diff --git a/tools/docker/scripts/plot_performance.py b/tools/docker/scripts/plot_performance.py
index 896ee0267b..9df1b7d62c 100755
--- a/tools/docker/scripts/plot_performance.py
+++ b/tools/docker/scripts/plot_performance.py
@@ -8,29 +8,39 @@ from collections import OrderedDict
 
 # ======================================================================================
 def main():
-    if len(sys.argv) != 5:
-        print("Usage: plot_performance.py <title> <label> <output_omp> <output_mpi>")
+    if len(sys.argv) != 6:
+        print(
+            "Usage: plot_performance.py <nprocs> <title> <label> <output_omp> <output_mpi>"
+        )
         sys.exit(1)
-    title, label, output_omp, output_mpi = sys.argv[1:]
+    nprocs, title, label, output_omp, output_mpi = sys.argv[1:]
 
     timings_omp = parse_timings(output_omp)
     timings_mpi = parse_timings(output_mpi)
-    routines = list(timings_omp.keys())[:6]
+    routines = list(set(list(timings_omp.keys())[:6] + list(timings_mpi.keys())[:6]))
+    routines.remove("total")
+    routines.sort(reverse=True, key=lambda r: timings_omp.get(r, 0.0))
 
-    full_title = f"Timings of {title} with OpenMP"
-    plot = f"{label}_timings"
+    sum_routines_omp = sum([timings_omp.get(r, 0.0) for r in routines])
+    sum_routines_mpi = sum([timings_mpi.get(r, 0.0) for r in routines])
+    timings_omp["rest"] = timings_omp["total"] - sum_routines_omp
+    timings_mpi["rest"] = timings_mpi["total"] - sum_routines_mpi
+
+    full_title = f"Timings of {title} with {nprocs} OpenMP Threads"
+    plot = f"{label}_timings_{nprocs}omp"
     print(f'Plot: name="{plot}", title="{full_title}", ylabel="time [s]"')
-    for r in routines:
+    for r in ["rest"] + routines:
         t = timings_omp[r]
         print(f'PlotPoint: plot="{plot}", name="{r}", label="{r}", y={t}, yerr=0.0')
     print("")
 
-    full_title = f"Parallel Efficiency of {title}"
-    plot = f"{label}_scaling"
-    print(f'Plot: name="{plot}", title="{full_title}", ylabel="MPI / OMP time"')
-    for r in routines:
-        e = timings_mpi[r] / timings_omp[r]
-        print(f'PlotPoint: plot="{plot}", name="{r}", label="{r}", y={e}, yerr=0.0')
+    full_title = f"Timings of {title} with {nprocs} MPI Ranks"
+    plot = f"{label}_timings_{nprocs}mpi"
+    print(f'Plot: name="{plot}", title="{full_title}", ylabel="time [s]"')
+    for r in ["rest"] + routines:
+        t = timings_mpi[r]
+        print(f'PlotPoint: plot="{plot}", name="{r}", label="{r}", y={t}, yerr=0.0')
+    print("")
 
 
 # ======================================================================================
@@ -38,7 +48,9 @@ def parse_timings(out_fn):
     output = open(out_fn, encoding="utf8").read()
 
     pattern = r"\n( -+\n - +-\n - +T I M I N G +-\n([^\n]*\n){4}.*? -+)\n"
-    report_lines = re.search(pattern, output, re.DOTALL).group(1).split("\n")[7:-1]
+    match = re.search(pattern, output, re.DOTALL)
+    report_lines = match.group(1).split("\n")[7:-1]
+    print("\nFrom {}:\n{}\n".format(out_fn, match.group(0).strip()))
 
     # Extract average self time.
     timings = {}
diff --git a/tools/docker/scripts/test_performance.sh b/tools/docker/scripts/test_performance.sh
index e2ed2d4320..eaf8967b10 100755
--- a/tools/docker/scripts/test_performance.sh
+++ b/tools/docker/scripts/test_performance.sh
@@ -60,7 +60,7 @@ for INPUT in "${BENCHMARKS[@]}"; do
   run_benchmark 32 1 "${INPUT_BASENAME}" "${OUTPUT_OMP}"
   cd ..
   echo ""
-  /workspace/plot_performance.py "${LABEL}" "${LABEL}" "${OUTPUT_OMP}" "${OUTPUT_MPI}"
+  /workspace/plot_performance.py 32 "${LABEL}" "${LABEL}" "${OUTPUT_OMP}" "${OUTPUT_MPI}"
   echo ""
 done