the "HOWTO use MOSIX/MOSIXVIEW with SSH?" at this website.
If you want to copy files to a node in this diskless-cluster you have now
two possibilities. You can use rcp or scp for copying remote or you can
use just cp and copy files on the master to the cluster-filesystem of one
node. The following two commands are equal:
rcp /etc/hosts 192.168.45.45./etc
cp /etc/hosts /tftpboot/192.168.45.45/etc/ |
