Additional Documentation:

In this additional documentation page, I will log all the small tid-bits of configuration, tweaking, software installation etc. I have done; which may not have been important to the working of the cluster as a whole but which perhaps makes life a lot easier when you have to live with the cluster.
 

• The cluster comprises a private network (subnet) with IP no.s ranging from 10.0.0.1 (for the master) to 10.0.0.8. The master, with two NICs, has two IP no.s: one private to the cluster and the other connected to the LAN. A routing problem of "bootp" packets had to be taken care of by adding the following line at the end of /etc/rc.d/rc.local file of the master.

 

 

route add -host 255.255.255.255 eth0
(eth0 is the network interface of the master to the cluster subnet while eth1 connects to the LAN).
 

• Installed and configured rsh services, for seamless (read password-less) logging into all nodes from any other node (including master).

• Created file-system on the secondary IDE Hard-disk with "mke2fs -b 4096 /dev/hda1". This 12 Gb disk is meant for secondary storage of codes and results. Since the codes are all going to be large files, the block size of 4kb was used to obtain faster raw reads, less fragmentation etc. Could not enable multiword DMA and 32-bit disk access modes because the IBM disk did not seem to support the parameters passed by hdparm. (kamadhenu hung when I tried, the only time!)

• Installed prsh, a parallel shell specially meant for Beowulf Clusters. It enables sending the same command parallely to all the nodes of the cluster, without logging into the each of the nodes individually. After installation, it requires that every user of the cluster has the "PRSH_HOSTS" environment variable set through his shell initialization file, so that he can type the command

           to send the <command> parallely to all nodes. However, it is advisable that the user sets his "PRSH_HOSTS" variable only to the slave nodes so that he won't bring disaster upon himself by mistakenly typing

bash$prsh -- reboot

Alternatively, one can specify all the node names (to which he wants to send the command to) in the command-line itself as:

bash$prsh node3 node6 node8 -- tkill
 

• Installed bWatch, a Tcl/Tk cluster monitoring tool, that gives CPU load, free memory etc. of the nodes in a neat GUI, without being resource hungry.

• Installed LAM-MPI 6.3.2 downloaded from University of Notre-dam and also ran the LAM-TEST suite also available from the same site.

• Installed "Smile Queueing Management System" a cluster management software from Thailand that enables queueing of jobs in a cluster when there are multiple users involved.

• Installed "Smile Cluster Management Suite", a cluster management suite again from Thailand, that is meant for system administration of a cluster. However, this software was found to be extremely resource hungry on the master, and was not really used.

• Applied the TCP patch to the 2.2.14-5.0 kernel that adjusts the TCP timout values and acknowledgement packet requests for a private cluster, as is the case here.

Considerable improvement, as regards the bandwidth and latency values were obtained as evident in the performance results page.
 
• Installed Code-Crusader, the Integrated Development Environment that is fantastically configurable and hence allows one to do a professional job at code development in scalar as well as parallel environment. I would suggest people doing scalar code development also give it a try because its really good.(You would want to throw away the Turbo C compiler and stuff!). It also supposedly has tight integration with Code-Medic, a graphical Debugging Environment. By the way, the people who developed Code-Crusader are not paying me to write all this !(they believe in Open Source anyway, so they can't be so cheap, right?)

• MPICH 1.2.1 was installed.

• Compiled and installed the Linux 2.2.16-3 kernel with Josip Loncaric's TCP patch.

Goto Starting Page...
Goto Configuration page...
Goto Benchmarking page...
Goto performance results page...

---

Page last updated: 27th Jan, 2001. 

---