diff -Nur hpl/HISTORY hpl-patched/HISTORY
--- hpl/HISTORY 2004-01-22 00:13:09.000000000 -0500
+++ hpl-patched/HISTORY 2006-12-14 10:19:19.000000000 -0500
@@ -20,4 +20,7 @@
Various building problems on the T3E have been fixed; Thanks
to Edward Anderson.
+ - 06/26/06 applied the "native MPI_Bcast() patch for HPL"
+ - 06/30/06 applied the "non-blocking collective operations patch to HPL"
+
==============================================================
diff -Nur hpl/include/hpl_comm.h hpl-patched/include/hpl_comm.h
--- hpl/include/hpl_comm.h 2004-01-22 00:13:17.000000000 -0500
+++ hpl-patched/include/hpl_comm.h 2006-12-14 10:06:53.000000000 -0500
@@ -64,7 +64,9 @@
HPL_2RING = 403, /* Increasing 2-ring */
HPL_2RING_M = 404, /* Increasing 2-ring (modified) */
HPL_BLONG = 405, /* long broadcast */
- HPL_BLONG_M = 406 /* long broadcast (modified) */
+ HPL_BLONG_M = 406, /* long broadcast (modified) */
+ HPL_NATIVE = 407, /* native MPI_Bcast() */
+ HPL_NBC = 408 /* non-blocking NBC_Ibcast() */
} HPL_T_TOP;
/*
* ---------------------------------------------------------------------
@@ -155,6 +157,14 @@
int HPL_bcast_blonM STDC_ARGS( ( HPL_T_panel *, int * ) );
int HPL_bwait_blonM STDC_ARGS( ( HPL_T_panel * ) );
+int HPL_binit_naMPI STDC_ARGS( ( HPL_T_panel * ) );
+int HPL_bcast_naMPI STDC_ARGS( ( HPL_T_panel *, int * ) );
+int HPL_bwait_naMPI STDC_ARGS( ( HPL_T_panel * ) );
+
+int HPL_binit_nbc STDC_ARGS( ( HPL_T_panel * ) );
+int HPL_bcast_nbc STDC_ARGS( ( HPL_T_panel *, int * ) );
+int HPL_bwait_nbc STDC_ARGS( ( HPL_T_panel * ) );
+
#endif
/*
* End of hpl_comm.h
diff -Nur hpl/makes/Make.comm hpl-patched/makes/Make.comm
--- hpl/makes/Make.comm 2004-01-22 00:13:14.000000000 -0500
+++ hpl-patched/makes/Make.comm 2006-12-14 10:05:37.000000000 -0500
@@ -59,7 +59,9 @@
HPL_2rinM.o HPL_blong.o HPL_blonM.o \
HPL_packL.o HPL_copyL.o HPL_binit.o \
HPL_bcast.o HPL_bwait.o HPL_send.o \
- HPL_recv.o HPL_sdrv.o
+ HPL_recv.o HPL_sdrv.o HPL_naMPI.o \
+ HPL_nbc.o
+
#
## Targets #############################################################
#
@@ -86,6 +88,10 @@
$(CC) -o $@ -c $(CCFLAGS) ../HPL_blong.c
HPL_blonM.o : ../HPL_blonM.c $(INCdep)
$(CC) -o $@ -c $(CCFLAGS) ../HPL_blonM.c
+HPL_naMPI.o : ../HPL_naMPI.c $(INCdep)
+ $(CC) -o $@ -c $(CCFLAGS) ../HPL_naMPI.c
+HPL_nbc.o : ../HPL_nbc.c $(INCdep)
+ $(CC) -o $@ -c $(CCFLAGS) ../HPL_nbc.c
HPL_packL.o : ../HPL_packL.c $(INCdep)
$(CC) -o $@ -c $(CCFLAGS) ../HPL_packL.c
HPL_copyL.o : ../HPL_copyL.c $(INCdep)
diff -Nur hpl/src/comm/HPL_bcast.c hpl-patched/src/comm/HPL_bcast.c
--- hpl/src/comm/HPL_bcast.c 2004-01-22 00:13:43.000000000 -0500
+++ hpl-patched/src/comm/HPL_bcast.c 2006-12-14 10:06:26.000000000 -0500
@@ -108,6 +108,8 @@
case HPL_2RING : ierr = HPL_bcast_2ring( PANEL, IFLAG ); break;
case HPL_BLONG_M : ierr = HPL_bcast_blonM( PANEL, IFLAG ); break;
case HPL_BLONG : ierr = HPL_bcast_blong( PANEL, IFLAG ); break;
