charm.git
17 years agobugfix
Filippo Gioachin [Tue, 10 Aug 2004 19:52:58 +0000 (19:52 +0000)]
bugfix

17 years agoSeries of bug fixes:
Orion Lawlor [Tue, 10 Aug 2004 19:23:42 +0000 (19:23 +0000)]
Series of bug fixes:
  1.) MPI_Isend with MPI_COMM_SELF didn't work, but MPI_Send did.
   Why?  Because MPI_Send has a special case for MPI_COMM_SELF
   missing from MPI_Isend (and MPI_Scan, and possibly others).
   Fixed by making ampi::getRank()/getSize() take a communicator,
   so the special case goes in one place instead of scattered through
   the source.

  2.) Make all the arithmetic reduction operations work with all the MPI datatypes.
   Removed horrible copy-and-paste code (at least from arithmetic ops).

  3.) Added (disabled) tracing statements, used to track down the above
   problems.

17 years agoUseless changes to support inheritance.
Eric Bohm [Tue, 10 Aug 2004 18:57:24 +0000 (18:57 +0000)]
Useless changes to support inheritance.
Checking them in because they'll come in handy later on when we figure out
the deeper changes required for inheritance.

These should make no difference to any existing code.  And might make multiple inheritance from non-posers work, but that capacity is untested.

17 years agoAdded sequential endtime approximation.
Terry L. Wilmarth [Tue, 10 Aug 2004 18:43:04 +0000 (18:43 +0000)]
Added sequential endtime approximation.

17 years agoFixed a bug in srtable:Restructure. Removed debug messages.
Terry L. Wilmarth [Tue, 10 Aug 2004 18:31:00 +0000 (18:31 +0000)]
Fixed a bug in srtable:Restructure.  Removed debug messages.

17 years agoAdding out of bounds check for CkVec. This is only enabled when CMK_PARANOID is tuned on.
Sameer Kumar [Tue, 10 Aug 2004 17:47:20 +0000 (17:47 +0000)]
Adding out of bounds check for CkVec. This is only enabled when CMK_PARANOID is tuned on.

17 years agoTwo bug fixes:
Orion Lawlor [Tue, 10 Aug 2004 17:20:25 +0000 (17:20 +0000)]
Two bug fixes:
  1.) Use CMK_LD to test shared library creation.  This is what
      charmc does, and with icc on the SuSE machines, CMK_CXX
      doesn't work (missing link options) but CMK_LD does.

  2.) Remove garbage (chunk of Makefile!) at end of configure.in.
     It's not clear if Chao checked this in by accident, or if
     CVS or the filesystem somehow screwed it up.

17 years ago- Moved 4 files (ComlibStrategy.*, ComlibArrayListener.*) from ck-core to the
Filippo Gioachin [Tue, 10 Aug 2004 07:20:14 +0000 (07:20 +0000)]
- Moved 4 files (ComlibStrategy.*, ComlibArrayListener.*) from ck-core to the
better place ck-com.
- Updated the benchmark for pipeBroadcast to test more strategies simulataneously
- Fixed a bug in comlib for which the variable "myInstanceID" did not get set
correctly when a Charm strategy embedded a Converse strategy in it (the converse
strategy variable was not set)
- There is still a TODO: usually Comlib remember the last strategy used (and set
through cinst.beginIteration()), but not at the very beginning of the program,
because when the StrategyTable is unPupped, it overwrite the currStrategy index
to the last inserted strategy.

17 years agodefined the missing CMK_MOD_EXT definition.
Gengbin Zheng [Mon, 9 Aug 2004 21:40:13 +0000 (21:40 +0000)]
defined the missing CMK_MOD_EXT definition.

17 years agoFix silly IDXL registration bug.
Orion Lawlor [Mon, 9 Aug 2004 17:02:18 +0000 (17:02 +0000)]
Fix silly IDXL registration bug.
Previously, would re-register after migration
*always*, even if no previous value existed.

IDXL registration is needed so that users can keep
IDXL list indices even across migration, like:
int list=FEM_Comm_ghost(mesh,FEM_ELEM);
        FEM_Migate();
        IDXL_...(list,...);
To do this, FEM needs to re-register its lists
in the same order on the new processor.

