pika
pika copied to clipboard
pika-org
Reduce debug bloat
This is not ok:
> ls -lh test/unit/eigensolver/test_tridiag_solver_full
-rwxr-xr-x 1 simbergm csstaff 1.5G Sep 6 11:49 test/unit/eigensolver/test_tridiag_solver_full
Is the debug bloat coming from pika?
Yes, at least to some extent. Here are the 50 longest strings in the test executable from https://github.com/eth-cscs/DLA-Future/pull/617: bloat.txt.
A lot of it is simply nested senders and receivers. Some of it seems to be boost's small vector. Some of it is possibly unnecessary lambdas. There's probably more.
A slightly more useful namespace change than the namespace detail to namespace impl change that I previously suggested is to collapse sender adaptor implementation namespaces to the top-level namespace. E.g.
namespace pika::execution::experimental::when_all_detail`
to
namespace pika::when_all_detail`
or at the very least
namespace pika::execution::when_all_detail`
. I did this change on the tridiagonal solver test in DLA-Future and saw a roughly 1% reduction in debug bloat. We have more than just one sender adaptor so I would expect this change to reduce the size by up to about 10%. The change is also very localized so it's fairly simple to do. It's a bit of a dumb change, but I'll take 10%.
Other changes will hopefully let us shave a bit more off the the bloat.
#446 is a start. I would re-evaluate this on the pika side once we've done #293. I think there's some room for improvement also on the DLA-Future side.
Some more work has been done in https://github.com/eth-cscs/DLA-Future/pull/674.