+ case HPL_NATIVE : ierr = HPL_bcast_naMPI( PANEL, IFLAG ); break;
+ case HPL_NBC : ierr = HPL_bcast_nbc( PANEL, IFLAG ); break;
default : ierr = HPL_SUCCESS;
}
diff -Nur hpl/src/comm/HPL_binit.c hpl-patched/src/comm/HPL_binit.c
--- hpl/src/comm/HPL_binit.c 2004-01-22 00:13:43.000000000 -0500
+++ hpl-patched/src/comm/HPL_binit.c 2006-12-14 10:07:23.000000000 -0500
@@ -98,6 +98,8 @@
case HPL_2RING : ierr = HPL_binit_2ring( PANEL ); break;
case HPL_BLONG_M : ierr = HPL_binit_blonM( PANEL ); break;
case HPL_BLONG : ierr = HPL_binit_blong( PANEL ); break;
+ case HPL_NATIVE : ierr = HPL_binit_naMPI( PANEL ); break;
+ case HPL_NBC : ierr = HPL_binit_nbc( PANEL ); break;
default : ierr = HPL_SUCCESS;
}
diff -Nur hpl/src/comm/HPL_bwait.c hpl-patched/src/comm/HPL_bwait.c
--- hpl/src/comm/HPL_bwait.c 2004-01-22 00:13:43.000000000 -0500
+++ hpl-patched/src/comm/HPL_bwait.c 2006-12-14 10:07:57.000000000 -0500
@@ -99,6 +99,8 @@
case HPL_2RING : ierr = HPL_bwait_2ring( PANEL ); break;
case HPL_BLONG_M : ierr = HPL_bwait_blonM( PANEL ); break;
case HPL_BLONG : ierr = HPL_bwait_blong( PANEL ); break;
+ case HPL_NATIVE : ierr = HPL_bwait_naMPI( PANEL ); break;
+ case HPL_NBC : ierr = HPL_bwait_nbc( PANEL ); break;
default : ierr = HPL_SUCCESS;
}
diff -Nur hpl/src/comm/HPL_naMPI.c hpl-patched/src/comm/HPL_naMPI.c
--- hpl/src/comm/HPL_naMPI.c 1969-12-31 19:00:00.000000000 -0500
+++ hpl-patched/src/comm/HPL_naMPI.c 2006-12-01 05:37:06.000000000 -0500
@@ -0,0 +1,190 @@
+/*
+ * derived from HPL_1ring.c; 7th algorithm which uses native MPI_Bcast()
+ */
+/*
+ * -- High Performance Computing Linpack Benchmark (HPL)
+ * HPL - 1.0a - January 20, 2004
+ * Antoine P. Petitet
+ * University of Tennessee, Knoxville
+ * Innovative Computing Laboratories
+ * (C) Copyright 2000-2004 All Rights Reserved
+ *
+ * -- Copyright notice and Licensing terms:
+ *
+ * Redistribution and use in source and binary forms, with or without
+ * modification, are permitted provided that the following conditions
+ * are met:
+ *
+ * 1. Redistributions of source code must retain the above copyright
+ * notice, this list of conditions and the following disclaimer.
+ *
+ * 2. Redistributions in binary form must reproduce the above copyright
+ * notice, this list of conditions, and the following disclaimer in the
+ * documentation and/or other materials provided with the distribution.
+ *
+ * 3. All advertising materials mentioning features or use of this
+ * software must display the following acknowledgement:
+ * This product includes software developed at the University of
+ * Tennessee, Knoxville, Innovative Computing Laboratories.
+ *
+ * 4. The name of the University, the name of the Laboratory, or the
+ * names of its contributors may not be used to endorse or promote
+ * products derived from this software without specific written
+ * permission.
