Fix issues in ib loopback benchmark (#369)

Fix several issues in ib loopback benchmark:
* use `--report_gbits` and divide by 8 to get GB/s, previous results are
  MiB/s / 1000
* use the ib_write_bw binary built in third_party instead of system path
* update the metrics name so that different hca indices have same metric

docs/user-tutorial/benchmarks/micro-benchmarks.md

@@ -220,11 +220,11 @@ Measure the InfiniBand loopback verbs bandwidth, performed by
 
 #### Metrics
 
-| Name                                        | Unit             | Description                                                  |
-|---------------------------------------------|------------------|--------------------------------------------------------------|
-| ib-loopback/ib_write_${msg_size}_ib[0-9]_bw | bandwidth (GB/s) | InfiniBand loopback write bandwidth with given message size. |
-| ib-loopback/ib_read_${msg_size}_ib[0-9]_bw  | bandwidth (GB/s) | InfiniBand loopback read bandwidth with given message size.  |
-| ib-loopback/ib_send_${msg_size}_ib[0-9]_bw  | bandwidth (GB/s) | InfiniBand loopback send bandwidth with given message size.  |
+| Name                                | Unit             | Description                                                  |
+|-------------------------------------|------------------|--------------------------------------------------------------|
+| ib-loopback/ib_write_bw_${msg_size} | bandwidth (GB/s) | InfiniBand loopback write bandwidth with given message size. |
+| ib-loopback/ib_read_bw_${msg_size}  | bandwidth (GB/s) | InfiniBand loopback read bandwidth with given message size.  |
+| ib-loopback/ib_send_bw_${msg_size}  | bandwidth (GB/s) | InfiniBand loopback send bandwidth with given message size.  |
 
 ### `nccl-bw` / `rccl-bw`
 

superbench/benchmarks/micro_benchmarks/ib_loopback_performance.py

@@ -161,12 +161,13 @@ def _preprocess(self):
                         server_core = int(numa_cores[-1])
                         client_core = int(numa_cores[-2])
                     command += ' ' + str(server_core) + ' ' + str(client_core)
-                    command += ' ' + self.__support_ib_commands[ib_command]
+                    command += ' ' + os.path.join(self._args.bin_dir, self.__support_ib_commands[ib_command])
                     command += command_mode + ' -F'
                     command += ' --iters=' + str(self._args.iters)
                     command += ' -d ' + network.get_ib_devices()[self._args.ib_index].split(':')[0]
                     command += ' -p ' + str(network.get_free_port())
                     command += ' -x ' + str(self._args.gid_index)
+                    command += ' --report_gbits'
                     self._commands.append(command)
                 except BaseException as e:
                     self._result.set_return_code(ReturnCode.MICROBENCHMARK_DEVICE_GETTING_FAILURE)
@@ -197,13 +198,13 @@ def _process_raw_result(self, cmd_idx, raw_output):
         metric_set = set()
         for line in content:
             try:
-                values = list(filter(None, line.split(' ')))
+                values = list(filter(None, line.split()))
                 if len(values) != 5:
                     continue
                 # Extract value from the line
                 size = int(values[0])
-                avg_bw = float(values[-2]) / 1000
-                metric = 'ib_{}_{}_ib{}_bw'.format(self._args.commands[cmd_idx], size, str(self._args.ib_index))
+                avg_bw = float(values[-2]) / 8.0
+                metric = f'{self.__support_ib_commands[self._args.commands[cmd_idx]]}_{size}:{self._args.ib_index}'
                 # Filter useless value in client output
                 if metric not in metric_set:
                     metric_set.add(metric)

tests/benchmarks/micro_benchmarks/test_ib_loopback_performance.py

@@ -76,7 +76,8 @@ def test_ib_loopback_all_sizes(self, raw_output, mock_ib_devices, mock_numa_core
         ret = benchmark._preprocess()
         assert (ret)
 
-        expect_command = 'run_perftest_loopback 3 1 ib_write_bw -a -F --iters=2000 -d mlx5_0 -p 10000 -x 0'
+        expect_command = 'run_perftest_loopback 3 1 ' + benchmark._args.bin_dir + \
+            '/ib_write_bw -a -F --iters=2000 -d mlx5_0 -p 10000 -x 0 --report_gbits'
         command = benchmark._bin_name + benchmark._commands[0].split(benchmark._bin_name)[1]
         assert (command == expect_command)
 
@@ -87,7 +88,7 @@ def test_ib_loopback_all_sizes(self, raw_output, mock_ib_devices, mock_numa_core
         metric_list = []
         for ib_command in benchmark._args.commands:
             for size in ['8388608', '4194304', '1024', '2']:
-                metric = 'ib_{}_{}_ib{}_bw'.format(ib_command, size, str(benchmark._args.ib_index))
+                metric = 'ib_{}_bw_{}:{}'.format(ib_command, size, str(benchmark._args.ib_index))
                 metric_list.append(metric)
         for metric in metric_list:
             assert (metric in benchmark.result)
@@ -145,7 +146,8 @@ def test_ib_loopback_8M_size(self, raw_output, mock_ib_devices, mock_numa_cores,
         ret = benchmark._preprocess()
         assert (ret)
 
-        expect_command = 'run_perftest_loopback 3 1 ib_write_bw -s 8388608 -F --iters=2000 -d mlx5_0 -p 10000 -x 0'
+        expect_command = 'run_perftest_loopback 3 1 ' + benchmark._args.bin_dir + \
+            '/ib_write_bw -s 8388608 -F --iters=2000 -d mlx5_0 -p 10000 -x 0 --report_gbits'
         command = benchmark._bin_name + benchmark._commands[0].split(benchmark._bin_name)[1]
         assert (command == expect_command)
 
@@ -155,7 +157,7 @@ def test_ib_loopback_8M_size(self, raw_output, mock_ib_devices, mock_numa_cores,
         # Positive case - valid raw output.
         metric_list = []
         for ib_command in benchmark._args.commands:
-            metric = 'ib_{}_8388608_ib{}_bw'.format(ib_command, str(benchmark._args.ib_index))
+            metric = 'ib_{}_bw_8388608:{}'.format(ib_command, str(benchmark._args.ib_index))
             metric_list.append(metric)
         for metric in metric_list:
             assert (metric in benchmark.result)