X-Git-Url: https://git.arvados.org/arvados.git/blobdiff_plain/f6bb371a170d0a74db6abf9df0f65aabe08d7cf9..refs/heads/10472-crunchstat-summary-job-components:/tools/crunchstat-summary/crunchstat_summary/summarizer.py diff --git a/tools/crunchstat-summary/crunchstat_summary/summarizer.py b/tools/crunchstat-summary/crunchstat_summary/summarizer.py index 8b9bfa32b0..f1c97c282c 100644 --- a/tools/crunchstat-summary/crunchstat_summary/summarizer.py +++ b/tools/crunchstat-summary/crunchstat_summary/summarizer.py @@ -3,12 +3,15 @@ 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 @@ -19,6 +22,11 @@ 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 @@ -36,8 +44,7 @@ class Summarizer(object): # 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)) @@ -51,10 +58,25 @@ class Summarizer(object): # 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_child(self, 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) + return + 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.starttime = self.starttime + child_summarizer.run() + logger.debug('%s: done %s', self.label, uuid) 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\d+) job_task (?P\S+)$', line) if m: @@ -64,7 +86,7 @@ class Summarizer(object): logger.debug('%s: seq %d is task %s', self.label, seq, uuid) continue - m = re.search(r'^\S+ \S+ \d+ (?P\d+) success in (?P\d+) seconds', line) + m = re.search(r'^\S+ \S+ \d+ (?P\d+) (success in|failure \(#., permanent\) after) (?P\d+) seconds', line) if m: task_id = self.seq_to_uuid[int(m.group('seq'))] elapsed = int(m.group('elapsed')) @@ -73,101 +95,99 @@ class Summarizer(object): self.stats_max['time']['elapsed'] = elapsed continue + # Old style job logs only - newer style uses job['components'] + uuid = None m = re.search(r'^\S+ \S+ \d+ (?P\d+) stderr Queued job (?P\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 = 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) + self.run_child(m.group('uuid')) continue m = re.search(r'^(?P[^\s.]+)(\.\d+)? (?P\S+) \d+ (?P\d+) stderr crunchstat: (?P\S+) (?P.*?)( -- interval (?P.*))?\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 - logger.debug('%s: done parsing', self.label) + (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 log', self.label) + + # Enabling this will roll up stats for all subjobs into the parent job + if False and 'components' in self.job: + for cname, component in self.job['components'].iteritems(): + self.run_child(component) self.job_tot = collections.defaultdict( functools.partial(collections.defaultdict, int)) @@ -195,6 +215,8 @@ class Summarizer(object): 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" @@ -224,17 +246,30 @@ class Summarizer(object): 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 @@ -245,7 +280,8 @@ class Summarizer(object): 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%""" @@ -254,7 +290,7 @@ class Summarizer(object): 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 ( @@ -317,6 +353,24 @@ class Summarizer(object): 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. @@ -327,67 +381,10 @@ class Summarizer(object): return '{}'.format(val) -class CollectionReader(object): - def __init__(self, collection_id): - 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))) - self._reader = collection.open(filenames[0]) - - def __iter__(self): - return iter(self._reader) - - -class LiveLogReader(object): - def __init__(self, job_uuid): - logger.debug('load stderr events for job %s', job_uuid) - self._filters = [ - ['object_uuid', '=', job_uuid], - ['event_type', '=', 'stderr']] - self._buffer = collections.deque() - self._got = 0 - self._got_all = False - self._label = job_uuid - self._last_id = 0 - - def __iter__(self): - return self - - def next(self): - if self._buffer is None: - raise StopIteration - if len(self._buffer) == 0: - if self._got_all: - raise StopIteration - page = arvados.api().logs().index( - limit=1000, - order=['id asc'], - filters=self._filters + [['id','>',str(self._last_id)]], - ).execute() - self._got += len(page['items']) - logger.debug('%s: received %d of %d log events', self._label, self._got, self._got + page['items_available'] - len(page['items'])) - if len(page['items']) == 0: - self._got_all = True - self._buffer = None - raise StopIteration - elif len(page['items']) == page['items_available']: - # Don't try to fetch any more after this page - self._got_all = True - for i in page['items']: - for line in i['properties']['text'].split('\n'): - self._buffer.append(line) - self._last_id = i['id'] - return self._buffer.popleft() + '\n' - - class CollectionSummarizer(Summarizer): def __init__(self, collection_id, **kwargs): super(CollectionSummarizer, self).__init__( - CollectionReader(collection_id), **kwargs) + crunchstat_summary.reader.CollectionReader(collection_id), **kwargs) self.label = collection_id @@ -398,11 +395,17 @@ class JobSummarizer(Summarizer): self.job = arv.jobs().get(uuid=job).execute() else: self.job = job - if self.job['log']: - rdr = CollectionReader(self.job['log']) - label = self.job['uuid'] - else: - rdr = LiveLogReader(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 @@ -411,29 +414,38 @@ class JobSummarizer(Summarizer): class PipelineSummarizer(object): def __init__(self, pipeline_instance_uuid, **kwargs): - arv = arvados.api('v1', model=OrderedJsonModel()) - instance = arv.pipeline_instances().get( + self.arv = arvados.api('v1', model=OrderedJsonModel()) + instance = self.arv.pipeline_instances().get( uuid=pipeline_instance_uuid).execute() self.summarizers = 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']) - summarizer = JobSummarizer(component['job'], **kwargs) - summarizer.label = cname - self.summarizers[cname] = summarizer + self.summarize_job(cname, component['job'], **kwargs) self.label = pipeline_instance_uuid + def summarize_job(self, cname, job, **kwargs): + uuid = job['uuid'] + logger.info("%s: job %s", cname, uuid) + summarizer = JobSummarizer(job, **kwargs) + summarizer.label = '{} {}'.format(cname, uuid) + self.summarizers[cname] = summarizer + if 'components' in job: + for cname, uuid in job['components'].iteritems(): + subjob = self.arv.jobs().get(uuid=uuid).execute() + self.summarize_job(cname, subjob, **kwargs) + 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 = ''