+ *
+ * -- Disclaimer:
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+ * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+ * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+ * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE UNIVERSITY
+ * OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+ * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+ * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+ * DATA OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+ * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ * ---------------------------------------------------------------------
+ */
+/*
+ * Include files
+ */
+#include "hpl.h"
+
+#ifdef HPL_NO_MPI_DATATYPE /* The user insists to not use MPI types */
+#ifndef HPL_COPY_L /* and also want to avoid the copy of L ... */
+#define HPL_COPY_L /* well, sorry, can not do that: force the copy */
+#endif
+#endif
+
+#ifdef STDC_HEADERS
+int HPL_binit_naMPI
+(
+ HPL_T_panel * PANEL
+)
+#else
+int HPL_binit_naMPI( PANEL )
+ HPL_T_panel * PANEL;
+#endif
+{
+#ifdef HPL_USE_MPI_DATATYPE
+/*
+ * .. Local Variables ..
+ */
+ int ierr;
+#endif
+/* ..
+ * .. Executable Statements ..
+ */
+ if( PANEL == NULL ) { return( HPL_SUCCESS ); }
+ if( PANEL->grid->npcol <= 1 ) { return( HPL_SUCCESS ); }
+#ifdef HPL_USE_MPI_DATATYPE
+#ifdef HPL_COPY_L
+/*
+ * Copy the panel into a contiguous buffer
+ */
+ HPL_copyL( PANEL );
+#endif
+/*
+ * Create the MPI user-defined data type
+ */
+ ierr = HPL_packL( PANEL, 0, PANEL->len, 0 );
+
+ return( ( ierr == MPI_SUCCESS ? HPL_SUCCESS : HPL_FAILURE ) );
+#else
+/*
+ * Force the copy of the panel into a contiguous buffer
+ */
+ HPL_copyL( PANEL );
+
+ return( HPL_SUCCESS );
+#endif
+}
+
+#ifdef HPL_USE_MPI_DATATYPE
+
+#define _M_BUFF PANEL->buffers[0]
+#define _M_COUNT PANEL->counts[0]
+#define _M_TYPE PANEL->dtypes[0]
+
+#else
+
+#define _M_BUFF (void *)(PANEL->L2)
+#define _M_COUNT PANEL->len
+#define _M_TYPE MPI_DOUBLE
+
+#endif
+
+#ifdef STDC_HEADERS
+int HPL_bcast_naMPI
+(
+ HPL_T_panel * PANEL,
+ int * IFLAG
+)
+#else
+int HPL_bcast_naMPI( PANEL, IFLAG )
+ HPL_T_panel * PANEL;
+ int * IFLAG;
+#endif
+{
+/*
+ * .. Local Variables ..
+ */
+ MPI_Comm comm;
+ int ierr, root, size;
+/* ..
+ * .. Executable Statements ..
+ */
+ if( PANEL == NULL ) { *IFLAG = HPL_SUCCESS; return( HPL_SUCCESS ); }
+ if( ( size = PANEL->grid->npcol ) <= 1 )
+ { *IFLAG = HPL_SUCCESS; return( HPL_SUCCESS ); }
+/*
+ * The root process spreads the panel. Each non root process waits until
+ * it receives the panel (no non-blocking feature anymore!)
+ */
+ comm = PANEL->grid->row_comm;
+ root = PANEL->pcol;
+
+ ierr = MPI_Bcast( _M_BUFF, _M_COUNT, _M_TYPE, root, comm);
+
+/*
+ * If the message was received or being forwarded, return HPL_SUCCESS.
+ * If an error occured in an MPI call, return HPL_FAILURE.
+ */
+ *IFLAG = ( ierr == MPI_SUCCESS ? HPL_SUCCESS : HPL_FAILURE );
+
+ return( *IFLAG );
+}
+
+#ifdef STDC_HEADERS
+int HPL_bwait_naMPI
+(
+ HPL_T_panel * PANEL
+)
+#else
+int HPL_bwait_naMPI( PANEL )
+ HPL_T_panel * PANEL;
+#endif
+{
+#ifdef HPL_USE_MPI_DATATYPE
+/*
+ * .. Local Variables ..
+ */
+ int ierr;
+#endif
+/* ..
+ * .. Executable Statements ..
