+# 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.dygraphs
import crunchstat_summary.reader
import datetime
import functools
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')
+
+
+WEBCHART_CLASS = crunchstat_summary.dygraphs.DygraphsChart
+
+
class Task(object):
def __init__(self):
self.starttime = None
class Summarizer(object):
- def __init__(self, logdata, label=None, skip_child_jobs=False):
+ def __init__(self, logdata, label=None, skip_child_jobs=False, uuid=None, **kwargs):
self._logdata = logdata
+ self.uuid = uuid
self.label = label
self.starttime = None
self.finishtime = None
# 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))
# constructor will overwrite this with something useful.
self.existing_constraints = {}
- logger.debug("%s: logdata %s", self.label, repr(logdata))
+ 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)
+ detected_crunch1 = False
for line in self._logdata:
- m = re.search(r'^\S+ \S+ \d+ (?P<seq>\d+) job_task (?P<task_uuid>\S+)$', line)
- if m:
- seq = int(m.group('seq'))
- uuid = m.group('task_uuid')
- self.seq_to_uuid[seq] = uuid
- logger.debug('%s: seq %d is task %s', self.label, seq, uuid)
- continue
+ if not detected_crunch1 and '-8i9sb-' in line:
+ detected_crunch1 = True
+
+ if detected_crunch1:
+ m = re.search(r'^\S+ \S+ \d+ (?P<seq>\d+) job_task (?P<task_uuid>\S+)$', line)
+ if m:
+ seq = int(m.group('seq'))
+ uuid = m.group('task_uuid')
+ self.seq_to_uuid[seq] = uuid
+ 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)
- if m:
- task_id = self.seq_to_uuid[int(m.group('seq'))]
- elapsed = int(m.group('elapsed'))
- self.task_stats[task_id]['time'] = {'elapsed': elapsed}
- if elapsed > self.stats_max['time']['elapsed']:
- self.stats_max['time']['elapsed'] = elapsed
- continue
+ 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'))
+ self.task_stats[task_id]['time'] = {'elapsed': elapsed}
+ if elapsed > self.stats_max['time']['elapsed']:
+ self.stats_max['time']['elapsed'] = elapsed
+ continue
- m = re.search(r'^\S+ \S+ \d+ (?P<seq>\d+) stderr Queued job (?P<uuid>\S+)$', line)
- if m:
- uuid = m.group('uuid')
- if self._skip_child_jobs:
- logger.warning('%s: omitting stats from child job %s'
- ' because --skip-child-jobs flag is on',
- self.label, uuid)
+ m = re.search(r'^\S+ \S+ \d+ (?P<seq>\d+) stderr Queued job (?P<uuid>\S+)$', line)
+ if m:
+ uuid = m.group('uuid')
+ if self._skip_child_jobs:
+ logger.warning('%s: omitting stats from child job %s'
+ ' because --skip-child-jobs flag is on',
+ self.label, uuid)
+ continue
+ logger.debug('%s: follow %s', self.label, uuid)
+ child_summarizer = ProcessSummarizer(uuid)
+ 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
- logger.debug('%s: follow %s', self.label, uuid)
- child_summarizer = JobSummarizer(uuid)
- child_summarizer.stats_max = self.stats_max
- child_summarizer.task_stats = self.task_stats
- child_summarizer.tasks = self.tasks
- child_summarizer.run()
- logger.debug('%s: done %s', self.label, uuid)
- continue
- 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
+ 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$', line)
+ if not m:
+ continue
+ else:
+ # crunch2
+ m = re.search(r'^(?P<timestamp>\S+) (?P<category>\S+) (?P<current>.*?)( -- interval (?P<interval>.*))?\n$', 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)
+ try:
+ self.label = m.group('job_uuid')
+ except IndexError:
+ self.label = 'container'
if m.group('category').endswith(':'):
# "stderr crunchstat: notice: ..."
continue
elif m.group('category') in ('error', 'caught'):
continue
- elif m.group('category') == 'read':
+ elif m.group('category') in ['read', 'open', 'cgroup', 'CID']:
