diff --git a/projects/rocshmem/tests/functional_tests/tester.cpp b/projects/rocshmem/tests/functional_tests/tester.cpp
index e802c0b289..8d081d976b 100644
--- a/projects/rocshmem/tests/functional_tests/tester.cpp
+++ b/projects/rocshmem/tests/functional_tests/tester.cpp
@@ -82,66 +82,66 @@ std::vector<Tester*> Tester::create(TesterArguments args) {
   CHECK_HIP(hipGetDeviceProperties(&deviceProps, device_id));
   numWarps = (args.wg_size - 1) / deviceProps.warpSize + 1;
 
-  if (rank == 0) std::cout << "*** Creating Test: ";
+  if (rank == 0) std::cout << "### Creating Test: ";
 
   TestType type = (TestType)args.algorithm;
 
   switch (type) {
     case InitTestType:
-      if (rank == 0) std::cout << "Init ***" << std::endl;
+      if (rank == 0) std::cout << "Init ###" << std::endl;
       testers.push_back(new EmptyTester(args));
       return testers;
     case GetTestType:
-      if (rank == 0) std::cout << "Blocking Gets***" << std::endl;
+      if (rank == 0) std::cout << "Blocking Gets ###" << std::endl;
       testers.push_back(new PrimitiveTester(args));
       return testers;
     case GetNBITestType:
-      if (rank == 0) std::cout << "Non-Blocking Gets***" << std::endl;
+      if (rank == 0) std::cout << "Non-Blocking Gets ###" << std::endl;
       testers.push_back(new PrimitiveTester(args));
       return testers;
     case PutTestType:
-      if (rank == 0) std::cout << "Blocking Puts***" << std::endl;
+      if (rank == 0) std::cout << "Blocking Puts ###" << std::endl;
       testers.push_back(new PrimitiveTester(args));
       return testers;
     case PutNBITestType:
-      if (rank == 0) std::cout << "Non-Blocking Puts***" << std::endl;
+      if (rank == 0) std::cout << "Non-Blocking Puts ###" << std::endl;
       testers.push_back(new PrimitiveTester(args));
       return testers;
     case TeamCtxInfraTestType:
-      if (rank == 0) std::cout << "Team Ctx Infra test***" << std::endl;
+      if (rank == 0) std::cout << "Team Ctx Infra test ###" << std::endl;
       testers.push_back(new TeamCtxInfraTester(args));
       return testers;
     case TeamCtxGetTestType:
-      if (rank == 0) std::cout << "Blocking Team Ctx Gets***" << std::endl;
+      if (rank == 0) std::cout << "Blocking Team Ctx Gets ###" << std::endl;
       testers.push_back(new TeamCtxPrimitiveTester(args));
       return testers;
     case TeamCtxGetNBITestType:
-      if (rank == 0) std::cout << "Non-Blocking Team Ctx Gets***" << std::endl;
+      if (rank == 0) std::cout << "Non-Blocking Team Ctx Gets ###" << std::endl;
       testers.push_back(new TeamCtxPrimitiveTester(args));
       return testers;
     case TeamCtxPutTestType:
-      if (rank == 0) std::cout << "Blocking Team Ctx Puts***" << std::endl;
+      if (rank == 0) std::cout << "Blocking Team Ctx Puts ###" << std::endl;
       testers.push_back(new TeamCtxPrimitiveTester(args));
       return testers;
     case TeamCtxPutNBITestType:
-      if (rank == 0) std::cout << "Non-Blocking Team Ctx Puts***" << std::endl;
+      if (rank == 0) std::cout << "Non-Blocking Team Ctx Puts ###" << std::endl;
       testers.push_back(new TeamCtxPrimitiveTester(args));
       return testers;
     case PTestType:
-      if (rank == 0) std::cout << "P Test***" << std::endl;
+      if (rank == 0) std::cout << "P Test ###" << std::endl;
       testers.push_back(new PrimitiveTester(args));
       return testers;
     case GTestType:
-      if (rank == 0) std::cout << "G Test***" << std::endl;
+      if (rank == 0) std::cout << "G Test ###" << std::endl;
       testers.push_back(new PrimitiveTester(args));
       return testers;
     case GetSwarmTestType:
-      if (rank == 0) std::cout << "Get Swarm***" << std::endl;
+      if (rank == 0) std::cout << "Get Swarm ###" << std::endl;
       testers.push_back(new GetSwarmTester(args));
       return testers;
     case TeamReductionTestType:
       if (rank == 0)
-        std::cout << "All-to-All Team-based Reduction***" << std::endl;
+        std::cout << "All-to-All Team-based Reduction ###" << std::endl;
       testers.push_back(new TeamReductionTester<float, ROC_SHMEM_SUM>(
           args,
           [](float& f1, float& f2) {
@@ -157,7 +157,7 @@ std::vector<Tester*> Tester::create(TesterArguments args) {
           }));
       return testers;
     case ReductionTestType:
-      if (rank == 0) std::cout << "All-to-All Reduction***" << std::endl;
+      if (rank == 0) std::cout << "All-to-All Reduction ###" << std::endl;
 