+ */
+ if( PANEL == NULL ) { return( HPL_SUCCESS ); }
+ if( PANEL->grid->npcol <= 1 ) { return( HPL_SUCCESS ); }
+/*
+ * Release the arrays of request / status / data-types and buffers
+ */
+#ifdef HPL_USE_MPI_DATATYPE
+ ierr = MPI_Type_free( &PANEL->dtypes[0] );
+ return( ( ierr == MPI_SUCCESS ? HPL_SUCCESS : HPL_FAILURE ) );
+#else
+ return( HPL_SUCCESS );
+#endif
+}
diff -Nur hpl/src/comm/HPL_nbc.c hpl-patched/src/comm/HPL_nbc.c
--- hpl/src/comm/HPL_nbc.c 1969-12-31 19:00:00.000000000 -0500
+++ hpl-patched/src/comm/HPL_nbc.c 2006-12-14 10:09:45.000000000 -0500
@@ -0,0 +1,150 @@
+/*
+ * Include files
+ */
+#include "hpl.h"
+#include "nbc.h"
+
+#ifdef HPL_NO_MPI_DATATYPE /* The user insists to not use MPI types */
+#ifndef HPL_COPY_L /* and also want to avoid the copy of L ... */
+#define HPL_COPY_L /* well, sorry, can not do that: force the copy */
+#endif
+#endif
+
+#ifdef STDC_HEADERS
+int HPL_binit_nbc
+(
+ HPL_T_panel * PANEL
+)
+#else
+int HPL_binit_nbc( PANEL )
+ HPL_T_panel * PANEL;
+#endif
+{
+#ifdef HPL_USE_MPI_DATATYPE
+/*
+ * .. Local Variables ..
+ */
+ int ierr;
+#endif
+/* ..
+ * .. Executable Statements ..
+ */
+ if( PANEL == NULL ) { return( HPL_SUCCESS ); }
+ if( PANEL->grid->npcol <= 1 ) { return( HPL_SUCCESS ); }
+#ifdef HPL_USE_MPI_DATATYPE
+#ifdef HPL_COPY_L
+/*
+ * Copy the panel into a contiguous buffer
+ */
+ HPL_copyL( PANEL );
+#endif
+/*
+ * Create the MPI user-defined data type
+ */
+ ierr = HPL_packL( PANEL, 0, PANEL->len, 0 );
+
+ return( ( ierr == MPI_SUCCESS ? HPL_SUCCESS : HPL_FAILURE ) );
+#else
+/*
+ * Force the copy of the panel into a contiguous buffer
+ */
+ HPL_copyL( PANEL );
+
+ return( HPL_SUCCESS );
+#endif
+}
+
+#ifdef HPL_USE_MPI_DATATYPE
+
+#define _M_BUFF PANEL->buffers[0]
+#define _M_COUNT PANEL->counts[0]
+#define _M_TYPE PANEL->dtypes[0]
+
+#else
+
+#define _M_BUFF (void *)(PANEL->L2)
+#define _M_COUNT PANEL->len
+#define _M_TYPE MPI_DOUBLE
+
+#endif
+
+/* global handle */
+NBC_Handle handle;
+int sent=0;
+
+#ifdef STDC_HEADERS
+int HPL_bcast_nbc
+(
+ HPL_T_panel * PANEL,
+ int * IFLAG
+)
+#else
+int HPL_bcast_nbc( PANEL, IFLAG )
+ HPL_T_panel * PANEL;
+ int * IFLAG;
+#endif
+{
+/*
+ * .. Local Variables ..
+ */
+ MPI_Comm comm;
+ int ierr, root;
+
+/* ..
+ * .. Executable Statements ..
