Changes between Initial Version and Version 1 of Ticket #18987, comment 18


Ignore:
Timestamp:
08/09/15 20:56:04 (4 years ago)
Author:
slabbe
Comment:

Legend:

Unmodified
Added
Removed
Modified
  • Ticket #18987, comment 18

    initial v1  
    1 Replying to [comment:15 vdelecroix]:
    2 > Replying to [comment:13 slabbe]:
    3 > > > Why? If your parallization ends with a time better than `total_time / nb_cpus` then you should parallelize more often ;-)
    4 > >
    5 > > I have 240 subproblems each of them taking between 20 minutes and 10 hours of computation. But my machine at work only have 4 cores. So one way or the other, the computation takes days to finish since I do not have access to a super machine.
    6 >
    7 > This looks very bad. At the end, you might end up with only one core working on the biggest subinstance. And it can lasts several days even with 200 cores. Ideally, you should slice the problem in such way that each subinstance will not take longer than 1 hour (let say). This is why adopting a less naive strategy at the level of dancing links seems to me the best option since people already worked on it.
     1> This looks very bad. At the end, you might end up with only one core working on the biggest subinstance. And it can lasts several days even with 200 cores. Ideally, you should slice the problem in such way that each subinstance will not take longer than 1 hour (let say).
    82
    93It is not very bad as most of the 240 computations takes the same amount of time (about 2 to 3 hours). Maybe 5 of time takes more (10 hours). So I am using the four cores at least 95% of the time. In my case, I have very good subinstances to reuse your term.