       testers.push_back(new ReductionTester<float, ROC_SHMEM_SUM>(
           args,
@@ -245,7 +245,7 @@ std::vector<Tester*> Tester::create(TesterArguments args) {
       return testers;
     case BroadcastTestType:
       if (rank == 0) {
-        std::cout << "Broadcast Test***" << std::endl;
+        std::cout << "Broadcast Test ###" << std::endl;
       }
       testers.push_back(new BroadcastTester<long>(
           args,
@@ -282,7 +282,7 @@ std::vector<Tester*> Tester::create(TesterArguments args) {
       return testers;
     case TeamBroadcastTestType:
       if (rank == 0) {
-        std::cout << "Team Broadcast Test***" << std::endl;
+        std::cout << "Team Broadcast Test ###" << std::endl;
       }
       testers.push_back(new TeamBroadcastTester<long>(
           args,
@@ -319,7 +319,7 @@ std::vector<Tester*> Tester::create(TesterArguments args) {
       return testers;
     case AllToAllTestType:
       if (rank == 0) {
-        std::cout << "Alltoall Test***" << std::endl;
+        std::cout << "Alltoall Test ###" << std::endl;
       }
       testers.push_back(new AlltoallTester<int64_t>(
           args,
@@ -345,7 +345,7 @@ std::vector<Tester*> Tester::create(TesterArguments args) {
       return testers;
     case FCollectTestType:
       if (rank == 0) {
-        std::cout << "Fcollect Test***" << std::endl;
+        std::cout << "Fcollect Test ###" << std::endl;
       }
       testers.push_back(new FcollectTester<int64_t>(
           args,
@@ -365,188 +365,188 @@ std::vector<Tester*> Tester::create(TesterArguments args) {
           }));
       return testers;
     case AMO_FAddTestType:
-      if (rank == 0) std::cout << "AMO Fetch_Add***" << std::endl;
+      if (rank == 0) std::cout << "AMO Fetch_Add ###" << std::endl;
       testers.push_back(new AMOStandardTester<long long>(args));
       testers.push_back(new AMOStandardTester<long>(args));
       testers.push_back(new AMOStandardTester<int>(args));
       return testers;
     case AMO_FIncTestType:
-      if (rank == 0) std::cout << "AMO Fetch_Inc***" << std::endl;
+      if (rank == 0) std::cout << "AMO Fetch_Inc ###" << std::endl;
       testers.push_back(new AMOStandardTester<long long>(args));
       testers.push_back(new AMOStandardTester<long>(args));
       testers.push_back(new AMOStandardTester<int>(args));
       return testers;
     case AMO_FetchTestType:
-      if (rank == 0) std::cout << "AMO Fetch***" << std::endl;
+      if (rank == 0) std::cout << "AMO Fetch ###" << std::endl;
       testers.push_back(new AMOExtendedTester<long long>(args));
       testers.push_back(new AMOExtendedTester<long>(args));
       testers.push_back(new AMOExtendedTester<int>(args));
       return testers;
     case AMO_FCswapTestType:
-      if (rank == 0) std::cout << "AMO Fetch_CSWAP***" << std::endl;
+      if (rank == 0) std::cout << "AMO Fetch_CSWAP ###" << std::endl;
       testers.push_back(new AMOStandardTester<long long>(args));
       testers.push_back(new AMOStandardTester<long>(args));
       testers.push_back(new AMOStandardTester<int>(args));
       return testers;
     case AMO_AddTestType:
-      if (rank == 0) std::cout << "AMO Add***" << std::endl;
+      if (rank == 0) std::cout << "AMO Add ###" << std::endl;
       testers.push_back(new AMOStandardTester<long long>(args));
       testers.push_back(new AMOStandardTester<long>(args));
       testers.push_back(new AMOStandardTester<int>(args));
       return testers;
     case AMO_SetTestType:
-      if (rank == 0) std::cout << "AMO Set***" << std::endl;
+      if (rank == 0) std::cout << "AMO Set ###" << std::endl;
       testers.push_back(new AMOExtendedTester<long long>(args));
       testers.push_back(new AMOExtendedTester<long>(args));
       testers.push_back(new AMOExtendedTester<int>(args));
       return testers;
     case AMO_SwapTestType:
-      if (rank == 0) std::cout << "AMO Swap***" << std::endl;
+      if (rank == 0) std::cout << "AMO Swap ###" << std::endl;
       testers.push_back(new AMOExtendedTester<long long>(args));
       testers.push_back(new AMOExtendedTester<long>(args));
       testers.push_back(new AMOExtendedTester<int>(args));
       return testers;
     case AMO_FetchAndTestType:
-      if (rank == 0) std::cout << "AMO Fetch And***" << std::endl;
