Charm++ parallel programming system

[Abhinav Bhatele <bhatele AT illinoisalumni.org>]

Work that is not instrumented by the tracing framework shows up as empty spaces (default color: black) in the timeline. Idle time shows up as a thin bar (default: white) - this happens when the CPU is idle.

It is not easy to automatically find the reasons for idle time. It could be due to:

- load imbalance: this CPU is waiting for others to finish their time step

- communication delays: this delay could again be due to load imbalance or congestion on the network

Typically, you look at the interesting CPUs and do some sort of critical path analysis (manually in the timeline view) to see what is every one waiting for.

- Abhinav

On Wed, Jan 6, 2016 at 4:19 PM, Ronak Buch <ronakbuch AT gmail.com> wrote:

Hi Jozsef,

As described in the manual (http://charm.cs.illinois.edu/manuals/html/projections/manual-1p.html), idle time is time that the scheduler has nothing to run.  There is also overhead, which is time spent in uninstrumented code, such as the runtime system.

A user event should show up in the timeline graph by default.  Double check, if you could, that the appropriate boxes are ticked in the View -> Show & Hide User Events menu item.

Thanks,

Ronak

On Wed, Jan 6, 2016 at 4:56 PM, Jozsef Bakosi <jbakosi AT gmail.com> wrote:

Hi folks,

If I understand correctly, the projections tool collects information on entry methods only by default. If that is the case, how can I add timing of a non-entry method?

Also, what is the definition of "idle time" and how can I find out the reason for it? Does idle time include simply a CPU waiting for work or does it include some work that is not timed, i.e., not in the projections database?

I have added a user event for a section that I *think* ought to show up on the timeline but I don't see it on the timeline graph, only on the user events list.

Can someone please explain?

Thanks in advance,

Jozsef

--

Abhinav Bhatele, people.llnl.gov/bhatele
Center for Applied Scientific Computing, Lawrence Livermore National Laboratory