使用 lttng trace rgw 性能

本节介绍如何使用 rgw tracepoints 跟踪 rgw 数据传送、bi更新和头对象写的时间开销。

1. 打过patch后，获取 trace数据

操作和 [1] 中过程一样。不一样的处理如下：

- apply patch [2]
- 编译打开 lttng 支持（同[1]）
- ceph.conf 中配置项打开

  rgw rados tracing = true
  rgw op tracing = true

2. 处理获取到的数据

上脚本

#!/usr/bin/env python
# -*- coding: utf-8 -*
# vi:set tw=0 ts=4 sw=4 nowrap fdm=indent
# Yang Honggang <rtlinux@163.com>

import pdb
import json
import sys
from babeltrace import *

# all events
g_events = set()
g_summary = {}

def update_summary(evt_name, cost):
  g_events.add(evt_name)

  if evt_name not in g_summary:
    g_summary[evt_name] = {
          'max': cost,
          'min': cost,
          'sum': cost,
          'count': 1
        }
  else:
    e = g_summary[evt_name]
    if e['max'] < cost:
      e['max'] = cost
    if cost < e['min']:
      e['min'] = cost

    e['sum'] += cost
    e['count'] += 1

def update():
  for k, v in g_summary.items():
    v['avg'] = float(v['sum']) // v['count']

    v['avg'] /= float(1000000000)
    v['min'] /= float(1000000000)
    v['max'] /= float(1000000000)
    v['sum'] /= float(1000000000)

if __name__ == '__main__':

  traces = TraceCollection()
  ret = traces.add_trace(sys.argv[1], "ctf")

  event_type = 'all'
  if len(sys.argv) == 3:
    event_type = str(sys.argv[2])

  events_enter = {}
  events_enter_set = set()
  for event in traces.events:
    # filter
    if event_type != 'all' and event_type not in event.name:
      continue

    if event.name == 'rgw_op:before_data_receive':
      op_tag = event.name + str(event['req_id'])
      # print ('enqueue %s' % op_tag)
      if op_tag in events_enter_set:
        print ('reset %s' % op_tag)

      events_enter_set.add(op_tag)
      events_enter[op_tag] = {
            'ts': event.timestamp,
          }
      continue

    if event.name == 'rgw_op:after_data_receive':
      op_tag = 'rgw_op:before_data_receive' + str(event['req_id'])
      # print ('dequeue %s' % op_tag)
      try:
        e = events_enter[op_tag]
        cost = event.timestamp - e['ts']

        update_summary('rgw_op:data_receive', cost)
        del events_enter[op_tag]
        events_enter_set.remove(op_tag)
        print ('%s %d' % ('rgw_op:data_receive', cost))
      except KeyError:
        print ('drop %s' % str(event))

      continue

    if event.name.startswith('rgw_') and event.name.endswith('_enter'):
      op_tag = event.name + str(event['req_id'])
      # print ('enqueue %s' % op_tag)
      assert(op_tag not in events_enter_set), ("dump events_enter_set: %s" % str(pdb.set_trace()))
      if op_tag in events_enter_set:
          print ('reset %s' % op_tag)
      events_enter_set.add(op_tag)
      events_enter[op_tag] = {
              'ts': event.timestamp,
            }
      continue

    if event.name.startswith('rgw_') and event.name.endswith('_exit'):
      op_tag = event.name[:-4] # get rid of tail 'exit'
      op_tag += 'enter' + str(event['req_id'])
      # print ('dequeue %s' % op_tag)
      try:
        e = events_enter[op_tag]
        cost = event.timestamp - e['ts']

        update_summary(event.name[:-5], cost)
        del events_enter[op_tag]
        events_enter_set.remove(op_tag)

        print ('%s %d' % (event.name[:-5], cost))
      except KeyError:
        print ('drop %s' % str(event))

      continue

  # update avg
  update()
  print (json.dumps(g_summary, indent=4))

[1] filestore lttng trace

[2] rgw lttng tracepoints