+      if (rank == 0) std::cout << "AMO Fetch And ###" << std::endl;
       testers.push_back(new AMOBitwiseTester<unsigned long long>(args));
       testers.push_back(new AMOBitwiseTester<unsigned long>(args));
       testers.push_back(new AMOBitwiseTester<unsigned int>(args));
       return testers;
     case AMO_AndTestType:
-      if (rank == 0) std::cout << "AMO And***" << std::endl;
+      if (rank == 0) std::cout << "AMO And ###" << std::endl;
       testers.push_back(new AMOBitwiseTester<unsigned long long>(args));
       testers.push_back(new AMOBitwiseTester<unsigned long>(args));
       testers.push_back(new AMOBitwiseTester<unsigned int>(args));
       return testers;
     case AMO_FetchOrTestType:
-      if (rank == 0) std::cout << "AMO Fetch Or***" << std::endl;
+      if (rank == 0) std::cout << "AMO Fetch Or ###" << std::endl;
       testers.push_back(new AMOBitwiseTester<unsigned long long>(args));
       testers.push_back(new AMOBitwiseTester<unsigned long>(args));
       testers.push_back(new AMOBitwiseTester<unsigned int>(args));
       return testers;
     case AMO_OrTestType:
-      if (rank == 0) std::cout << "AMO Or***" << std::endl;
+      if (rank == 0) std::cout << "AMO Or ###" << std::endl;
       testers.push_back(new AMOBitwiseTester<unsigned long long>(args));
       testers.push_back(new AMOBitwiseTester<unsigned long>(args));
       testers.push_back(new AMOBitwiseTester<unsigned int>(args));
       return testers;
     case AMO_FetchXorTestType:
-      if (rank == 0) std::cout << "AMO Fetch Xor***" << std::endl;
+      if (rank == 0) std::cout << "AMO Fetch Xor ###" << std::endl;
       testers.push_back(new AMOBitwiseTester<unsigned long long>(args));
       testers.push_back(new AMOBitwiseTester<unsigned long>(args));
       testers.push_back(new AMOBitwiseTester<unsigned int>(args));
       return testers;
     case AMO_XorTestType:
-      if (rank == 0) std::cout << "AMO Xor***" << std::endl;
+      if (rank == 0) std::cout << "AMO Xor ###" << std::endl;
       testers.push_back(new AMOBitwiseTester<unsigned long long>(args));
       testers.push_back(new AMOBitwiseTester<unsigned long>(args));
       testers.push_back(new AMOBitwiseTester<unsigned int>(args));
       return testers;
     case AMO_IncTestType:
-      if (rank == 0) std::cout << "AMO Inc***" << std::endl;
+      if (rank == 0) std::cout << "AMO Inc ###" << std::endl;
       testers.push_back(new AMOStandardTester<long long>(args));
       testers.push_back(new AMOStandardTester<long>(args));
       testers.push_back(new AMOStandardTester<int>(args));
       return testers;
     case PingPongTestType:
-      if (rank == 0) std::cout << "PingPong***" << std::endl;
+      if (rank == 0) std::cout << "PingPong ###" << std::endl;
       testers.push_back(new PingPongTester(args));
       return testers;
     case PingAllTestType:
-      if (rank == 0) std::cout << "PingAll***" << std::endl;
+      if (rank == 0) std::cout << "PingAll ###" << std::endl;
       testers.push_back(new PingAllTester(args));
       return testers;
     case BarrierAllTestType:
-      if (rank == 0) std::cout << "Barrier_All***" << std::endl;
+      if (rank == 0) std::cout << "Barrier_All ###" << std::endl;
       testers.push_back(new BarrierAllTester(args));
       return testers;
     case SyncAllTestType:
-      if (rank == 0) std::cout << "SyncAll***" << std::endl;
+      if (rank == 0) std::cout << "SyncAll ###" << std::endl;
       testers.push_back(new SyncTester(args));
       return testers;
     case SyncTestType:
-      if (rank == 0) std::cout << "Sync***" << std::endl;
+      if (rank == 0) std::cout << "Sync ###" << std::endl;
       testers.push_back(new SyncTester(args));
       return testers;
     case RandomAccessTestType:
-      if (rank == 0) std::cout << "Random_Access***" << std::endl;
+      if (rank == 0) std::cout << "Random_Access ###" << std::endl;
       testers.push_back(new RandomAccessTester(args));
       return testers;
     case ShmemPtrTestType:
-      if (rank == 0) std::cout << "Shmem_Ptr***" << std::endl;
+      if (rank == 0) std::cout << "Shmem_Ptr ###" << std::endl;
       testers.push_back(new ShmemPtrTester(args));
       return testers;
     case WGGetTestType:
       if (rank == 0) {
         if (args.num_wgs > 1)
-          std::cout << "Tiled Blocking WG level Gets***" << std::endl;