17 years agodeleted an unused (and misleading) constructor
Filippo Gioachin [Sat, 7 Aug 2004 04:49:58 +0000 (04:49 +0000)]
deleted an unused (and misleading) constructor

17 years agofixed a wrong invocation of CmiRegisterHandler
Filippo Gioachin [Sat, 7 Aug 2004 04:48:25 +0000 (04:48 +0000)]
fixed a wrong invocation of CmiRegisterHandler

17 years agoChecking in some garbage because I need to debug elsewhere...
Terry L. Wilmarth [Fri, 6 Aug 2004 22:31:45 +0000 (22:31 +0000)]
Checking in some garbage because I need to debug elsewhere...

17 years agoSeveral optimizations to speed up srTable.
Terry L. Wilmarth [Fri, 6 Aug 2004 16:07:13 +0000 (16:07 +0000)]
Several optimizations to speed up srTable.

17 years agoCompleted adapt3 strategy improves on adapt2.
Terry L. Wilmarth [Thu, 5 Aug 2004 22:25:02 +0000 (22:25 +0000)]
Completed adapt3 strategy improves on adapt2.

17 years agoFixed make test line
Eric Bohm [Thu, 5 Aug 2004 17:10:41 +0000 (17:10 +0000)]
Fixed make test line

17 years agosnippet about -s flag
Eric Bohm [Thu, 5 Aug 2004 17:05:28 +0000 (17:05 +0000)]
snippet about -s flag

17 years agoUse -s option to etrans.pl to get sequential support.
Eric Bohm [Thu, 5 Aug 2004 17:03:09 +0000 (17:03 +0000)]
Use -s option to etrans.pl to get sequential support.
Note, you still have to define SEQUENTIAL_POSE in your pose_config

17 years agoReorganized code.
Yan Shi [Thu, 5 Aug 2004 14:40:10 +0000 (14:40 +0000)]
Reorganized code.
Fixed optimization of zgemm routine, and streaming at first phase.

17 years agoRearrange IBM SP charm versions:
Orion Lawlor [Thu, 5 Aug 2004 00:22:22 +0000 (00:22 +0000)]
Rearrange IBM SP charm versions:
   mpi-sp mpcc32 is always 32 bit
   mpi-sp mpcc64 is always 64 bit
   mpi-sp (no options) uses OBJECT_MODE to decide on the number of bits.

I think this is what most people will want: those who don't care
about the bit setting will use the default OBJECT_MODE in their dot files;
while those who do care have a sensible override.

17 years agoAdded integrated sequential simulation to POSE.
Terry L. Wilmarth [Wed, 4 Aug 2004 21:32:25 +0000 (21:32 +0000)]
Added integrated sequential simulation to POSE.

17 years agoAdapt3 updates, backwards commits.
Terry L. Wilmarth [Wed, 4 Aug 2004 17:17:12 +0000 (17:17 +0000)]
Adapt3 updates, backwards commits.

17 years agoforgot the change in translator - remove unused setImmediate().
Gengbin Zheng [Wed, 4 Aug 2004 14:37:22 +0000 (14:37 +0000)]
forgot the change in translator - remove unused setImmediate().

17 years agobugfixes for pipelined broadcast in converse
Filippo Gioachin [Tue, 3 Aug 2004 20:41:31 +0000 (20:41 +0000)]
bugfixes for pipelined broadcast in converse

17 years agovirtualizing a function
Filippo Gioachin [Tue, 3 Aug 2004 20:38:16 +0000 (20:38 +0000)]
virtualizing a function

17 years agouse special bg trace macro for mpi_barrier so that the simulator can identify the...
Gengbin Zheng [Tue, 3 Aug 2004 17:11:42 +0000 (17:11 +0000)]
use special bg trace macro for mpi_barrier so that the simulator can identify the barrier and swap it with its own machine implementation.

17 years agodefine a special trace for MPI_Barrier so that it potentially allows a simulator...
Gengbin Zheng [Tue, 3 Aug 2004 17:08:00 +0000 (17:08 +0000)]
define a special trace for MPI_Barrier so that it potentially allows a simulator to swap the implementation of barrier instead of using Charm++'s barrier implementation.

17 years agoremoved unused (obsolete) code about setting immediate bit in charm envelop.
Gengbin Zheng [Tue, 3 Aug 2004 16:43:05 +0000 (16:43 +0000)]
removed unused (obsolete) code about setting immediate bit in charm envelop.

