+# Copyright (C) The Arvados Authors. All rights reserved.
+#
+# SPDX-License-Identifier: AGPL-3.0
+
from __future__ import print_function
import arvados
import collections
import crunchstat_summary.chartjs
+import crunchstat_summary.reader
import datetime
import functools
import itertools
import math
import re
import sys
+import threading
+import _strptime
from arvados.api import OrderedJsonModel
from crunchstat_summary import logger
AVAILABLE_RAM_RATIO = 0.95
+# Workaround datetime.datetime.strptime() thread-safety bug by calling
+# it once before starting threads. https://bugs.python.org/issue7980
+datetime.datetime.strptime('1999-12-31_23:59:59', '%Y-%m-%d_%H:%M:%S')
+
+
class Task(object):
def __init__(self):
self.starttime = None
class Summarizer(object):
- existing_constraints = {}
-
def __init__(self, logdata, label=None, skip_child_jobs=False):
self._logdata = logdata
# stats_max: {category: {stat: val}}
self.stats_max = collections.defaultdict(
- functools.partial(collections.defaultdict,
- lambda: float('-Inf')))
+ functools.partial(collections.defaultdict, lambda: 0))
# task_stats: {task_id: {category: {stat: val}}}
self.task_stats = collections.defaultdict(
functools.partial(collections.defaultdict, dict))
self.seq_to_uuid = {}
self.tasks = collections.defaultdict(Task)
- logger.debug("%s: logdata %s", self.label, repr(logdata))
+ # We won't bother recommending new runtime constraints if the
+ # constraints given when running the job are known to us and
+ # are already suitable. If applicable, the subclass
+ # constructor will overwrite this with something useful.
+ self.existing_constraints = {}
+
+ logger.debug("%s: logdata %s", self.label, logdata)
def run(self):
- logger.debug("%s: parsing log data", self.label)
+ logger.debug("%s: parsing logdata %s", self.label, self._logdata)
for line in self._logdata:
m = re.search(r'^\S+ \S+ \d+ (?P<seq>\d+) job_task (?P<task_uuid>\S+)$', line)
if m:
logger.debug('%s: seq %d is task %s', self.label, seq, uuid)
continue
- m = re.search(r'^\S+ \S+ \d+ (?P<seq>\d+) success in (?P<elapsed>\d+) seconds', line)
+ m = re.search(r'^\S+ \S+ \d+ (?P<seq>\d+) (success in|failure \(#., permanent\) after) (?P<elapsed>\d+) seconds', line)
if m:
task_id = self.seq_to_uuid[int(m.group('seq'))]
elapsed = int(m.group('elapsed'))
child_summarizer.stats_max = self.stats_max
child_summarizer.task_stats = self.task_stats
child_summarizer.tasks = self.tasks
+ child_summarizer.starttime = self.starttime
child_summarizer.run()
logger.debug('%s: done %s', self.label, uuid)
continue
- m = re.search(r'^(?P<timestamp>\S+) (?P<job_uuid>\S+) \d+ (?P<seq>\d+) stderr crunchstat: (?P<category>\S+) (?P<current>.*?)( -- interval (?P<interval>.*))?\n', line)
+ m = re.search(r'^(?P<timestamp>[^\s.]+)(\.\d+)? (?P<job_uuid>\S+) \d+ (?P<seq>\d+) stderr crunchstat: (?P<category>\S+) (?P<current>.*?)( -- interval (?P<interval>.*))?\n', line)
if not m:
continue
- if self.label is None:
- self.label = m.group('job_uuid')
- logger.debug('%s: using job uuid as label', self.label)
- if m.group('category').endswith(':'):
- # "stderr crunchstat: notice: ..."
- continue
- elif m.group('category') in ('error', 'caught'):
- continue
- elif m.group('category') == 'read':
- # "stderr crunchstat: read /proc/1234/net/dev: ..."
- # (crunchstat formatting fixed, but old logs still say this)
- continue
- task_id = self.seq_to_uuid[int(m.group('seq'))]
- task = self.tasks[task_id]
-
- # Use the first and last crunchstat timestamps as
- # approximations of starttime and finishtime.
- timestamp = datetime.datetime.strptime(
- m.group('timestamp'), '%Y-%m-%d_%H:%M:%S')
- if not task.starttime:
- task.starttime = timestamp
- logger.debug('%s: task %s starttime %s',
- self.label, task_id, timestamp)
- task.finishtime = timestamp
-
- if not self.starttime:
- self.starttime = timestamp
- self.finishtime = timestamp
-
- this_interval_s = None
- for group in ['current', 'interval']:
- if not m.group(group):
+ try:
+ if self.label is None:
+ self.label = m.group('job_uuid')
+ logger.debug('%s: using job uuid as label', self.label)
+ if m.group('category').endswith(':'):
+ # "stderr crunchstat: notice: ..."
continue
- category = m.group('category')
- words = m.group(group).split(' ')
- stats = {}
- for val, stat in zip(words[::2], words[1::2]):
- try:
- if '.' in val:
- stats[stat] = float(val)
- else:
- stats[stat] = int(val)
- except ValueError as e:
- raise ValueError(
- 'Error parsing {} stat in "{}": {!r}'.format(
- stat, line, e))
- if 'user' in stats or 'sys' in stats:
- stats['user+sys'] = stats.get('user', 0) + stats.get('sys', 0)
- if 'tx' in stats or 'rx' in stats:
- stats['tx+rx'] = stats.get('tx', 0) + stats.get('rx', 0)
- for stat, val in stats.iteritems():
- if group == 'interval':
- if stat == 'seconds':
- this_interval_s = val
- continue
- elif not (this_interval_s > 0):
- logger.error(
- "BUG? interval stat given with duration {!r}".
