When I compare the "number of events processed" on the
status bar with the number of method calls that I specified
to trigger a detach, I see far fewer events than I expect.
Namely, at least 2 events per method call (entry/exit).
I have 2 test cases where case 1 yields about 2 events
per method call, and case 2 yields 0.003. The details:
Here's my setup:
- TPTP 4.2
- Analysis types selected:
- Profile execution flow (show execution flow graphical
details, boundary class depth=1 or 111, see below)
- Memory allocation.
This is my test (the exact test case is not relevant,
the idea is to test TPTP under heavy loads. Profiling TPTP
itself serves that purpose):
- Profile-launching of an Eclipse Application
- org.eclipse.tptp.* and org.eclipse.hyades.* included
- "show execution flow graphical details" is checked
- Boundary class depth set to 1 or 111 (see below)
- Profiling the tutorial example from the launched Eclipse
Application (so we're "profiling the profiler").
Here are some test cases and resulting numbers:
-----------------------------------------------------------
Case 1:
Boundary class depth=1, number of method calls=20,000:
- 42,000 processed events
- 5 button clicks of tutorial example before auto-detach
(about 8,000 events per click)
Case 2:
Boundary class depth=111, number of method calls=50,000,000:
- 150,000 processed events
- 4 to 5 button clicks of tutorial example before auto-detach
(about 35,000 events per click)
-----------------------------------------------------------
The processed events were read from the status bar.
>From these data, we get
- Number of events per method call:
- Case 1: about 2
- Case 2: about 0.003
There should be at least 2 events per method call (assuming
that all method calls are treated the same way). Case 1 is
almost normal (I would expect more events because of object
allocations) but case 2 is way below normal.
The only plausible explanation that I can think of is CPU
starvation under heavy loads. Is it possible that the agent
controller breaks down under heavy loads and starts skipping
events? Is there a way to detect if any events were missed?
Or am I simply misinterpreting the data?
Henk Aling
Wind River
_______________________________________________
tptp-tracing-profiling-tools-dev mailing list
tptp-tracing-profiling-tools-dev@eclipse.org
https://dev.eclipse.org/mailman/listinfo/tptp-tracing-profiling-tools-dev
