improve slurm epilog script

epilog will now skip on-node compression if only a single compute
node is used and will generate the final log file instead

darshan-runtime/lib/darshan-core.c

@@ -391,8 +391,7 @@ void darshan_core_shutdown()
     if(getenv("DARSHAN_INTERNAL_TIMING"))
         internal_timing_flag = 1;
 
-    if(internal_timing_flag)
-        start_log_time = DARSHAN_MPI_CALL(PMPI_Wtime)();
+    start_log_time = DARSHAN_MPI_CALL(PMPI_Wtime)();
 
     /* disable darhan-core while we shutdown */
     DARSHAN_CORE_LOCK();

darshan-runtime/share/darshan-mmap-epilog.sh.in

@@ -11,8 +11,6 @@ DARSHAN_INSTALL_DIR=@prefix@
 # use the log dir specified at configure time
 DARSHAN_LOG_DIR=@__DARSHAN_LOG_PATH@
 
-NODE_NAME=$(uname -n)
-
 # use the default mmap log directory (/tmp), unless the 
 # env variable is set to something
 if [ -z "$DARSHAN_MMAP_LOGPATH" ]; then
@@ -24,36 +22,54 @@ fi
 DARSHAN_MMAP_LOG_GLOB=${DARSHAN_MMAP_LOG_DIR}/*id${SLURM_JOB_ID}*.darshan
 
 # if no mmap logs found for this job, we have nothing to do
-DARSHAN_TEST_LOG=$(ls $DARSHAN_MMAP_LOG_GLOB 2>/dev/null | head -n 1)
-if [ -z $DARSHAN_TEST_LOG ]; then
+DARSHAN_MMAP_LOG1=$(ls $DARSHAN_MMAP_LOG_GLOB 2>/dev/null | head -n 1)
+if [ -z $DARSHAN_MMAP_LOG1 ]; then
     exit 0
 fi
 
 # get the job start time from the first log file
-JOB_START_DATE=$(${DARSHAN_INSTALL_DIR}/bin/darshan-parser $DARSHAN_TEST_LOG |
+JOB_START_DATE=$(${DARSHAN_INSTALL_DIR}/bin/darshan-parser $DARSHAN_MMAP_LOG1 |
     grep "# start_time_asci" | cut -d':' -f 2- | cut -d' ' -f 2-)
-OUTLOG_YEAR=$(date --date="$(printf "$JOB_START_DATE")" +"%Y")
-OUTLOG_MON=$(date --date="$(printf "$JOB_START_DATE")" +"%-m")
-OUTLOG_DAY=$(date --date="$(printf "$JOB_START_DATE")" +"%-d")
-OUTLOG_SECS=$((
+OUTPUT_YEAR=$(date --date="$(printf "$JOB_START_DATE")" +"%Y")
+OUTPUT_MON=$(date --date="$(printf "$JOB_START_DATE")" +"%-m")
+OUTPUT_DAY=$(date --date="$(printf "$JOB_START_DATE")" +"%-d")
+OUTPUT_SECS=$((
     ($(date --date="$(printf "$JOB_START_DATE")" +"%-H") * 60 * 60) +
     ($(date --date="$(printf "$JOB_START_DATE")" +"%-M") * 60) +
     ($(date --date="$(printf "$JOB_START_DATE")" +"%-S"))
 ))
 
-MMAP_LOG_PRE=$(basename $DARSHAN_TEST_LOG | cut -d'_' -f 1-3)
+LOG_NAME_PRE=$(basename $DARSHAN_MMAP_LOG1 | cut -d'_' -f 1-3)
+
+# construct full name of directory to store output log(s)
+OUTPUT_LOG_DIR=${DARSHAN_LOG_DIR}/${OUTPUT_YEAR}/${OUTPUT_MON}/${OUTPUT_DAY}/
+OUTPUT_NAME_PRE=${LOG_NAME_PRE}_${OUTPUT_MON}-${OUTPUT_DAY}-${OUTPUT_SECS}
+
+if [ $SLURM_NNODES -gt 1 ]; then
+    NODE_LOG_DIR=${OUTPUT_LOG_DIR}/${OUTPUT_NAME_PRE}
+    NODE_NAME=$(uname -n)
 
-# construct full output log directory name
-OUTLOG_DIR=${DARSHAN_LOG_DIR}/${OUTLOG_YEAR}/${OUTLOG_MON}/${OUTLOG_DAY}/
-OUTLOG_DIR=${OUTLOG_DIR}/${MMAP_LOG_PRE}_${OUTLOG_MON}-${OUTLOG_DAY}-${OUTLOG_SECS}
+    # multiple nodes, create a node log directory for everyone to write to
+    mkdir -p $NODE_LOG_DIR
 
-# TODO: do shared reduction and skip mkdir if just one compute node?
+    # construct the per-node log file and store in the output directory
+    $DARSHAN_INSTALL_DIR/bin/darshan-merge \
+        --output ${NODE_LOG_DIR}/${LOG_NAME_PRE}_${NODE_NAME}.darshan \
+        $DARSHAN_MMAP_LOG_GLOB
+else
+    TMP_LOG=${OUTPUT_NAME_PRE}.darshan
+
+    # single node, just create the final output darshan log
+    LOG_WRITE_START=$(date +%s)
+    $DARSHAN_INSTALL_DIR/bin/darshan-merge
+        --shared-redux --output ${OUTPUT_LOG_DIR}/${TMP_LOG} \
+        $DARSHAN_MMAP_LOG_GLOB
+    LOG_WRITE_END=$(date +%s)
 
-# create the output directory for this job
-mkdir -p $OUTLOG_DIR
+    WRITE_TM=$(($LOG_WRITE_END - $LOG_WRITE_START + 1))
+    FINAL_LOG=${OUTPUT_NAME_PRE}-${RANDOM}_${WRITE_TM}.darshan
 
-# construct the per-node log files and store in the output directory
-$DARSHAN_INSTALL_DIR/bin/darshan-merge --output ${OUTLOG_DIR}/${MMAP_LOG_PRE}_${NODE_NAME}.darshan \
-    $DARSHAN_MMAP_LOG_GLOB
+    mv ${OUTPUT_LOG_DIR}/${TMP_LOG} ${OUTPUT_LOG_DIR}/${FINAL_LOG}
+fi
 
 exit 0