17 years agoan optimized version to substantially reduced bigsim overhead of unnecessarily access...
Gengbin Zheng [Tue, 3 Aug 2004 16:41:10 +0000 (16:41 +0000)]
an optimized version to substantially reduced bigsim overhead of unnecessarily accessing thread private varaibles.

17 years agoupdated loadlog utility program.
Gengbin Zheng [Tue, 3 Aug 2004 16:34:59 +0000 (16:34 +0000)]
updated loadlog utility program.

17 years agoSimplified and accellerated Mgc volume computation routines,
Orion Lawlor [Tue, 3 Aug 2004 02:20:00 +0000 (02:20 +0000)]
Simplified and accellerated Mgc volume computation routines,
and added check against volume_planes.h routines.

17 years agoAdd volume computation routines for arbitrary sets of
Orion Lawlor [Tue, 3 Aug 2004 02:18:59 +0000 (02:18 +0000)]
Add volume computation routines for arbitrary sets of
halfspaces.  This is GPL code from the package "VINCI",
   http://www.lix.polytechnique.fr/Labo/Andreas.Enge/Vinci.html
although I've heavily modified it for speed (in 3D) and
portability.

To avoid running afoul of the GPL, this source file is not
built or linked in by default in Charm++.  If these methods
prove useful (and for now they're still to slow to compete
with the Mgc routines), we'll have to write a non-GPL version.

17 years agohardcoded -q32 to force 32 bit compilation.
Gengbin Zheng [Mon, 2 Aug 2004 18:18:08 +0000 (18:18 +0000)]
hardcoded -q32 to force 32 bit compilation.

17 years agoBug fix:
Orion Lawlor [Mon, 2 Aug 2004 17:10:02 +0000 (17:10 +0000)]
Bug fix:
   Really blow away charm/bin, lib, include directories, even
if they're not just softlinks.  This lets ./build work properly
even if somebody does "cp -r" on the charm source tree--without
this change, the build would succeed, but charm/bin wouldn't change,
leading to horribly confusing errors.

17 years agoSome optimization breaks running on IBM SP and BG/L - Wrong answers are given and...
Yan Shi [Sat, 31 Jul 2004 19:07:04 +0000 (19:07 +0000)]
Some optimization breaks running on IBM SP and BG/L  -  Wrong answers are given and causes NANs. So have to convert back to a non-opt version just to run.

17 years agoAdded subsection on etrans describing multiple input C file support.
Eric Bohm [Wed, 28 Jul 2004 15:21:41 +0000 (15:21 +0000)]
Added subsection on etrans describing multiple input C file support.

17 years agoReplace clash-prone macro names MAX, MIN, ABS, and LEASH with
Orion Lawlor [Wed, 28 Jul 2004 01:38:29 +0000 (01:38 +0000)]
Replace clash-prone macro names MAX, MIN, ABS, and LEASH with
BG_MAX, BG_MIN, etc.  According to Jim, MAX and MIN are already
defined in the net-ppc-darwin-xlC system headers.

17 years agoCommit "feature" fixed: commits all executed events at end of program.
Terry L. Wilmarth [Tue, 27 Jul 2004 21:50:21 +0000 (21:50 +0000)]
Commit "feature" fixed: commits all executed events at end of program.

17 years agoSupports multiple C files.
Eric Bohm [Tue, 27 Jul 2004 18:53:34 +0000 (18:53 +0000)]
Supports multiple C files.
Simply add them to the etrans.pl line after the module to which
they belong.

They will be translated and the output appended to the modulename_sim.C
file.  So its a primitive and collision prone support for multiple files.
But it will at least allow some logical breakup of big C files.

17 years agoAdded in header include for fync function.
Yan Shi [Tue, 27 Jul 2004 13:59:14 +0000 (13:59 +0000)]
Added in header include for fync function.

17 years agoAdd in #ifdef to avoid fortran convention conflict on underscore after.
Yan Shi [Tue, 27 Jul 2004 13:37:24 +0000 (13:37 +0000)]
Add in #ifdef to avoid fortran convention conflict on underscore after.