+ */
+ if( PANEL == NULL ) { *IFLAG = HPL_SUCCESS; return( HPL_SUCCESS ); }
+ if( ( PANEL->grid->npcol ) <= 1 )
+ { *IFLAG = HPL_SUCCESS; return( HPL_SUCCESS ); }
+
+ comm = PANEL->grid->row_comm;
+ root = PANEL->pcol;
+
+ //ierr = MPI_Bcast(_M_BUFF, _M_COUNT, _M_TYPE, root, comm);
+
+ if(sent == 0) {
+ ierr = NBC_Ibcast(_M_BUFF, _M_COUNT, _M_TYPE, root, comm, &handle);
+ sent = 1;
+ }
+
+ if(NBC_Test(&handle) == NBC_OK) {
+ *IFLAG = HPL_SUCCESS;
+ sent = 0;
+ }
+ else
+ *IFLAG = HPL_KEEP_TESTING;
+
+ return( *IFLAG );
+}
+
+#ifdef STDC_HEADERS
+int HPL_bwait_nbc
+(
+ HPL_T_panel * PANEL
+)
+#else
+int HPL_bwait_nbc( PANEL )
+ HPL_T_panel * PANEL;
+#endif
+{
+#ifdef HPL_USE_MPI_DATATYPE
+/*
+ * .. Local Variables ..
+ */
+ int ierr;
+#endif
+/* ..
+ * .. Executable Statements ..
+ */
+ if( PANEL == NULL ) { return( HPL_SUCCESS ); }
+ if( PANEL->grid->npcol <= 1 ) { return( HPL_SUCCESS ); }
+/*
+ * Release the arrays of request / status / data-types and buffers
+ */
+#ifdef HPL_USE_MPI_DATATYPE
+ ierr = MPI_Type_free( &PANEL->dtypes[0] );
+ return( ( ierr == MPI_SUCCESS ? HPL_SUCCESS : HPL_FAILURE ) );
+#else
+ return( HPL_SUCCESS );
+#endif
+}
diff -Nur hpl/testing/ptest/HPL.dat hpl-patched/testing/ptest/HPL.dat
--- hpl/testing/ptest/HPL.dat 2004-01-22 00:13:29.000000000 -0500
+++ hpl-patched/testing/ptest/HPL.dat 2006-12-14 10:10:19.000000000 -0500
@@ -20,7 +20,7 @@
3 # of recursive panel fact.
0 1 2 RFACTs (0=left, 1=Crout, 2=Right)
1 # of broadcast
-0 BCASTs (0=1rg,1=1rM,2=2rg,3=2rM,4=Lng,5=LnM)
+0 BCASTs (0=1rg,1=1rM,2=2rg,3=2rM,4=Lng,5=LnM,6=MPI,7=NBC)
1 # of lookahead depth
0 DEPTHs (>=0)
2 SWAP (0=bin-exch,1=long,2=mix)
diff -Nur hpl/testing/ptest/HPL_pddriver.c hpl-patched/testing/ptest/HPL_pddriver.c
--- hpl/testing/ptest/HPL_pddriver.c 2004-01-22 00:13:21.000000000 -0500
+++ hpl-patched/testing/ptest/HPL_pddriver.c 2006-12-14 10:10:34.000000000 -0500
@@ -137,7 +137,7 @@
* 3 # of recursive panel fact.
* 0 1 2 RFACTs (0=left, 1=Crout, 2=Right)
* 1 # of broadcast
- * 0 BCASTs (0=1rg,1=1rM,2=2rg,3=2rM,4=Lng,5=LnM)
+ * 0 BCASTs (0=1rg,1=1rM,2=2rg,3=2rM,4=Lng,5=LnM,6=MPI,7=NBC)
* 1 # of lookahead depth
* 0 DEPTHs (>=0)
* 2 SWAP (0=bin-exch,1=long,2=mix)
diff -Nur hpl/testing/ptest/HPL_pdinfo.c hpl-patched/testing/ptest/HPL_pdinfo.c
--- hpl/testing/ptest/HPL_pdinfo.c 2004-01-22 00:13:25.000000000 -0500
+++ hpl-patched/testing/ptest/HPL_pdinfo.c 2006-12-14 10:13:19.000000000 -0500
@@ -516,7 +516,7 @@
else RF[ i ] = HPL_RIGHT_LOOKING;
}
/*
- * Broadcast topology (TP) (0=rg, 1=2rg, 2=rgM, 3=2rgM, 4=L)
+ * Broadcast topology (TP) (0=1rg, 1=1rM, 2=2rg, 3=2rM, 4=Lng, 5=LnM, 6=MPI, 7=NBC)
*/
(void) fgets( line, HPL_LINE_MAX - 2, infp );
(void) sscanf( line, "%s", num ); *NTPS = atoi( num );
