Author malte
Recipients erez, malte, silvia
Date 2011-01-23.11:29:53
It's true that the LM graph storage is not particularly efficient, but that
cannot explain memory usage in the gigabytes. In fact, I'd be surprised if the
landmark graph contributed even a single megabyte here. We should really
optimize where it matters first, and that means reducing *per-state* memory
cost, not per-problem memory cost. That means we can safely ignore 1) and 2),
although 3) is indeed an issue.

Also, going from vector<int> to vector<state_t> etc. on vectors that only
usually have one or two elements as in point 1) saves next to nothing (or even
exactly nothing due to memory alignment issues); the vector and dynamic-memory
management overhead dwarfs everything else here. Feel free to try the
optimizations in 1) and 2) out, since seeing is believing :-).

Note that so far we have no evidence that this memory explosion only happens for
the landmark configurations; they might happen everywhere. Landmark
configurations were the only ones I had the chance to look at so far.
Date User Action Args
2011-01-23 11:29:53maltesetmessageid: <>
2011-01-23 11:29:53maltesetrecipients: + malte, erez, silvia
2011-01-23 11:29:53maltelinkissue213 messages
2011-01-23 11:29:53maltecreate