17 years agoAdapt3 continues: code to find largest timestamp in eq, minor adjustments to safeTime...
Terry L. Wilmarth [Tue, 27 Jul 2004 05:17:10 +0000 (05:17 +0000)]
Adapt3 continues: code to find largest timestamp in eq, minor adjustments to safeTime and commit.

17 years agotemp output added.
Terry L. Wilmarth [Tue, 27 Jul 2004 00:04:00 +0000 (00:04 +0000)]
temp output added.

17 years agoDetect if an empty communicator(local and remote) is to be created,
Chao Huang [Mon, 26 Jul 2004 22:55:38 +0000 (22:55 +0000)]
Detect if an empty communicator(local and remote) is to be created,
if so, complain and abort

17 years agoAdded rudimentary support for multiple C files in a module.
Eric Bohm [Fri, 23 Jul 2004 17:36:29 +0000 (17:36 +0000)]
Added rudimentary support for multiple C files in a module.

17 years agoMore changes to adapt3. Not working yet...
Terry L. Wilmarth [Fri, 23 Jul 2004 16:44:20 +0000 (16:44 +0000)]
More changes to adapt3.  Not working yet...

17 years agoAdd support for floating-point pixel images, as requested
Orion Lawlor [Thu, 22 Jul 2004 22:20:43 +0000 (22:20 +0000)]
Add support for floating-point pixel images, as requested
by the Cosmo folks.  Also took opportunity to lightly refactor
and heavily comment ImageData.[Ch].

17 years agoUse FEM_COORD to store node coordinates, instead of FEM_DATA.
Orion Lawlor [Thu, 22 Jul 2004 22:19:27 +0000 (22:19 +0000)]
Use FEM_COORD to store node coordinates, instead of FEM_DATA.
This is the new standard place to store node coordinates.

17 years agoUse CkVec to store nodes and elements of tet mesh.
Orion Lawlor [Thu, 22 Jul 2004 22:18:58 +0000 (22:18 +0000)]
Use CkVec to store nodes and elements of tet mesh.
This lets us easily add nodes and elements, which is
needed, e.g., to extrude out boundary faces.

17 years agoBug fix: don't "compensate" for zero volume by dividing by it!
Orion Lawlor [Thu, 22 Jul 2004 22:16:31 +0000 (22:16 +0000)]
Bug fix: don't "compensate" for zero volume by dividing by it!
This fix makes zero-volume transfers come out zero, not nan.

17 years agoMake pup routines #ifdef, to allow CkVector3d to be used
Orion Lawlor [Thu, 22 Jul 2004 22:15:28 +0000 (22:15 +0000)]
Make pup routines #ifdef, to allow CkVector3d to be used
outside of Charm++ (e.g., in liveViz3d client).

17 years agoAdded "const" getData accessor to CkReductionMsg.
Orion Lawlor [Thu, 22 Jul 2004 22:10:43 +0000 (22:10 +0000)]
Added "const" getData accessor to CkReductionMsg.
This lets you deal with "const CkReductionMsg *"'s sensibly.

17 years agoFixed a tiny bug in timing.
Terry L. Wilmarth [Thu, 22 Jul 2004 21:39:00 +0000 (21:39 +0000)]
Fixed a tiny bug in timing.

17 years agoAdded extensive, Doxygen-friendly comments.
Orion Lawlor [Thu, 22 Jul 2004 21:35:09 +0000 (21:35 +0000)]
Added extensive, Doxygen-friendly comments.

17 years agoChanges to convert to POSE_TimeType in generated code
Eric Bohm [Thu, 22 Jul 2004 21:21:18 +0000 (21:21 +0000)]
Changes to convert to POSE_TimeType in generated code

17 years agoAdd __cmi_assert, which fixes a link error on MPI-SP.
Orion Lawlor [Thu, 22 Jul 2004 16:36:33 +0000 (16:36 +0000)]
Add __cmi_assert, which fixes a link error on MPI-SP.

17 years agoRemoved // comments: they don't work with many C compilers.
Orion Lawlor [Thu, 22 Jul 2004 16:01:05 +0000 (16:01 +0000)]
Removed // comments: they don't work with many C compilers.

17 years agoHeader definition for CK_QUEUEING_LFIFO for LFIFO support.
Eric Bohm [Thu, 22 Jul 2004 15:16:10 +0000 (15:16 +0000)]
Header definition for CK_QUEUEING_LFIFO for LFIFO support.

