DAOS-14654 test: simplify ior_per_rank.py (#13346) (#13650)

- Only run with transfer size 1M
- Reduce stonewall to 15s

Signed-off-by: Dalton Bohning <[email protected]>

src/tests/ftest/deployment/ior_per_rank.py

@@ -5,6 +5,7 @@
 """
 
 from avocado.core.exceptions import TestFail
+from ClusterShell.NodeSet import NodeSet
 from general_utils import DaosTestError, percent_change
 from ior_test_base import IorTestBase
 from ior_utils import IorCommand, IorMetrics
@@ -32,53 +33,52 @@ def execute_ior_per_rank(self, rank):
         self.log.info("Running Test on rank: %s", rank)
         # create the pool on specified rank.
         self.add_pool(connect=False, target_list=[rank])
+        self.container = self.get_container(self.pool)
+
+        host = self.server_managers[0].get_host(rank)
+
+        # execute ior on given rank and collect the results
+        try:
+            self.ior_cmd.flags.update(self.write_flags)
+            dfs_out = self.run_ior_with_pool(create_cont=False, fail_on_warning=self.log.info)
+            dfs_perf_write = IorCommand.get_ior_metrics(dfs_out)
+            self.ior_cmd.flags.update(self.read_flags)
+            dfs_out = self.run_ior_with_pool(create_cont=False, fail_on_warning=self.log.info)
+            dfs_perf_read = IorCommand.get_ior_metrics(dfs_out)
+
+            # Destroy container, to be sure we use newly created container in next iteration
+            self.container.destroy()
+            self.container = None
+
+            # gather actual and expected perf data to be compared
+            dfs_max_write = float(dfs_perf_write[0][IorMetrics.MAX_MIB])
+            dfs_max_read = float(dfs_perf_read[0][IorMetrics.MAX_MIB])
+            actual_write_x = abs(percent_change(self.expected_bw, dfs_max_write))
+            actual_read_x = abs(percent_change(self.expected_bw, dfs_max_read))
+
+            # verify write performance
+            if actual_write_x > self.write_x:
+                if host not in self.failed_nodes:
+                    self.failed_nodes[host] = []
+                self.failed_nodes[host].append(
+                    f"rank {rank} low write perf. "
+                    f"BW: {dfs_max_write:.2f}/{self.expected_bw:.2f}; "
+                    f"percent diff: {actual_write_x:.2f}/{self.write_x:.2f}")
+
+            # verify read performance
+            if actual_read_x > self.read_x:
+                if host not in self.failed_nodes:
+                    self.failed_nodes[host] = []
+                self.failed_nodes[host].append(
+                    f"rank {rank} low read perf. "
+                    f"BW: {dfs_max_read:.2f}/{self.expected_bw:.2f}; "
+                    f"percent diff: {actual_read_x:.2f}/{self.read_x:.2f}")
+
+        except (TestFail, DaosTestError) as error:
+            if host not in self.failed_nodes:
+                self.failed_nodes[host] = []
+            self.failed_nodes[host].append(str(error))
 
