Space recycle

Do the recycling not cell-by-cell, but with lists of cells at a time.

Added 1 commit:

• 75a9d6f7 - Revert "move calls to MPI_Irecv to the enqueue_mapper."

Added 49 commits:

• 75a9d6f7...290e85f7 - 48 commits from branch master
• 981be8bc - Merge branch 'master' into space_recycle

Removed the unwanted commit from the MPI tests and merged from master, so this should be good for testing now.

So we have a leak of c->sort somewhere. I'd guess it is because space_rebuild_recycle used to call space_recycle which freed this. That make sense to you?

OK, I'll run this with ASAN and look into it. Thanks for catching this!

Added 3 commits:

• 6979ad47 - don't mess-up the list by clearing linked cells.
• 9dcdeb4a - no need to clean up linked cells as they are bzeroed before being returned in space_getcell.
• 47ed85d1 - properly destroy glocks as well.

PTAL, now only "leaks" as much as master

Status changed to merged

mentioned in commit 1ce1015d

Thanks and whoops. Didn't mean to merge just yet. Thick fingers and dodgy touchpad... I'll leave it unreverted for now and test some more.

Passed any tests, so will accept properly now. @jwillis might like to check the scaling. Don't see any improvements in my simple tests.

Doing the tcmalloc scaling at the moment, but I will do this one next.

The dark blue is the new master, it performs better than the old master:

Thanks.