Partitions implementation in reconverse

[ritvikrao]

Implements support for partitions (multiple instances of Charm++ running on the same communication and machine layer), including command-line definitions of partitions and messaging functions between partitions. Does not include topological-aware partitions.

[adityapb]

Where is the CMK_HAS_PARTITION defined?

[ritvikrao]

I just pushed a commit to add the definition in charm-config.h. It's really only needed in the charm build so I put it there.

[ericjbohm]

Looks like it should work. Has it been tested with anything that actually uses partitions?

[ritvikrao]

I am trying to run the partition test located in charm/tests/charm++/partitions, but the example command doesn't seem to make sense to me. It is ./hello +pe 4 10 2 +partitions 2. But that command is inherently on 1 node, so how can you request 2 partitions? As a result I get this error
------- Processor 0 Exiting: Called CmiAbort ------ Reason: Number of partitions does not evenly divide number of processes. Aborting

[ericjbohm]

I am trying to run the partition test located in charm/tests/charm++/partitions, but the example command doesn't seem to make sense to me. It is ./hello +pe 4 10 2 +partitions 2. But that command is inherently on 1 node, so how can you request 2 partitions? As a result I get this error ------- Processor 0 Exiting: Called CmiAbort ------ Reason: Number of partitions does not evenly divide number of processes. Aborting

It needs to be able to slice along process boundaries. So, if that was a non-smp build, that command line would work. You would need a more carefully constructed line with +p and +ppn to do the same in an smp build. It wouldn't work at all in multicore.

[ritvikrao]

That test program requires creating P/2 partitions if you have P PEs. I ran the test with 2 PEs on 1 process in standalone mode, and with 4 PEs on 2 processes with LCI, and both worked.

[ericjbohm]

Looks good.