Boinc Manager (Linux) shows Progress, Elapsed Time and Remaining Time for each active task. Currently all three values do not reflect what is actually happening.

eg, Progress jumps from <20% to 100%, Elapsed time drops considerably at task completion and Remaining time is very unreliable.

The calculation of these values deserves another look.

this is somewhat expected. and I'm not sure the jumping to 100% (and likely time estimate) can be fixed in a meaningful way since that would rely on knowing how computationally difficult every task is before hand, and we just dont know.

the reason this is happening is because there is an early abort of tasks when object count exceeds 2 million. this is the count of objects >10cm and when the counts get that high it's because a cascade has happened and caused the object count to explode like that. capping at 2 million was a resource guard, computational difficulty is directly proportional to how many objects there are, and as the count grows, the steps take longer and longer. there's no need to keep crunching and wasting more resources when the result was already clear for that scenario so it's not really necessary to keep the sim running the full 200 years at that point. especially for CPUs.

unfortunately there's no way to know which scenarios will be quick (low object counts for 200 years, or counts that explode and reach 2M very quickly) and which scenarios will be slow (large object counts below the cap that run the full 200) without running it to find out. one of the goals of this project is to help answer that question more specifically than what existing research shows.

The complexity is understood. Goal posts on 3 different fields and you can't pick a set until the game is over. Perhaps once calculation has started it might be possible to predict the most likely outcome and re-evaluate each decade.

For the Time estimates, is gpu concurrency factored in? I'm using 4x to 8x on my hosts and short duration (<60s) tasks were the basis for the previous observations. Most tasks today are in the 4-10min range and the reported times seem more reasonable.

the boinc client does "learn" over time. but it will shift as the runtimes change. i think it does take into account concurrency. but i havent looked into the specifics, just the behavior i've observed before.