- format(this_interval_s))
- continue
+ elif m.group('category') in ('error', 'caught'):
+ continue
+ elif m.group('category') == 'read':
+ # "stderr crunchstat: read /proc/1234/net/dev: ..."
+ # (crunchstat formatting fixed, but old logs still say this)
+ continue
+ task_id = self.seq_to_uuid[int(m.group('seq'))]
+ task = self.tasks[task_id]
+
+ # Use the first and last crunchstat timestamps as
+ # approximations of starttime and finishtime.
+ timestamp = datetime.datetime.strptime(
+ m.group('timestamp'), '%Y-%m-%d_%H:%M:%S')
+ if not task.starttime:
+ task.starttime = timestamp
+ logger.debug('%s: task %s starttime %s',
+ self.label, task_id, timestamp)
+ task.finishtime = timestamp
+
+ if not self.starttime:
+ self.starttime = timestamp
+ self.finishtime = timestamp
+
+ this_interval_s = None
+ for group in ['current', 'interval']:
+ if not m.group(group):
+ continue
+ category = m.group('category')
+ words = m.group(group).split(' ')
+ stats = {}
+ for val, stat in zip(words[::2], words[1::2]):
+ try:
+ if '.' in val:
+ stats[stat] = float(val)
+ else:
+ stats[stat] = int(val)
+ except ValueError as e:
+ raise ValueError(
+ 'Error parsing {} stat: {!r}'.format(
+ stat, e))
+ if 'user' in stats or 'sys' in stats:
+ stats['user+sys'] = stats.get('user', 0) + stats.get('sys', 0)
+ if 'tx' in stats or 'rx' in stats:
+ stats['tx+rx'] = stats.get('tx', 0) + stats.get('rx', 0)
+ for stat, val in stats.iteritems():
+ if group == 'interval':
+ if stat == 'seconds':
+ this_interval_s = val
+ continue
+ elif not (this_interval_s > 0):
+ logger.error(
+ "BUG? interval stat given with duration {!r}".
+ format(this_interval_s))
+ continue
+ else:
+ stat = stat + '__rate'
+ val = val / this_interval_s
+ if stat in ['user+sys__rate', 'tx+rx__rate']:
+ task.series[category, stat].append(
+ (timestamp - self.starttime, val))
else:
- stat = stat + '__rate'
- val = val / this_interval_s
- if stat in ['user+sys__rate', 'tx+rx__rate']:
+ if stat in ['rss']:
task.series[category, stat].append(
- (timestamp - task.starttime, val))
- else:
- if stat in ['rss']:
- task.series[category, stat].append(
- (timestamp - task.starttime, val))
- self.task_stats[task_id][category][stat] = val
- if val > self.stats_max[category][stat]:
- self.stats_max[category][stat] = val
+ (timestamp - self.starttime, val))
+ self.task_stats[task_id][category][stat] = val
+ if val > self.stats_max[category][stat]:
+ self.stats_max[category][stat] = val
+ except Exception as e:
+ logger.info('Skipping malformed line: {}Error was: {}\n'.format(line, e))
logger.debug('%s: done parsing', self.label)
self.job_tot = collections.defaultdict(
return label
def text_report(self):
+ if not self.tasks:
+ return "(no report generated)\n"
return "\n".join(itertools.chain(
self._text_report_gen(),
self._recommend_gen())) + "\n"
lambda x: x * 100),
('Overall CPU usage: {}%',
self.job_tot['cpu']['user+sys'] /
- self.job_tot['time']['elapsed'],
+ self.job_tot['time']['elapsed']
+ if self.job_tot['time']['elapsed'] > 0 else 0,
lambda x: x * 100),
('Max memory used by a single task: {}GB',
self.stats_max['mem']['rss'],
lambda x: x / 1e9),
('Max network traffic in a single task: {}GB',
- self.stats_max['net:eth0']['tx+rx'],
+ self.stats_max['net:eth0']['tx+rx'] +
+ self.stats_max['net:keep0']['tx+rx'],
lambda x: x / 1e9),
('Max network speed in a single interval: {}MB/s',
- self.stats_max['net:eth0']['tx+rx__rate'],
- lambda x: x / 1e6)):
+ self.stats_max['net:eth0']['tx+rx__rate'] +
+ self.stats_max['net:keep0']['tx+rx__rate'],
+ lambda x: x / 1e6),
+ ('Keep cache miss rate {}%',
+ (float(self.job_tot['keepcache']['miss']) /
+ float(self.job_tot['keepcalls']['get']))
+ if self.job_tot['keepcalls']['get'] > 0 else 0,
+ lambda x: x * 100.0),
+ ('Keep cache utilization {}%',
+ (float(self.job_tot['blkio:0:0']['read']) /
+ float(self.job_tot['net:keep0']['rx']))
+ if self.job_tot['net:keep0']['rx'] > 0 else 0,
+ lambda x: x * 100.0)):
format_string, val, transform = args
if val == float('-Inf'):
continue
def _recommend_gen(self):
