Do Beowulf clusters have to use entirely identical PCs?

[Nukemann]

I am building a Linux-based Beowulf cluster for a computer science course. The computers will be provided by my school, but do the computers have to be entirely identical in specs? Ex. One is a P4 3 GHz and 512MB RAM but the other one is a Core 2 Duo 2.0 GHz, 2GB RAM. Would that work in a Linux Beowulf? All the guides I found mention two identical PCs, but how identical do they have to be, hardware-wise?

 

[ch424]

They can be different if the scheduler knows they're different, or if the workload is in lots of small chunks.

For example, imagine you had a simulation to run, and you gave half of the work to each of the computers. The C2D would finish much faster, and you would have to wait for the much slower P4 to finish while the C2D sat idle.

However, if you could chop the work into 100 small sections, and give the computers a new section each time they finished the last, there would be less waiting time. For example, the C2D might end up doing 60 of the jobs while the P4 only managed 40, but they would finish at around the same time as each other.

What are you using as the scheduler? Or will you write it yourself?

 

[easp]

A two node cluster with obsolete equipment is just a learning exercise anyway, so don't worry too much about it.

The bottom line is that it should work, so get it working, and then go from there. You may need to worry about the scheduling concerns that ch424 mentions, but that is almost certainly going to be an optimization. What is it I heard about premature optimization...