-        # execute ior on given rank for different transfer sizes and collect the results
-        for idx, transfer_size in enumerate(self.transfer_sizes):
-            try:
-                self.ior_cmd.transfer_size.update(transfer_size)
-                self.ior_cmd.flags.update(self.write_flags)
-                dfs_out = self.run_ior_with_pool(fail_on_warning=self.log.info)
-                dfs_perf_write = IorCommand.get_ior_metrics(dfs_out)
-                self.ior_cmd.flags.update(self.read_flags)
-                dfs_out = self.run_ior_with_pool(create_cont=False, fail_on_warning=self.log.info)
-                dfs_perf_read = IorCommand.get_ior_metrics(dfs_out)
-
-                # Destroy container, to be sure we use newly created container in next iteration
-                self.container.destroy()
-                self.container = None
-
-                # gather actual and expected perf data to be compared
-                if idx == 0:
-                    dfs_max_write = float(dfs_perf_write[0][IorMetrics.MAX_MIB])
-                    dfs_max_read = float(dfs_perf_read[0][IorMetrics.MAX_MIB])
-                    actual_write_x = percent_change(dfs_max_write, self.expected_bw)
-                    actual_read_x = percent_change(dfs_max_read, self.expected_bw)
-                else:
-                    dfs_max_write = float(dfs_perf_write[0][IorMetrics.MAX_OPS])
-                    dfs_max_read = float(dfs_perf_read[0][IorMetrics.MAX_OPS])
-                    actual_write_x = percent_change(dfs_max_write, self.expected_iops)
-                    actual_read_x = percent_change(dfs_max_read, self.expected_iops)
-
-                # compare actual and expected perf data
-                self.assertLessEqual(abs(actual_write_x), self.write_x,
-                                     "Max Write Diff too large for rank: {}".format(rank))
-                self.assertLessEqual(abs(actual_read_x), self.read_x,
-                                     "Max Read Diff too large for rank: {}".format(rank))
-                # collect list of good nodes
-                good_node = self.server_managers[0].get_host(rank)
-                if ((good_node not in self.good_nodes)
-                        and (good_node not in self.failed_nodes)):
-                    self.good_nodes.append(good_node)
-            except (TestFail, DaosTestError):
-                # collect bad nodes
-                failed_node = self.server_managers[0].get_host(rank)
-                if failed_node not in self.failed_nodes:
-                    self.failed_nodes[failed_node] = [rank]
-                else:
-                    self.failed_nodes[failed_node].append(rank)
-                if failed_node in self.good_nodes:
-                    self.good_nodes.remove(failed_node)
         # Destroy pool, to be sure we use newly created pool in next iteration
         self.pool.destroy()
         self.pool = None
@@ -100,8 +100,6 @@ def test_ior_per_rank(self):
 
         # test params
         self.failed_nodes = {}
-        self.good_nodes = []
-        self.transfer_sizes = self.params.get("transfer_sizes", self.ior_cmd.namespace)
         self.write_flags = self.params.get("write_flags", self.ior_cmd.namespace)
         self.read_flags = self.params.get("read_flags", self.ior_cmd.namespace)
 
@@ -122,13 +120,15 @@ def test_ior_per_rank(self):
         for rank in rank_list:
             self.execute_ior_per_rank(rank)
 
-        # list of good nodes
-        if self.good_nodes:
-            self.log.info("List of good nodes: %s", self.good_nodes)
+        # the good nodes are any that did not fail
+        good_nodes = self.hostlist_servers - NodeSet.fromlist(self.failed_nodes.keys())
+        if good_nodes:
+            self.log.info("Good nodes: %s", good_nodes)
 
         # list the failed node and the rank number associated with that node
         if self.failed_nodes:
-            self.log.info("List of failed ranks with corresponding nodes")
-            for node, rank in self.failed_nodes.items():
-                self.log.info("Node: %s, Rank: %s", node, rank)
+            self.log.info("List of failed nodes with corresponding ranks")
+            for node, reason_list in self.failed_nodes.items():
+                for reason in reason_list:
+                    self.log.info("%s: %s", node, reason)
             self.fail("Performance check failed for one or more nodes")

src/tests/ftest/deployment/ior_per_rank.yaml

@@ -23,26 +23,24 @@ server_config:
 pool:
   mode: 146
   size: 750G # Cannot use percentage, as it does not work when using pool create for per rank.
-  control_method: dmg
   properties: ec_cell_sz:128KiB
 container:
   type: POSIX
   properties: cksum:crc16,cksum_size:16384,srv_cksum:on
   control_method: daos
+  oclass: SX
 ior:
   client_processes:
     ppn: 32
   api: DFS
   test_file: testFile
   write_flags: "-w -C -e -g -G 27 -k -Q 1"
   read_flags: "-r -R -C -e -g -G 27 -k -Q 1"
-  sw_deadline: 30
+  sw_deadline: 15
   sw_wearout: 1
   sw_status_file: "/var/tmp/daos_testing/stoneWallingStatusFile"
-  dfs_oclass: 'SX'
-  transfer_sizes:
-    - 1M
-    - 256B
+  dfs_oclass: SX
+  transfer_size: 1M
   block_size: 150G
   # 0.5 only for CI, due to the varying nature of different clusters in CI.
   # Change it to 15% (0.15) for Aurora.