17 years agoMaking sure adapt2 is the default strategy for ASIM.
Terry L. Wilmarth [Thu, 22 Jul 2004 15:13:51 +0000 (15:13 +0000)]
Making sure adapt2 is the default strategy for ASIM.

17 years agoset CMK_QT to aix to resolve missing alloca.
Gengbin Zheng [Thu, 22 Jul 2004 14:24:36 +0000 (14:24 +0000)]
set CMK_QT to aix to resolve missing alloca.

17 years agosuggested by IBM of using CMK_NO_OUTSTANDING_SENDS on bgl
Gengbin Zheng [Thu, 22 Jul 2004 14:21:48 +0000 (14:21 +0000)]
suggested by IBM of using CMK_NO_OUTSTANDING_SENDS on bgl

17 years agoconfigure for using native bgl compilers instead of cross compiler (provided by Bob...
Gengbin Zheng [Thu, 22 Jul 2004 14:20:59 +0000 (14:20 +0000)]
configure for using native bgl compilers instead of cross compiler (provided by Bob Walkup at IBM)

17 years agorenamed Makefile_lb to Make.lb.
Gengbin Zheng [Thu, 22 Jul 2004 03:32:16 +0000 (03:32 +0000)]
renamed Makefile_lb to Make.lb.

17 years agorename Makefile_lb to Make.lb
Gengbin Zheng [Thu, 22 Jul 2004 03:31:12 +0000 (03:31 +0000)]
rename Makefile_lb to Make.lb

17 years agorenamed Makefile_lb to Make.lb. This is due to some weird things happened on net...
Gengbin Zheng [Thu, 22 Jul 2004 03:30:54 +0000 (03:30 +0000)]
renamed Makefile_lb to Make.lb. This is due to some weird things happened on net-win32 where it keep trying to regenerate Makefile_lb by cat Makefile_lb.sh to it due to some weird default make rule.

17 years agoExpanded adapt3. Event queue keeps track of largest timestamp of unexecuted events.
Terry L. Wilmarth [Thu, 22 Jul 2004 03:27:13 +0000 (03:27 +0000)]
Expanded adapt3. Event queue keeps track of largest timestamp of unexecuted events.

17 years agoUpdated main to use new initialization method.
Terry L. Wilmarth [Thu, 22 Jul 2004 01:50:20 +0000 (01:50 +0000)]
Updated main to use new initialization method.

17 years agoUpdated adapt3 to use POSE_TimeType.
Terry L. Wilmarth [Thu, 22 Jul 2004 01:47:18 +0000 (01:47 +0000)]
Updated adapt3 to use POSE_TimeType.

17 years agoAdded adapt3; updated initialization and timing method.
Terry L. Wilmarth [Thu, 22 Jul 2004 01:39:22 +0000 (01:39 +0000)]
Added adapt3; updated initialization and timing method.

17 years agosupport for USE_LONG_TIMESTAMPS
Eric Bohm [Wed, 21 Jul 2004 22:43:53 +0000 (22:43 +0000)]
support for USE_LONG_TIMESTAMPS

17 years agoAdded support for 64 bit timestamps.
Eric Bohm [Wed, 21 Jul 2004 22:43:05 +0000 (22:43 +0000)]
Added support for 64 bit timestamps.

Set USE_LONG_TIMESTAMPS =1 in pose_config.h to use.
You must also use POSE_TimeType for any variables containing timestamps in your user code.  It would be advisable to adopt that type even if you don't use long timestamps for readability.  Similarly, I/O routines should use %lld for timestamps under USE_LONG_TIMESTAMPS conditions.

POSE_TimeType will be set correctly based on the value of USE_LONG_TIMESTAMPS.

This feature works fine with ASIM, but has not been exhaustively tested under battle conditions.

Converting all of POSE's timestamps to POSE_TimeType meant touching a LOT of files.  So adopt this set of changes carefully.  Your POSE compile will definitely break if you neglect to merge with the new pose_config.h.

17 years agoAdded CK_QUEUEING_LFIFO and CK_QUEUEING_LLIFO (with corresponding changes
Eric Bohm [Wed, 21 Jul 2004 22:30:19 +0000 (22:30 +0000)]
Added CK_QUEUEING_LFIFO and CK_QUEUEING_LLIFO (with corresponding changes
for CQS_*) to support long (64 bit) signed integer priorities.

