From: Laura Flores <lflores@redhat.com>
Date: Fri, 23 Jul 2021 17:40:27 +0000 (+0000)
Subject: mgr, mgr/telemetry: add access to osd commands in mgr and workload metrics to telemetry
X-Git-Tag: v17.1.0~1008^2~12
X-Git-Url: http://git-server-git.apps.pok.os.sepia.ceph.com/?a=commitdiff_plain;h=3b27bac6eec81f33e51b1da3e587c11448d1be22;p=ceph.git

mgr, mgr/telemetry: add access to osd commands in mgr and workload metrics to telemetry

In the manager module (src/pybind/mgr/mgr_module.py), there was a command that allowed access to monitor commands, but there was nothing to access osd commands. I added a new method called osd_command that allows access to the output of commands like "ceph osd.0 perf histogram dump".

In the telemetry module (src/pybind/mgr/telemetry/module.py), I added information that may help to characterize a workload. These changes include:
	- io rate: delta in pg stats since last report, including a pg stats summary and some store stats
	- stats sum per pool: summed up stats from every pg in a pool
	- perf histograms: histogram dump of latency and request size for each osd

Signed-off-by: Laura Flores <lflores@redhat.com>
---

diff --git a/src/pybind/mgr/mgr_module.py b/src/pybind/mgr/mgr_module.py
index f7da34b7713..60a405847d2 100644
--- a/src/pybind/mgr/mgr_module.py
+++ b/src/pybind/mgr/mgr_module.py
@@ -1499,6 +1499,20 @@ class MgrModule(ceph_module.BaseMgrModule, MgrModuleLoggingMixin):
 
         return r
 
+    def osd_command(self, cmd_dict: dict, inbuf: Optional[str] = None) -> Tuple[int, str, str]:
+
+        t1 = time.time()
+        result = CommandResult()
+        self.send_command(result, "osd", cmd_dict['id'], json.dumps(cmd_dict), "", inbuf)
+        r = result.wait()
+        t2 = time.time()
+
+        self.log.debug("osd_command: '{0}' -> {1} in {2:.3f}s".format(
+            cmd_dict['prefix'], r[0], t2 - t1
+        ))
+
+        return r
+
     def send_command(
             self,
             result: CommandResult,
diff --git a/src/pybind/mgr/telemetry/module.py b/src/pybind/mgr/telemetry/module.py
index 3e795bd51b1..e40ee9083a8 100644
--- a/src/pybind/mgr/telemetry/module.py
+++ b/src/pybind/mgr/telemetry/module.py
@@ -302,6 +302,92 @@ class Module(MgrModule):
             'active_changed': sorted(list(active)),
         }
 
+    def get_stat_sum_per_pool(self) -> List[dict]:
+        # Initialize 'result' list
+        result: List[dict] = []
+
+        # Create a list of tuples containing pool ids and their associated names
+        # that will later act as a queue, i.e.:
+        #   pool_queue = [('1', '.mgr'), ('2', 'cephfs.a.meta'), ('3', 'cephfs.a.data')]
+        osd_map = self.get('osd_map')
+        pool_queue: List[tuple] = []
+        for pool in osd_map['pools']:
+            pool_queue.append((str(pool['pool']), pool['pool_name']))
+
+        # Populate 'result', i.e.:
+        #   {
+        #       'pool_id': '1'
+        #       'pool_name': '.mgr'
+        #       'stats_sum': {
+        #           'num_bytes': 36,
+        #           'num_bytes_hit_set_archive': 0,
+        #           ...
+        #           'num_write_kb': 0
+        #           }
+        #       }
+        #   }
+        while pool_queue:
+            # Pop a pool out of pool_queue
+            curr_pool = pool_queue.pop(0)
+
+            # Get the current pool's id and name
+            curr_pool_id, curr_pool_name = curr_pool[0], curr_pool[1]
+
+            # Initialize a dict that will hold aggregated stats for the current pool
+            compiled_stats_dict: Dict[str, dict] = defaultdict(lambda: defaultdict(int))
+
+            # Find out which pgs belong to the current pool and add up
+            # their stats
+            pg_dump = self.get('pg_dump')
+            for pg in pg_dump['pg_stats']:
+                pool_id = pg['pgid'][0:1]
+                if pool_id == curr_pool_id:
+                    for metric in pg['stat_sum']:
+                        compiled_stats_dict['pool_id'] = int(pool_id)
+                        compiled_stats_dict['pool_name'] = curr_pool_name
+                        compiled_stats_dict['stats_sum'][metric] += pg['stat_sum'][metric]
+                else:
+                    continue
+            # 'compiled_stats_dict' now holds all stats pertaining to
+            # the current pool. Adding it to the list of results.
+            result.append(compiled_stats_dict)
+
+        return result
+
+    def get_osd_histograms(self) -> Dict[str, dict]:
+        # Initialize result dict
+        result: Dict[str, dict] = defaultdict()
+
+        # You can get a list of osd ids from the metadata
+        osd_metadata = self.get('osd_metadata')
+        
+        # Grab output from the "osd.x perf histogram dump" command
+        for osd_id in osd_metadata:
+            cmd_dict = {
+                'prefix': 'perf histogram dump',
+                'id': str(osd_id),
+                'format': 'json'
+            }
+            r, outb, outs = self.osd_command(cmd_dict)
+            # Check for invalid calls
+            if r != 0:
+                continue
+            else:
+                try:
+                    # This is where the histograms will land
+                    # if there are any
+                    dump = json.loads(outb)
+                    result['osd.'+osd_id] = dump
+                # Sometimes, json errors occur if
+                # you give it an empty string
+                except json.decoder.JSONDecodeError:
+                    continue
+
+        return result
+
+    def get_io_rate(self) -> dict:
+        return self.get('io_rate')
+
     def gather_crashinfo(self) -> List[Dict[str, str]]:
         crashlist: List[Dict[str, str]] = list()
         errno, crashids, err = self.remote('crash', 'ls')
@@ -731,6 +817,9 @@ class Module(MgrModule):
 
         if 'perf' in channels:
             report['perf_counters'] = self.gather_perf_counters()
+            report['stat_sum_per_pool'] = self.get_stat_sum_per_pool()
+            report['io_rate'] = self.get_io_rate()
+            report['osd_perf_histograms'] = self.get_osd_histograms()
 
         # NOTE: We do not include the 'device' channel in this report; it is
         # sent to a different endpoint.