diff --git a/Testing/perf/na/na_bw_get.c b/Testing/perf/na/na_bw_get.c
index 5e7612ab..0eb4444b 100644
--- a/Testing/perf/na/na_bw_get.c
+++ b/Testing/perf/na/na_bw_get.c
@@ -35,9 +35,6 @@ na_perf_run(struct na_perf_info *info, size_t buf_size, size_t skip)
     na_return_t ret;
     size_t i, j;
 
-    if (info->na_test_info.mpi_comm_size > 1)
-        NA_Test_barrier(&info->na_test_info);
-
     /* Actual benchmark */
     for (i = 0; i < skip + (size_t) info->na_test_info.loop; i++) {
         struct na_perf_rma_info rma_info = {.request = info->request,
@@ -96,13 +93,9 @@ na_perf_run(struct na_perf_info *info, size_t buf_size, size_t skip)
         }
     }
 
-    if (info->na_test_info.mpi_comm_size > 1)
-        NA_Test_barrier(&info->na_test_info);
-
     hg_time_get_current(&t2);
 
-    if (info->na_test_info.mpi_comm_rank == 0)
-        na_perf_print_bw(info, buf_size, hg_time_subtract(t2, t1));
+    na_perf_print_bw(info, buf_size, hg_time_subtract(t2, t1));
 
     return NA_SUCCESS;
 
@@ -129,8 +122,7 @@ main(int argc, char *argv[])
         "na_perf_mem_handle_recv() failed (%s)", NA_Error_to_string(na_ret));
 
     /* Header info */
-    if (info.na_test_info.mpi_comm_rank == 0)
-        na_perf_print_header_bw(&info, BENCHMARK_NAME);
+    na_perf_print_header_bw(&info, BENCHMARK_NAME);
 
     /* Msg with different sizes */
     for (size = info.rma_size_min; size <= info.rma_size_max; size *= 2) {
@@ -142,8 +134,7 @@ main(int argc, char *argv[])
     }
 
     /* Finalize interface */
-    if (info.na_test_info.mpi_comm_rank == 0)
-        na_perf_send_finalize(&info);
+    na_perf_send_finalize(&info);
 
     na_perf_cleanup(&info);
 
diff --git a/Testing/perf/na/na_bw_put.c b/Testing/perf/na/na_bw_put.c
index 7507dd41..479c4a71 100644
--- a/Testing/perf/na/na_bw_put.c
+++ b/Testing/perf/na/na_bw_put.c
@@ -35,9 +35,6 @@ na_perf_run(struct na_perf_info *info, size_t buf_size, size_t skip)
     na_return_t ret;
     size_t i, j;
 
-    if (info->na_test_info.mpi_comm_size > 1)
-        NA_Test_barrier(&info->na_test_info);
-
     /* Actual benchmark */
     for (i = 0; i < skip + (size_t) info->na_test_info.loop; i++) {
         struct na_perf_rma_info rma_info = {.request = info->request,
@@ -112,13 +109,9 @@ na_perf_run(struct na_perf_info *info, size_t buf_size, size_t skip)
         }
     }
 
-    if (info->na_test_info.mpi_comm_size > 1)
-        NA_Test_barrier(&info->na_test_info);
-
     hg_time_get_current(&t2);
 
-    if (info->na_test_info.mpi_comm_rank == 0)
-        na_perf_print_bw(info, buf_size, hg_time_subtract(t2, t1));
+    na_perf_print_bw(info, buf_size, hg_time_subtract(t2, t1));
 
     return NA_SUCCESS;
 
@@ -150,8 +143,7 @@ main(int argc, char *argv[])
         "na_perf_mem_handle_recv() failed (%s)", NA_Error_to_string(na_ret));
 
     /* Header info */
-    if (info.na_test_info.mpi_comm_rank == 0)
-        na_perf_print_header_bw(&info, BENCHMARK_NAME);
+    na_perf_print_header_bw(&info, BENCHMARK_NAME);
 
     /* Msg with different sizes */
     for (size = info.rma_size_min; size <= info.rma_size_max; size *= 2) {
@@ -163,8 +155,7 @@ main(int argc, char *argv[])
     }
 
     /* Finalize interface */
-    if (info.na_test_info.mpi_comm_rank == 0)
-        na_perf_send_finalize(&info);
+    na_perf_send_finalize(&info);
 
     na_perf_cleanup(&info);
 
diff --git a/Testing/perf/na/na_lat.c b/Testing/perf/na/na_lat.c
index 4569064c..b24f1c33 100644
--- a/Testing/perf/na/na_lat.c
+++ b/Testing/perf/na/na_lat.c
@@ -62,11 +62,8 @@ na_perf_run(struct na_perf_info *info, size_t buf_size, size_t skip)
 
