The TONS idea

The TONS system is an attempt to create for scientific computing, an abstraction that frees the user from being bothered by parallelization issues. The two main tasks the system will take care of, that are interesting to us in this report, are:

• Automatic parallelization of code
• Scheduling, distribution and execution of that code

One could argue, that this is no different from using a parallelizing compiler and a batch queuing system. I disagree, of course. A parallelizing compiler does not parallelize the code with regard to the current state of the parallel machine. TONS does that to some extent already^1.2. A queuing system won't start a parallel program requiring $N$ CPUs, before $N$ CPUs are available, which in turn means that $N-1$ CPUs have been idle for some time. A system like TONS has the potential to perform much smarter scheduling, although this is not currently implemented.

I cannot see how such functionality could reasonably be implemented using compilers and queuing systems as we know them today. Since the parallelization involves code changes, depending on the current circumstances, traditional tools would probably require a timely compile of the user code, every time it was to run. TONS does not build upon compilers and queuing systems, rather it bends the rules by not compiling the code, which in turn makes smarter scheduling possible.

TONS is a system that executes interpreted code. It is a distributed virtual machine, that will execute code on a cluster of machines.