Lines of Code
[root]/Packages/Uintah/CCA/Components/Schedulers
| Author | Changes | Lines of Code | Lines per Change |
|---|---|---|---|
| Totals | 130 (100.0%) | 1838 (100.0%) | 14.1 |
| luitjens | 92 (70.8%) | 1422 (77.4%) | 15.4 |
| dav | 20 (15.4%) | 384 (20.9%) | 19.2 |
| jas | 12 (9.2%) | 17 (0.9%) | 1.4 |
| worthen | 5 (3.8%) | 15 (0.8%) | 3.0 |
| harman | 1 (0.8%) | 0 (0.0%) | 0.0 |
Readded code that allows particle sets to exist in the newDW pre-relocate.
0 lines of code changed in 2 files:
undid 42618 this appears to be causing some problems and will need further scruitiny.
2 lines of code changed in 1 file:
Allow particle sets to exist prior to relocate. Previously particle sets did not exist for preRelocate variables in the newDW until after the relocate code was called. This causes unknown variable errors.
2 lines of code changed in 1 file:
Disabled some code which is no longer needed
3 lines of code changed in 1 file:
Added a count variable to the PSPatchMatlGhost structure. This variable is used to record how many times a particle set has been added/erased in the process of combining. Doing this allows us to avoid deleting particle sets that are still needed and will hopefully resolve the issues i've been seeing reguarding particle sendsets.
50 lines of code changed in 1 file:
Fix another particle sendset bug.
The root of these problems lie in the fact that dependencies are scheduled per variable and sendsets are created once for all variables. If you have tasks that share some variables but not all dependencies can combine in diffent ways. This causes sendsets to be deleted even though they are still needed by other dependencies.
The current fix for this is to ensure that dependencies combine consistantly. The current logic for combining dependences is the following:
First if a new dependency is completly included in another dependency then combine with the first one of those you find. If no dependencies completly include the new dependency then combine it with the first dependency that can be extended to include this dependency without increasing the amount of communication. If this cannot be done then create a new dependency.
This seems to fix the sendset!=old_dw assert (for now).
24 lines of code changed in 1 file:
Output the amount of compression used in the sus spew.
0 lines of code changed in 2 files:
Fix for memory leaks
6 lines of code changed in 1 file:
Do not pack/unpack zero sized messages. Packing 0 sized messages increases the size of the send buffers. This should reduce communication.
Do not send 0 sized messages. There is no reason to send a message that has no data. Doing so only creates a syncronization point.
Since 0 sized messages are no longer being sent I have undone the zero sized compression check.
48 lines of code changed in 1 file:
Enabled gzip message compression for messages sent via the task graph.
By default this is disabled until some large scale testing and parameter studies can be performed.
To enable add this to your input file:
<Scheduler>
<compressionLevel>5</compressionLevel>
<compressionThreshold>50</compressionThreshold>
</Scheduler>
compressionLevel specifies the gzip compression level which ranges from 0 (no compression) to 9 (maximum compression)
compressionThreshold specifies the threshold for message size in bytes in which message compression will be attempted. That is any messages that are of size <=compressionThreshold will not be compressed.
22 lines of code changed in 5 files:
Added the mpi status to donesome.
9 lines of code changed in 2 files:
Removed some unnessary code.
1 lines of code changed in 1 file:
Tasks that require data from a fine level are now profiled sperately from normal tasks. This will hopefully allow the profiling to be more accurate.
1 lines of code changed in 3 files:
get method for the profile type.
1 lines of code changed in 1 file:
Added the profile type to DetailedTasks. The profile type can be either Normal or Fine.
Normal means that that the task executes proportional to the number of cells/particles in the patch.
Fine means that the task executes proportional to the number of cells/particles on the fine level below it (for example coarsening).
10 lines of code changed in 2 files:
Pass the task into the profiler instead of the patches that the task ran on. This will be used to profile the coarse interfaces seperately.
1 lines of code changed in 1 file:
Added a SCI_DEBUG called MPIStats. This debug will printout the number of messages that the task graph sends and the total message volume.
45 lines of code changed in 3 files:
Reworked the dependency combining code. I'm unsure the current code will work for all cases. I have now changed the dependency combing code to hopefully always consistently combine and be quicker.
44 lines of code changed in 2 files:
pruned some unused code.
0 lines of code changed in 1 file:
Fixed a bug where sendstates and dependencies could become inconsistent causing all kinds of bad and semi random things to happen
27 lines of code changed in 1 file:
(46 more)