     /* Actual benchmark */
     for (i = 0; i < skip + (size_t) info->na_test_info.loop; i++) {
-        if (i == skip) {
-            if (info->na_test_info.mpi_comm_size > 1)
-                NA_Test_barrier(&info->na_test_info);
+        if (i == skip)
             hg_time_get_current(&t1);
-        }
 
         hg_request_reset(info->request);
 
@@ -105,13 +102,9 @@ na_perf_run(struct na_perf_info *info, size_t buf_size, size_t skip)
         }
     }
 
-    if (info->na_test_info.mpi_comm_size > 1)
-        NA_Test_barrier(&info->na_test_info);
-
     hg_time_get_current(&t2);
 
-    if (info->na_test_info.mpi_comm_rank == 0)
-        na_perf_print_lat(info, buf_size, hg_time_subtract(t2, t1));
+    na_perf_print_lat(info, buf_size, hg_time_subtract(t2, t1));
 
     return NA_SUCCESS;
 
@@ -135,15 +128,13 @@ main(int argc, char *argv[])
     /* Init data */
     na_perf_init_data(info.msg_unexp_buf, info.msg_unexp_size_max,
         info.msg_unexp_header_size);
-    if (info.na_test_info.mpi_comm_rank == 0)
-        na_perf_send_init(&info);
+    na_perf_send_init(&info);
 
     min_size =
         (info.msg_unexp_header_size > 0) ? info.msg_unexp_header_size : 1;
 
     /* Header info */
-    if (info.na_test_info.mpi_comm_rank == 0)
-        na_perf_print_header_lat(&info, BENCHMARK_NAME, min_size);
+    na_perf_print_header_lat(&info, BENCHMARK_NAME, min_size);
 
     /* Msg with different sizes */
     for (size = min_size; size <= info.msg_unexp_size_max; size *= 2) {
@@ -155,8 +146,7 @@ main(int argc, char *argv[])
     }
 
     /* Finalize interface */
-    if (info.na_test_info.mpi_comm_rank == 0)
-        na_perf_send_finalize(&info);
+    na_perf_send_finalize(&info);
 
     na_perf_cleanup(&info);
 
diff --git a/Testing/perf/na/na_perf.c b/Testing/perf/na/na_perf.c
index a27586f9..faf2633f 100644
--- a/Testing/perf/na/na_perf.c
+++ b/Testing/perf/na/na_perf.c
@@ -130,6 +130,10 @@ na_perf_init(int argc, char *argv[], bool listen, struct na_perf_info *info)
         error, ret, "NA_Test_init() failed (%s)", NA_Error_to_string(ret));
     info->na_class = info->na_test_info.na_class;
 
+    /* Parallel is not supported */
+    NA_TEST_CHECK_ERROR(info->na_test_info.mpi_comm_size > 1, error, ret,
+        NA_OPNOTSUPPORTED, "Not a parallel test");
+
     /* Multi-recv */
     multi_recv =
         (listen && NA_Has_opt_feature(info->na_class, NA_OPT_MULTI_RECV) &&
@@ -398,10 +402,9 @@ void
 na_perf_print_lat(const struct na_perf_info *info, size_t buf_size, hg_time_t t)
 {
     double msg_lat;
-    size_t loop = (size_t) info->na_test_info.loop,
-           mpi_comm_size = (size_t) info->na_test_info.mpi_comm_size;
+    size_t loop = (size_t) info->na_test_info.loop;
 
-    msg_lat = hg_time_to_double(t) * 1e6 / (double) (loop * 2 * mpi_comm_size);
+    msg_lat = hg_time_to_double(t) * 1e6 / (double) (loop * 2);
 
     printf("%-*zu%*.*f\n", 10, buf_size, NWIDTH, NDIGITS, msg_lat);
 }
@@ -432,14 +435,11 @@ void
 na_perf_print_bw(const struct na_perf_info *info, size_t buf_size, hg_time_t t)
 {
     size_t loop = (size_t) info->na_test_info.loop,
-           mpi_comm_size = (size_t) info->na_test_info.mpi_comm_size,
            buf_count = (size_t) info->rma_count;
     double avg_time, avg_bw;
 
-    avg_time = hg_time_to_double(t) * 1e6 /
-               (double) (loop * mpi_comm_size * buf_count);
-    avg_bw = (double) (buf_size * loop * mpi_comm_size * buf_count) /
-             hg_time_to_double(t);
+    avg_time = hg_time_to_double(t) * 1e6 / (double) (loop * buf_count);
+    avg_bw = (double) (buf_size * loop * buf_count) / hg_time_to_double(t);
 
     if (info->na_test_info.mbps)
         avg_bw /= 1e6; /* MB/s, matches OSU benchmarks */