@@ -538,6 +538,8 @@
else if( j == 3 ) TP[ i ] = HPL_2RING_M;
else if( j == 4 ) TP[ i ] = HPL_BLONG;
else if( j == 5 ) TP[ i ] = HPL_BLONG_M;
+ else if( j == 6 ) TP[ i ] = HPL_NATIVE;
+ else if( j == 7 ) TP[ i ] = HPL_NBC;
else TP[ i ] = HPL_1RING_M;
}
/*
@@ -696,6 +698,8 @@
else if( TP[i] == HPL_2RING_M ) iwork[j] = 3;
else if( TP[i] == HPL_BLONG ) iwork[j] = 4;
else if( TP[i] == HPL_BLONG_M ) iwork[j] = 5;
+ else if( TP[i] == HPL_NATIVE ) iwork[j] = 6;
+ else if( TP[i] == HPL_NBC ) iwork[j] = 7;
j++;
}
for( i = 0; i < *NDHS; i++ ) { iwork[j] = DH[i]; j++; }
@@ -739,6 +743,8 @@
else if( iwork[j] == 3 ) TP[i] = HPL_2RING_M;
else if( iwork[j] == 4 ) TP[i] = HPL_BLONG;
else if( iwork[j] == 5 ) TP[i] = HPL_BLONG_M;
+ else if( iwork[j] == 6 ) TP[i] = HPL_NATIVE;
+ else if( iwork[j] == 7 ) TP[i] = HPL_NBC;
j++;
}
for( i = 0; i < *NDHS; i++ ) { DH[i] = iwork[j]; j++; }
@@ -994,6 +1000,10 @@
HPL_fprintf( TEST->outfp, " Blong " );
else if( TP[i] == HPL_BLONG_M )
HPL_fprintf( TEST->outfp, " BlongM " );
+ else if( TP[i] == HPL_NATIVE )
+ HPL_fprintf( TEST->outfp, " natMPI " );
+ else if( TP[i] == HPL_NBC )
+ HPL_fprintf( TEST->outfp, " NBC " );
}
if( *NTPS > 8 )
{
@@ -1012,6 +1022,10 @@
HPL_fprintf( TEST->outfp, " Blong " );
else if( TP[i] == HPL_BLONG_M )
HPL_fprintf( TEST->outfp, " BlongM " );
+ else if( TP[i] == HPL_NATIVE )
+ HPL_fprintf( TEST->outfp, " natMPI " );
+ else if( TP[i] == HPL_NBC )
+ HPL_fprintf( TEST->outfp, " NBC " );
}
if( *NTPS > 16 )
{
@@ -1030,6 +1044,10 @@
HPL_fprintf( TEST->outfp, " Blong " );
else if( TP[i] == HPL_BLONG_M )
HPL_fprintf( TEST->outfp, " BlongM " );
+ else if( TP[i] == HPL_NATIVE )
+ HPL_fprintf( TEST->outfp, " natMPI " );
+ else if( TP[i] == HPL_NBC )
+ HPL_fprintf( TEST->outfp, " NBC " );
}
}
}
diff -Nur hpl/testing/ptest/HPL_pdtest.c hpl-patched/testing/ptest/HPL_pdtest.c
--- hpl/testing/ptest/HPL_pdtest.c 2004-01-22 00:13:29.000000000 -0500
+++ hpl-patched/testing/ptest/HPL_pdtest.c 2006-12-14 10:13:58.000000000 -0500
@@ -235,7 +235,10 @@
else if( ALGO->btopo == HPL_2RING ) ctop = '2';
else if( ALGO->btopo == HPL_2RING_M ) ctop = '3';
else if( ALGO->btopo == HPL_BLONG ) ctop = '4';
- else /* if( ALGO->btopo == HPL_BLONG_M ) */ ctop = '5';
+ else if( ALGO->btopo == HPL_BLONG_M ) ctop = '5';
+ else if( ALGO->btopo == HPL_NATIVE ) ctop = '6';
+ else if( ALGO->btopo == HPL_NBC ) ctop = '7';
+ else ctop = '?';
if( wtime[0] > HPL_rzero )
HPL_fprintf( TEST->outfp,
diff -Nur hpl/TUNING hpl-patched/TUNING
--- hpl/TUNING 2004-01-22 00:13:09.000000000 -0500
+++ hpl-patched/TUNING 2006-12-14 10:02:11.000000000 -0500
@@ -217,13 +217,13 @@
a good choice. Lines 22-23: (Example 1):
1 # of broadcast
- 1 BCASTs (0=1rg,1=1rM,2=2rg,3=2rM,4=Lng,5=LnM)
