diff --git a/src/parallel b/src/parallel
index 1d0d0600..cafad983 100755
--- a/src/parallel
+++ b/src/parallel
@@ -1297,15 +1297,18 @@ sub start_more_jobs {
             }
         }
 
+	my $sleep = 0.0001; # 0.01 ms - better performance on highend
         for my $sshlogin (values %Global::host) {
             debug("Running jobs before on ".$sshlogin->string().": ".$sshlogin->jobs_running()."\n");
 	    while ($sshlogin->jobs_running() < $sshlogin->max_jobs_running()) {
 		if($opt::load and $sshlogin->loadavg_too_high()) {
 		    # The load is too high or unknown
+		    $sleep = ::reap_usleep($sleep);
 		    next;
 		}
 		if($opt::noswap and $sshlogin->swapping()) {
 		    # The server is swapping
+		    $sleep = ::reap_usleep($sleep);
 		    next;
 		}
 		if($sshlogin->too_fast_remote_login()) {
@@ -2482,7 +2485,7 @@ sub loadavg {
         # load average: 0.76, 1.53, 1.45
         if($uptime_out =~ /load averages?: (\d+.\d+)/) {
             $self->{'loadavg'} = $1;
-            ::debug("New loadavg: ".$self->{'loadavg'});
+            ::debug("New loadavg: ".$self->{'loadavg'}."\n");
         } else {
 	    ::die_bug("loadavg_invalid_content: $uptime_out");
 	}
diff --git a/src/parallel.texi b/src/parallel.texi
index bc3b9ba1..24527930 100644
--- a/src/parallel.texi
+++ b/src/parallel.texi
@@ -1645,6 +1645,14 @@ treated as the relative path to your home dir. This means that if your
 home dir is different on remote computers (e.g. if your login is
 different) the relative path will still be relative to your home dir.
 
+To see the difference try:
+
+@strong{parallel -S server pwd ::: ""}
+
+@strong{parallel --wd . -S server pwd ::: ""}
+
+@strong{parallel --wd ... -S server pwd ::: ""}
+
 @item @strong{--wait}
 @anchor{@strong{--wait}}
 
diff --git a/testsuite/tests-to-run/parallel-local2.sh b/testsuite/tests-to-run/parallel-local2.sh
index 9394fe44..b1f367ba 100644
--- a/testsuite/tests-to-run/parallel-local2.sh
+++ b/testsuite/tests-to-run/parallel-local2.sh
@@ -1,13 +1,36 @@
 #!/bin/bash
 
+highload ()
+{
+  # Force load > #cpus
+  CPUS=$(parallel --number-of-cores)
+  seq 0 0.1 $CPUS | nice nice parallel -j0 timeout 50 burnP6 2>/dev/null &
+  perl -e 'do{$a=`uptime`} while($a=~/average: *(\S+)/ and $1 < '$CPUS')'
+  # Load is now > $CPUS
+  # Kill off burnP6 and the parent parallel
+  kill %1; sleep 0.1; kill %1; killall burnP6; sleep 0.3; kill -9 %1 2>/dev/null
+}
+
+highload 2>/dev/null
+
+cat <<'EOF' | parallel -j0 -k -L1
+echo "bug #38441: CPU usage goes to 100% if load is higher than --load at first job"
+/usr/bin/time -f %e parallel --load 100% true ::: a 2>&1 | 
+  perl -ne '$_ > 1 and print "More than 1 secs wall clock: OK\n"'
+
+/usr/bin/time -f %U parallel --load 100% true ::: a 2>&1 | 
+  perl -ne '$_ < 1 and print "Less than 1 secs user time: OK\n"'
+
 echo '### Test slow arguments generation - https://savannah.gnu.org/bugs/?32834'
-seq 1 3 | parallel -j1 "sleep 2; echo {}" | parallel -kj2 echo
+  seq 1 3 | parallel -j1 "sleep 2; echo {}" | parallel -kj2 echo
 
 echo '### Test too slow spawning'
-killall -9 burnP6 2>/dev/null
-seq `parallel --number-of-cores` | parallel -j200% -N0 timeout -k 25 26 burnP6 &
-sleep 1
-seq 1 1000 |
-stdout nice nice  parallel -s 100 -uj0 true |
-perl -pe '/parallel: Warning: Starting \d+ processes took/ and do {close STDIN; `killall -9 burnP6`; print "OK\n"; exit }'
+# Let the commands below run during high load
+seq `parallel --number-of-cores` | parallel -j200% -N0 timeout -k 25 26 burnP6 & 
+  sleep 1; 
+  seq 1 1000 | stdout nice nice  parallel -s 100 -uj0 true | 
+  perl -pe '/parallel: Warning: Starting \d+ processes took/ and do {close STDIN; `killall -9 burnP6`; print "OK\n"; exit }'
+EOF
+
+# Make sure we got all the burnP6 killed
 killall -9 burnP6 2>/dev/null
diff --git a/testsuite/wanted-results/parallel-local2 b/testsuite/wanted-results/parallel-local2
index 55a91f5e..07ee806d 100644
--- a/testsuite/wanted-results/parallel-local2
+++ b/testsuite/wanted-results/parallel-local2
@@ -1,3 +1,6 @@
+bug #38441: CPU usage goes to 100% if load is higher than --load at first job
+More than 1 secs wall clock: OK
+Less than 1 secs user time: OK
 ### Test slow arguments generation - https://savannah.gnu.org/bugs/?32834
 1
 2