TODO

This file records things that should be done to improve the tutorial
presentation. Remove them from this lest when you apply the changes to
the content.

From Phil at ICS 2011:

We used the slide deck from the similar-length tutorial session from
the (2011) workshop. The timing was OK but not great, with the
pre-intermission segment fitting in the 2-hour session before the
break, and the post-intermission segment wrapping up somewhat early in
the remaining 1.5 hours. We then switched to discussing broader
questions, returning to some earlier discussions, and talking about
prospective applications.

The selection of content was mostly good, but a bit weird in places.

- The deep dive on migratable threads might be sensible with a (long)
discussion of AMPI, but that was explicitly out of scope so the
threads discussion didn't really fit in.

- One missing piece that I'd like to see is an early slide or two on
the principles and philosophy of Charm++ programming: virtualization,
overdecomposition, migratability, and the various things that are
thereby intrinsic or extrinsic to application logic (respectively:
decomposition, communication, control flow; mapping, load balancing,
fault tolerance).

- There were some points where the slides were not in a good order
pedagogically, with the content creating forward references in the
lecture, or leading the students to ask questions whose answers were
forward references. This breaks up the natural flow of the lectures,
and sometimes led to confusion.

Here are a few points that I recall from looking over the slides now:
0. [name of Blue Waters/BGQ communication API] is mentioned on the
portability slides, while even that is apparently still NDA. LAPI is
also mentioned in reference to the Blue Waters software stack.
1. Chare array broadcasts should be discussed after the Hello World
example showing point-to-point
2. Prioritized execution should mention *critical paths* - the big use
in NAMD, OpenAtom, ChaNGa
3. The detail of the runtime queue being implemented in three chunks
is utterly irrelevant detail. It's a priority queue, period.
4. Initcall is deprecated in favor of initnode/initproc (per chamxi's
own output), so shouldn't be mentioned. If it's still used in NAMD, we
should fix that.
5. Exclusive entry methods refer to node groups, which are discussed
afterward. Perhaps those should be reordered.
6. The definitions of group and nodegroup examples inherit from Group
and NodeGroup, respectively. For consistency, we should always use
CBase_foo. This also means that if a group is changed to or from an
array, that part of its definition need not change.
7. When mentioning fault tolerance, emphasize that message delivery is
always reliable at the application layer. Developers don't need to
follow the crazy "resilient MPI programming" movement that's starting.
This should be explicit and clear in papers and presentations on
Charm++ FT as well.