+ 1 BCASTs (0=1rg,1=1rM,2=2rg,3=2rM,4=Lng,5=LnM,6=MPI,7=NBC)
This will cause HPL to broadcast the current panel using the
increasing ring modified topology. Lines 22-23: (Example 2):
2 # of broadcast
- 0 4 BCASTs (0=1rg,1=1rM,2=2rg,3=2rM,4=Lng,5=LnM)
+ 0 4 BCASTs (0=1rg,1=1rM,2=2rg,3=2rM,4=Lng,5=LnM,6=MPI,7=NBC)
This will cause HPL to broadcast the current panel using the
increasing ring virtual topology and the long message algori-
@@ -359,12 +359,17 @@
wing for the lines 22-23:
2 # of broadcast
- 1 3 BCASTs (0=1rg,1=1rM,2=2rg,3=2rM,4=Lng,5=LnM)
+ 1 3 BCASTs (0=1rg,1=1rM,2=2rg,3=2rM,4=Lng,5=LnM,6=MPI,7=NBC)
The best broadcast depends on your problem size and harware
performance. My take is that 4 or 5 may be competitive for
machines featuring very fast nodes comparatively to the
network.
+ Two new broadcast versions are added to the original HPL code. One uses
+ the MPI_Bcast() to perform the broadcast operation (option 6). The
+ second one uses the non-blocking broadcast implemented in LibNBC
+ (http://www.unixer.de/NBC) to benefit from overlapping computation and
+ communication.
6) Look-ahead depth: as mentioned above 0 or 1 are likely to
be the best choices. This also depends on the problem size
diff -Nur hpl/www/tuning.html hpl-patched/www/tuning.html
--- hpl/www/tuning.html 2004-01-22 00:13:08.000000000 -0500
+++ hpl-patched/www/tuning.html 2006-12-14 10:15:34.000000000 -0500
@@ -252,14 +252,14 @@
Lines 22-23: (Example 1)
1 # of broadcast
-1 BCASTs (0=1rg,1=1rM,2=2rg,3=2rM,4=Lng,5=LnM)
+1 BCASTs (0=1rg,1=1rM,2=2rg,3=2rM,4=Lng,5=LnM,6=MPI,7=NBC)
This will cause HPL to broadcast the current panel using the
increasing ring modified topology.
Lines 22-23: (Example 2)
2 # of broadcast
-0 4 BCASTs (0=1rg,1=1rM,2=2rg,3=2rM,4=Lng,5=LnM)
+0 4 BCASTs (0=1rg,1=1rM,2=2rg,3=2rM,4=Lng,5=LnM,6=MPI,7=NBC)
This will cause HPL to broadcast the current panel using the
increasing ring virtual topology and the long message
@@ -394,12 +394,18 @@
following for the lines 22-23:
2 # of broadcast
-1 3 BCASTs (0=1rg,1=1rM,2=2rg,3=2rM,4=Lng,5=LnM)
+1 3 BCASTs (0=1rg,1=1rM,2=2rg,3=2rM,4=Lng,5=LnM,6=MPI,7=NBC)
The best broadcast depends on your problem size and harware
performance. My take is that 4 or 5 may be competitive for
machines featuring very fast nodes comparatively to the
-network.
+network.
+Two new broadcast versions are added to the original HPL code. One uses
+the MPI_Bcast() to perform the broadcast operation (option 6). The
+second one uses the non-blocking broadcast implemented in LibNBC
+(http://www.unixer.de/NBC) to benefit from overlapping computation and
+communication.
+
Look-ahead depth: as mentioned above 0 or 1 are likely to
be the best choices. This also depends on the problem size