-        else std::cout << "Blocking WG level Gets***" << std::endl;
+          std::cout << "Tiled Blocking WG level Gets ###" << std::endl;
+        else std::cout << "Blocking WG level Gets ###" << std::endl;
       }
       testers.push_back(new ExtendedPrimitiveTester(args));
       return testers;
     case WGGetNBITestType:
       if (rank == 0) {
         if (args.num_wgs > 1)
-          std::cout << "Tiled Non-Blocking WG level Gets***" << std::endl;
-        else std::cout << "Non-Blocking WG level Gets***" << std::endl;
+          std::cout << "Tiled Non-Blocking WG level Gets ###" << std::endl;
+        else std::cout << "Non-Blocking WG level Gets ###" << std::endl;
       }
       testers.push_back(new ExtendedPrimitiveTester(args));
       return testers;
     case WGPutTestType:
       if (rank == 0) {
         if (args.num_wgs > 1)
-          std::cout << "Tiled Blocking WG level Puts***" << std::endl;
-        else std::cout << "Blocking WG level Puts***" << std::endl;
+          std::cout << "Tiled Blocking WG level Puts ###" << std::endl;
+        else std::cout << "Blocking WG level Puts ###" << std::endl;
       }
       testers.push_back(new ExtendedPrimitiveTester(args));
       return testers;
     case WGPutNBITestType:
       if (rank == 0) {
         if (args.num_wgs > 1)
-          std::cout << "Tiled Non-Blocking WG level Puts***" << std::endl;
-        else std::cout << "Non-Blocking WG level Puts***" << std::endl;
+          std::cout << "Tiled Non-Blocking WG level Puts ###" << std::endl;
+        else std::cout << "Non-Blocking WG level Puts ###" << std::endl;
       }
       testers.push_back(new ExtendedPrimitiveTester(args));
       return testers;
     case PutNBIMRTestType:
       if (rank == 0)
-        std::cout << "Non-Blocking Put message rate***" << std::endl;
+        std::cout << "Non-Blocking Put message rate ###" << std::endl;
       testers.push_back(new PrimitiveMRTester(args));
       return testers;
     case WAVEGetTestType:
       if (rank == 0) {
         if (args.num_wgs > 1 || numWarps > 1)
-          std::cout << "Tiled Blocking WAVE level Gets***" << std::endl;
-        else std::cout << "Blocking WAVE level Gets***" << std::endl;
+          std::cout << "Tiled Blocking WAVE level Gets ###" << std::endl;
+        else std::cout << "Blocking WAVE level Gets ###" << std::endl;
       }
       testers.push_back(new WaveLevelPrimitiveTester(args));
       return testers;
     case WAVEGetNBITestType:
       if (rank == 0) {
         if (args.num_wgs > 1 || numWarps > 1)
-          std::cout << "Tiled Non-Blocking WAVE level Gets***" << std::endl;
-        else std::cout << "Non-Blocking WAVE level Gets***" << std::endl;
+          std::cout << "Tiled Non-Blocking WAVE level Gets ###" << std::endl;
+        else std::cout << "Non-Blocking WAVE level Gets ###" << std::endl;
       }
       testers.push_back(new WaveLevelPrimitiveTester(args));
       return testers;
     case WAVEPutTestType:
       if (rank == 0) {
         if (args.num_wgs > 1 || numWarps > 1)
-          std::cout << "Tiled Blocking WAVE level Puts***" << std::endl;
-        else std::cout << "Blocking WAVE level Puts***" << std::endl;
+          std::cout << "Tiled Blocking WAVE level Puts ###" << std::endl;
+        else std::cout << "Blocking WAVE level Puts ###" << std::endl;
       }
       testers.push_back(new WaveLevelPrimitiveTester(args));
       return testers;
     case WAVEPutNBITestType:
       if (rank == 0) {
         if (args.num_wgs > 1 || numWarps > 1)
-          std::cout << "Tiled Non-Blocking WAVE level Puts***" << std::endl;
-        else std::cout << "Non-Blocking WAVE level Puts***" << std::endl;
+          std::cout << "Tiled Non-Blocking WAVE level Puts ###" << std::endl;
+        else std::cout << "Non-Blocking WAVE level Puts ###" << std::endl;
       }
       testers.push_back(new WaveLevelPrimitiveTester(args));
       return testers;
     default:
-      if (rank == 0) std::cout << "Unknown***" << std::endl;
+      if (rank == 0) std::cout << "Unknown ###" << std::endl;
       testers.push_back(new PrimitiveTester(args));
       return testers;
   }
@@ -688,14 +688,20 @@ void Tester::print(uint64_t size) {
   int field_width = 20;
   int float_precision = 2;
 