Use CmiInt8 for portability rather than long long or long long int.

This addition was motivated by the desire for many billions of timesteps in
POSE based simulations.  Since POSE uses both positive and negative values
BFIFO could not be used and still retain proper signed integer priority
queueing behavior.

megatest's priolongtest does rudimentary validation of this feature.

17 years agofsync is not available in win32, so skip it by forcing #if.
Gengbin Zheng [Wed, 21 Jul 2004 20:21:26 +0000 (20:21 +0000)]
fsync is not available in win32, so skip it by forcing #if.

17 years agooriginally expedited message implies inline, now make it separate. However one can...
Gengbin Zheng [Wed, 21 Jul 2004 20:02:07 +0000 (20:02 +0000)]
originally expedited message implies inline, now make it separate. However one can still send message in both expedited and inline fashion.

17 years agouse new expedited message to speed up multicast and reduction in spanning tree.
Gengbin Zheng [Wed, 21 Jul 2004 19:24:15 +0000 (19:24 +0000)]
use new expedited message to speed up multicast and reduction in spanning tree.

17 years agofixed a minor file dependence.
Gengbin Zheng [Wed, 21 Jul 2004 19:01:59 +0000 (19:01 +0000)]
fixed a minor file dependence.

17 years agofixed double int i declaration.
Gengbin Zheng [Wed, 21 Jul 2004 19:01:10 +0000 (19:01 +0000)]
fixed double int i declaration.

17 years agoput a '-' before romio make rule to ignore error code.
Gengbin Zheng [Wed, 21 Jul 2004 17:47:16 +0000 (17:47 +0000)]
put a '-' before romio make rule to ignore error code.

17 years agoget rid of extern "C" in C code.
Gengbin Zheng [Wed, 21 Jul 2004 16:50:41 +0000 (16:50 +0000)]
get rid of extern "C" in C code.

17 years agochanged to use [notrace]
Gengbin Zheng [Wed, 21 Jul 2004 06:28:07 +0000 (06:28 +0000)]
changed to use [notrace]

17 years agouse [notrace] instead of CkDisableTrace calls.
Gengbin Zheng [Wed, 21 Jul 2004 06:19:38 +0000 (06:19 +0000)]
use [notrace] instead of CkDisableTrace calls.

17 years agoadded [nokeep] and [notrace]
Gengbin Zheng [Wed, 21 Jul 2004 05:37:31 +0000 (05:37 +0000)]
added [nokeep] and [notrace]

17 years agoreverted the previous checkin.
Gengbin Zheng [Wed, 21 Jul 2004 05:18:54 +0000 (05:18 +0000)]
reverted the previous checkin.

17 years agofixed a typo.
Gengbin Zheng [Wed, 21 Jul 2004 05:17:22 +0000 (05:17 +0000)]
fixed a typo.

17 years agoname changes to reflex [expedited].
Gengbin Zheng [Wed, 21 Jul 2004 05:17:03 +0000 (05:17 +0000)]
name changes to reflex [expedited].

17 years agochanged [skipscheduler] to [expedited].
Gengbin Zheng [Wed, 21 Jul 2004 05:14:57 +0000 (05:14 +0000)]
changed [skipscheduler] to [expedited].
added [notrace] attribe to disable tracing of an EP.

17 years agoactivated by "+ringexit", program is shutdown in a token ring fashion to avoid networ...
Gengbin Zheng [Wed, 21 Jul 2004 04:51:07 +0000 (04:51 +0000)]
activated by "+ringexit", program is shutdown in a token ring fashion to avoid network congestion caused by projections log generation at exit.

17 years agowhen closing logs, fsync files to disk also. This is for +ringexit to make sure one...
Gengbin Zheng [Wed, 21 Jul 2004 04:47:46 +0000 (04:47 +0000)]
when closing logs, fsync files to disk also. This is for +ringexit to make sure one log file on a processor is completed before switching to another processor.

17 years agodisable ampi and ampiParent constructor tracing to avoid nested prjections/summary...
Gengbin Zheng [Wed, 21 Jul 2004 03:23:29 +0000 (03:23 +0000)]
disable ampi and ampiParent constructor tracing to avoid nested prjections/summary events.

