Which distribution kit to choose for a computing cluster?

Question

Clusters

Given: an institute, a bunch of different semi-working rubbish.
Task: send all the megahertz to the calculation.
Let's say I have around 20 old system workers, completely different in configuration, workers. There is a desire to collect them together, introduce them to each other, and what would they consider together, for example, prime numbers.
Can you recommend the best tool for this task? I have been reading about Red Hat for a long time, is it relevant now? I myself am familiar with Linux at the user level, but I didn’t go deep into this topic, so I will be very glad to see any really useful links with the mathematical part!
Thanks in advance.

Answer 1

This is how it works here:
- RHEL 6.4 and CentOS 6.4
- servers with Ubuntu 14.04
- 1Gb/10Gb ethernet network + infiniband FDR/QDR (20/
40Gb
) NFS (mostly by IB, but there are also clients for 1Gb)
- openldap + kerberos for logins/passwords and SSO
- the calculations themselves are based on different libraries from OpenMPI to proprietary solutions.
- deployed the Sun Grid Engine to automate tasks on the cluster
In general, the conclusions are:
1) a separate network for internal computing 1/10Gb (infiniband)
2) a separate network for ethernet access 1/10Gb
3) a shared file storage for both networks
4) SSO based on kerberos
Well, it’s good that it’s all the same to run on a cluster, as long as it suits the task ...

Answer 2

it is logical to choose two mainstream distributions ubuntu and centos (free redhat).
Well, as always, count on one of the MPI implementations.

Answer 3

IMHO, in our time of a rise in the price of energy resources, it is not advisable to collect a cluster from trash.
Unless for tests and training.
The new hardware has an order of magnitude better ratio of gigaflops per watt, especially with accelerators.