# "stderr crunchstat: read /proc/1234/net/dev: ..."
- # (crunchstat formatting fixed, but old logs still say this)
+ # (old logs are less careful with unprefixed error messages)
continue
- task_id = self.seq_to_uuid[int(m.group('seq'))]
+
+ if detected_crunch1:
+ task_id = self.seq_to_uuid[int(m.group('seq'))]
+ else:
+ task_id = 'container'
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')
+ timestamp = m.group('timestamp')
+ if timestamp[10:11] == '_':
+ timestamp = datetime.datetime.strptime(
+ timestamp, '%Y-%m-%d_%H:%M:%S')
+ elif timestamp[10:11] == 'T':
+ timestamp = datetime.datetime.strptime(
+ timestamp[:19], '%Y-%m-%dT%H:%M:%S')
+ else:
+ raise ValueError("Cannot parse timestamp {!r}".format(
+ timestamp))
+
if not task.starttime:
task.starttime = timestamp
logger.debug('%s: task %s starttime %s',
category = m.group('category')
words = m.group(group).split(' ')
stats = {}
- for val, stat in zip(words[::2], words[1::2]):
- try:
+ try:
+ for val, stat in zip(words[::2], words[1::2]):
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))
+ except ValueError as e:
+ logger.warning('Error parsing {} stat: {!r}'.format(
+ stat, e))
+ continue
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:
val = val / this_interval_s
if stat in ['user+sys__rate', 'tx+rx__rate']:
task.series[category, stat].append(
- (timestamp - task.starttime, val))
+ (timestamp - self.starttime, val))
else:
if stat in ['rss']:
task.series[category, stat].append(
- (timestamp - task.starttime, 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
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"
def html_report(self):
- return crunchstat_summary.chartjs.ChartJS(self.label, [self]).html()
+ return WEBCHART_CLASS(self.label, [self]).html()
def _text_report_gen(self):
yield "\t".join(['category', 'metric', 'task_max', 'task_max_rate', 'job_total'])
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))
+ 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 (
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.
self.label = collection_id
-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
- if self.job['log']:
- rdr = crunchstat_summary.reader.CollectionReader(self.job['log'])
- label = self.job['uuid']
- else:
- rdr = crunchstat_summary.reader.LiveLogReader(self.job['uuid'])
- label = self.job['uuid'] + ' (partial)'
- super(JobSummarizer, self).__init__(rdr, **kwargs)
+def NewSummarizer(process, **kwargs):
+ """Construct with the appropriate subclass for this uuid/object."""
+
+ if not isinstance(process, dict):
+ uuid = process
+ process = None
+ arv = arvados.api('v1', model=OrderedJsonModel())
+
+ if '-dz642-' in uuid:
+ if process is None:
+ process = arv.containers().get(uuid=uuid).execute()
+ klass = ContainerTreeSummarizer
+ elif '-xvhdp-' in uuid:
+ if process is None:
+ process = arv.container_requests().get(uuid=uuid).execute()
+ klass = ContainerTreeSummarizer
+ elif '-8i9sb-' in uuid:
+ if process is None:
+ process = arv.jobs().get(uuid=uuid).execute()
+ klass = JobSummarizer
+ elif '-d1hrv-' in uuid:
+ if process is None:
+ process = arv.pipeline_instances().get(uuid=uuid).execute()
+ klass = PipelineSummarizer
+ elif '-4zz18-' in uuid:
+ return CollectionSummarizer(collection_id=uuid)
+ else:
+ raise ArgumentError("Unrecognized uuid %s", uuid)
+ return klass(process, uuid=uuid, **kwargs)
+
+
+class ProcessSummarizer(Summarizer):
+ """Process is a job, pipeline, container, or container request."""