-  printf("\n##### Message Size %lu #####\n", size);
+  if (_print_header) {
+    printf("%-*s%*s%*s%*s",
+           10, "# Size (B)",
+           field_width, "Latency (us)",
+           field_width, "Bandwidth (GB/s)",
+           field_width + 1, "Msg Rate (Msg/s)\n");
+    _print_header = 0;
+  }
 
-  printf("%*s%*s%*s\n", field_width + 1, "Latency AVG (us)", field_width + 1,
-         "Bandwidth (GB/s)", field_width + 1, "Avg Message rate (Messages/s)");
-
-  printf("%*.*f %*.*f %*.*f\n", field_width, float_precision, latency_avg,
-         field_width, float_precision, bandwidth_avg_gbs, field_width,
-         float_precision, avg_msg_rate);
+  printf("%-*lu%*.*f%*.*f%*.*f\n",
+         10, size,
+         field_width, float_precision, latency_avg,
+         field_width, float_precision, bandwidth_avg_gbs,
+         field_width, float_precision, avg_msg_rate);
 
   fflush(stdout);
 }
diff --git a/projects/rocshmem/tests/functional_tests/tester.hpp b/projects/rocshmem/tests/functional_tests/tester.hpp
index e0482868e8..ad639a63ae 100644
--- a/projects/rocshmem/tests/functional_tests/tester.hpp
+++ b/projects/rocshmem/tests/functional_tests/tester.hpp
@@ -132,6 +132,7 @@ class Tester {
   uint64_t *timer = nullptr;
 
  private:
+  bool _print_header = 1;
   void print(uint64_t size);
 
   void barrier();