return itertools.chain(
self._recommend_cpu(),
- self._recommend_ram())
+ self._recommend_ram(),
+ self._recommend_keep_cache())
def _recommend_cpu(self):
"""Recommend asking for 4 cores if max CPU usage was 333%"""
if cpu_max_rate == float('-Inf'):
logger.warning('%s: no CPU usage data', self.label)
return
- used_cores = int(math.ceil(cpu_max_rate))
- asked_cores = self.existing_constraints.get('min_cores_per_node')
+ used_cores = max(1, int(math.ceil(cpu_max_rate)))
+ asked_cores = self.existing_constraints.get('min_cores_per_node')
if asked_cores is None or used_cores < asked_cores:
yield (
'#!! {} max CPU usage was {}% -- '
int(used_mib),
int(math.ceil(nearlygibs(used_mib))*AVAILABLE_RAM_RATIO*1024))
+ def _recommend_keep_cache(self):
+ """Recommend increasing keep cache if utilization < 80%"""
+ if self.job_tot['net:keep0']['rx'] == 0:
+ return
+ utilization = (float(self.job_tot['blkio:0:0']['read']) /
+ float(self.job_tot['net:keep0']['rx']))
+ asked_mib = self.existing_constraints.get('keep_cache_mb_per_task', 256)
+
+ if utilization < 0.8:
+ yield (
+ '#!! {} Keep cache utilization was {:.2f}% -- '
+ 'try runtime_constraints "keep_cache_mb_per_task":{} (or more)'
+ ).format(
+ self.label,
+ utilization * 100.0,
+ asked_mib*2)
+
+
def _format(self, val):
"""Return a string representation of a stat.
class CollectionSummarizer(Summarizer):
def __init__(self, collection_id, **kwargs):
- logger.debug('load collection %s', collection_id)
- collection = arvados.collection.CollectionReader(collection_id)
- filenames = [filename for filename in collection]
- if len(filenames) != 1:
- raise ValueError(
- "collection {} has {} files; need exactly one".format(
- collection_id, len(filenames)))
super(CollectionSummarizer, self).__init__(
- collection.open(filenames[0]), **kwargs)
+ crunchstat_summary.reader.CollectionReader(collection_id), **kwargs)
self.label = collection_id
-class JobSummarizer(CollectionSummarizer):
+class JobSummarizer(Summarizer):
def __init__(self, job, **kwargs):
arv = arvados.api('v1')
if isinstance(job, basestring):
self.job = arv.jobs().get(uuid=job).execute()
else:
self.job = job
- self.label = self.job['uuid']
+ rdr = None
+ if self.job.get('log'):
+ try:
+ rdr = crunchstat_summary.reader.CollectionReader(self.job['log'])
+ except arvados.errors.NotFoundError as e:
+ logger.warning("Trying event logs after failing to read "
+ "log collection %s: %s", self.job['log'], e)
+ else:
+ label = self.job['uuid']
+ if rdr is None:
+ rdr = crunchstat_summary.reader.LiveLogReader(self.job['uuid'])
+ label = self.job['uuid'] + ' (partial)'
+ super(JobSummarizer, self).__init__(rdr, **kwargs)
+ self.label = label
self.existing_constraints = self.job.get('runtime_constraints', {})
- if not self.job['log']:
- raise ValueError(
- "job {} has no log; live summary not implemented".format(
- self.job['uuid']))
- super(JobSummarizer, self).__init__(self.job['log'], **kwargs)
- self.label = self.job['uuid']
-class PipelineSummarizer():
+class PipelineSummarizer(object):
def __init__(self, pipeline_instance_uuid, **kwargs):
arv = arvados.api('v1', model=OrderedJsonModel())
instance = arv.pipeline_instances().get(
if 'job' not in component:
logger.warning(
"%s: skipping component with no job assigned", cname)
- elif component['job'].get('log') is None:
- logger.warning(
- "%s: skipping job %s with no log available",
- cname, component['job'].get('uuid'))
else:
logger.info(
- "%s: logdata %s", cname, component['job']['log'])
+ "%s: job %s", cname, component['job']['uuid'])
summarizer = JobSummarizer(component['job'], **kwargs)
- summarizer.label = cname
+ summarizer.label = '{} {}'.format(
+ cname, component['job']['uuid'])
self.summarizers[cname] = summarizer
self.label = pipeline_instance_uuid
def run(self):
+ threads = []
for summarizer in self.summarizers.itervalues():
- summarizer.run()
+ t = threading.Thread(target=summarizer.run)
+ t.daemon = True
+ t.start()
+ threads.append(t)
+ for t in threads:
+ t.join()
def text_report(self):
txt = ''
projects / arvados.git / blobdiff