X-Git-Url: https://git.arvados.org/arvados.git/blobdiff_plain/6d788bb6ab248cf5aef02a049912fbcd42da514f..55aafbb07904ca24390dd47ea960eae7cb2b909a:/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 5caf83472f..9b8410e9aa 100644 --- a/tools/crunchstat-summary/crunchstat_summary/summarizer.py +++ b/tools/crunchstat-summary/crunchstat_summary/summarizer.py @@ -1,3 +1,7 @@ +# Copyright (C) The Arvados Authors. All rights reserved. +# +# SPDX-License-Identifier: AGPL-3.0 + from __future__ import print_function import arvados @@ -11,6 +15,7 @@ import math import re import sys import threading +import _strptime from arvados.api import OrderedJsonModel from crunchstat_summary import logger @@ -21,6 +26,11 @@ 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 @@ -65,7 +75,7 @@ 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')) @@ -96,79 +106,82 @@ class Summarizer(object): 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 - self.starttime, val)) - else: - if stat in ['rss']: - task.series[category, stat].append( - (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 + 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( @@ -197,6 +210,8 @@ 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" @@ -226,7 +241,8 @@ 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'], @@ -238,7 +254,17 @@ class Summarizer(object): ('Max network speed in a single interval: {}MB/s', self.stats_max['net:eth0']['tx+rx__rate'] + self.stats_max['net:keep0']['tx+rx__rate'], - lambda x: x / 1e6)): + 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 @@ -249,7 +275,8 @@ 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%""" @@ -321,6 +348,24 @@ 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. @@ -345,10 +390,16 @@ class JobSummarizer(Summarizer): 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 = 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) @@ -366,15 +417,12 @@ class PipelineSummarizer(object): 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