17 years agobetter description of FuturePredictor
Filippo Gioachin [Wed, 21 Jul 2004 00:27:37 +0000 (00:27 +0000)]
better description of FuturePredictor

17 years agosmall fix
Filippo Gioachin [Tue, 20 Jul 2004 21:46:56 +0000 (21:46 +0000)]
small fix

17 years agoadded two functions to support asynchronous load balancing, i.e. when receiving migra...
Gengbin Zheng [Tue, 20 Jul 2004 19:53:51 +0000 (19:53 +0000)]
added two functions to support asynchronous load balancing, i.e. when receiving migration decision, it only remembers it but only migrate when AMPI_Allow_Migrate() is called.

17 years agoupdated the instruction for setting VC directories.
Gengbin Zheng [Tue, 20 Jul 2004 13:50:09 +0000 (13:50 +0000)]
updated the instruction for setting VC directories.

17 years agoI changed two of the settings for default message send strategy boundaries
Greg Koenig [Tue, 20 Jul 2004 09:23:15 +0000 (09:23 +0000)]
I changed two of the settings for default message send strategy boundaries
earlier today when I checked in changes.  These new boundaries were to
optimize the machine layer for machines with faster memory controllers.
These values work fine for all the Converse and Charm++ tests as well as
with NAMD, but when testing with leanMD I discovered that they are too
aggressive.  Replaced with earlier values.  These can be overridden at
runtime with environment variables anyway.

CMI_VMI_SMALL_MESSAGE_BOUNDARY   1024
CMI_VMI_MEDIUM_MESSAGE_BOUNDARY  16384

17 years agoAdded new flavor of "accumulate": returns an lvalue the user
Orion Lawlor [Tue, 20 Jul 2004 01:24:06 +0000 (01:24 +0000)]
Added new flavor of "accumulate": returns an lvalue the user
can accumulate to.

In particular, this feature will be used during FEM mesh
partitioning, where "accumulate" means merge two FEM meshes,
and we'd like to be able to do:
    meshMsa.accumulate(chunk).node.push_back(...);
    meshMsa.accumulate(chunk).elem[e].push_back(...);
Because we'd like to touch *both* nodes *and* elements, it'd
be weird and painful to write a single element type that can
handle both.  Hence this new API.

17 years agolink Makefile_lb to tmp.
Gengbin Zheng [Mon, 19 Jul 2004 21:41:41 +0000 (21:41 +0000)]
link Makefile_lb to tmp.

17 years agoadded Makefile for ck-ldb
Gengbin Zheng [Mon, 19 Jul 2004 21:38:17 +0000 (21:38 +0000)]
added Makefile for ck-ldb

17 years agoThese modifications are a better solution to the problem of spanning tree
Greg Koenig [Mon, 19 Jul 2004 21:37:42 +0000 (21:37 +0000)]
These modifications are a better solution to the problem of spanning tree
broadcasts which does not involve forwarding to the spanning children from
within CmiGetNonLocal().

The problem before was that a child in the spanning tree would receive an
RDMA broadcast message and attempt to forward the message to its children
in the spanning tree from within the RDMA receive handler.  Unfortunately
VMI never invoked the RDMA send completion handler because an underlying
event on Myrinet (and only Myrinet; everything worked fine on Ethernet and
on Infiniband) never got delievered.  This resulted in the node spinning
forever.

To fix this correctly, I realized that I could set a "disposition" on each
send handle for what to do in the send handler (nothing, free the message,
enqueue the message for delivery to Converse) and then forward the message
to the spanning tree children asynchronously instead of synchronously.
In this scenario, the send completion handler gets invoked after the
receive handler exits, and all is happy.

17 years agouse include Makefile_lb in main Makefile.
Gengbin Zheng [Mon, 19 Jul 2004 21:37:41 +0000 (21:37 +0000)]
use include Makefile_lb in main Makefile.

17 years agoUpdates to refineClient interface: updated collapse, added nodeUpdate, nodeReplace...
Terry L. Wilmarth [Mon, 19 Jul 2004 21:07:05 +0000 (21:07 +0000)]
Updates to refineClient interface: updated collapse, added nodeUpdate, nodeReplace, and nodeDelete.