-* segfault with ./loba cluster1000.xml -N64 -a fairstrategy
+* In run-all:
+ - Fix compression of output files (make it optional?).
+ - Use more legible names for generated files.
+ Current problems are:
+ . file names are too long.
+ . MORE_ARGS do not appear in file name.
+ Why not simply using file names "plain.out", and "bookkeeping.out"?
- this is a deadlock occuring when:
- - a process is in the finalize stage;
- - all processes but one are blocked on receive;
- - the process that is still running owns all the remaining load,
- and sends it all to the finalizing process, and then goes in
- blocking receive.
- The finalizing process receives the load, and blocks again,
- waiting for a close message.
- All processes are then blocked, and non-one is able to see that
- there is no more load in the system!
+* Support heterogeneous platforms?
+ Not sure yet.
+ Should be doable if each process also sends its speed to its neighbors.
-* verify bookkeeping version.
+* Add options -j/-J: minimum number of iterations?
-* add several metrics
- - flops (total, avg, stddev)
- - message exchanges (number, volume)
+* Add options -f/-F: termination based on flops achieved?
-* add options -j/-J : minimum number of iterations ?
+* Reindent source code.
+ Mainly loba_{bulk,lln}.cpp.
-* add a variant to (not) change neighbor load information at send.
+* Choose a license before making the code public.
+ CeCILL-B?
-* implement loba_* algorithms (start with some trivial one)
+* Translate french parts in english (README, QUESTIONS, ALGORITHMS, ...)
+ Not urgent.
-* for automatic process topology,
- -> implement some random initial distribution of load
+* Add synchronized mode?
+ No.
-* add synchronized mode
-
-* translate README file?
+* Add a variant to (not) change neighbor load information at send?
+ No.