+
+ def __init__(self, process, label=None, **kwargs):
+ rdr = None
+ self.process = process
+ if label is None:
+ label = self.process.get('name', self.process['uuid'])
+ if self.process.get('log'):
+ try:
+ rdr = crunchstat_summary.reader.CollectionReader(self.process['log'])
+ except arvados.errors.NotFoundError as e:
+ logger.warning("Trying event logs after failing to read "
+ "log collection %s: %s", self.process['log'], e)
+ if rdr is None:
+ rdr = crunchstat_summary.reader.LiveLogReader(self.process['uuid'])
+ label = label + ' (partial)'
+ super(ProcessSummarizer, self).__init__(rdr, label=label, **kwargs)
+ self.existing_constraints = self.process.get('runtime_constraints', {})
+
+
+class JobSummarizer(ProcessSummarizer):
+ pass
+
+
+class ContainerSummarizer(ProcessSummarizer):
+ pass
+
+
+class MultiSummarizer(object):
+ def __init__(self, children={}, label=None, threads=1, **kwargs):
+ self.throttle = threading.Semaphore(threads)
+ self.children = children
self.label = label
- self.existing_constraints = self.job.get('runtime_constraints', {})
+ def run_and_release(self, target, *args, **kwargs):
+ try:
+ return target(*args, **kwargs)
+ finally:
+ self.throttle.release()
-class PipelineSummarizer(object):
- def __init__(self, pipeline_instance_uuid, **kwargs):
- arv = arvados.api('v1', model=OrderedJsonModel())
- instance = arv.pipeline_instances().get(
- uuid=pipeline_instance_uuid).execute()
- self.summarizers = collections.OrderedDict()
+ def run(self):
+ threads = []
+ for child in self.children.itervalues():
+ self.throttle.acquire()
+ t = threading.Thread(target=self.run_and_release, args=(child.run, ))
+ t.daemon = True
+ t.start()
+ threads.append(t)
+ for t in threads:
+ t.join()
+
+ def text_report(self):
+ txt = ''
+ for cname, child in self.children.iteritems():
+ if len(self.children) > 1:
+ txt += '### Summary for {} ({})\n'.format(
+ cname, child.process['uuid'])
+ txt += child.text_report()
+ txt += '\n'
+ return txt
+
+ def html_report(self):
+ return WEBCHART_CLASS(self.label, self.children.itervalues()).html()
+
+
+class PipelineSummarizer(MultiSummarizer):
+ def __init__(self, instance, **kwargs):
+ children = collections.OrderedDict()
for cname, component in instance['components'].iteritems():
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
- self.summarizers[cname] = summarizer
- self.label = pipeline_instance_uuid
+ summarizer.label = '{} {}'.format(
+ cname, component['job']['uuid'])
+ children[cname] = summarizer
+ super(PipelineSummarizer, self).__init__(
+ children=children,
+ label=instance['uuid'],
+ **kwargs)
- def run(self):
- for summarizer in self.summarizers.itervalues():
- summarizer.run()
- def text_report(self):
- txt = ''
- for cname, summarizer in self.summarizers.iteritems():
- txt += '### Summary for {} ({})\n'.format(
- cname, summarizer.job['uuid'])
- txt += summarizer.text_report()
- txt += '\n'
- return txt
+class ContainerTreeSummarizer(MultiSummarizer):
+ def __init__(self, root, **kwargs):
+ arv = arvados.api('v1', model=OrderedJsonModel())
- def html_report(self):
- return crunchstat_summary.chartjs.ChartJS(
- self.label, self.summarizers.itervalues()).html()
+ label = kwargs.pop('label', None) or root.get('name') or root['uuid']
+ root['name'] = label
+
+ children = collections.OrderedDict()
+ todo = collections.deque((root, ))
+ while len(todo) > 0:
+ current = todo.popleft()
+ label = current['name']
+ if current['uuid'].find('-xvhdp-') > 0:
+ current = arv.containers().get(uuid=current['container_uuid']).execute()
+ children[current['uuid']] = ContainerSummarizer(
+ current, label=label, **kwargs)
+ page_filters = []
+ while True:
+ items = arv.container_requests().index(
+ order=['uuid asc'],
+ filters=page_filters+[
+ ['requesting_container_uuid', '=', current['uuid']]],
+ ).execute()['items']
+ if not items:
+ break
+ page_filters = [['uuid', '>', items[-1]['uuid']]]
+ for cr in items:
+ if cr['container_uuid']:
+ logger.debug('%s: container req %s', current['uuid'], cr['uuid'])
+ cr['name'] = cr.get('name') or cr['uuid']
+ todo.append(cr)
+ super(ContainerTreeSummarizer, self).__init__(
+ children=children,
+ label=root['name'],
+ **kwargs)
projects / arvados.git / blobdiff