X-Git-Url: https://git.arvados.org/arvados.git/blobdiff_plain/f6bb371a170d0a74db6abf9df0f65aabe08d7cf9..80d28abfb972c34d7769fbfeac5e3b67049f216b:/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..f422501b10 100644 --- a/tools/crunchstat-summary/crunchstat_summary/summarizer.py +++ b/tools/crunchstat-summary/crunchstat_summary/summarizer.py @@ -3,12 +3,14 @@ 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 from arvados.api import OrderedJsonModel from crunchstat_summary import logger @@ -36,8 +38,7 @@ 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 +52,10 @@ 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(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 +65,7 @@ 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')) @@ -86,6 +87,7 @@ class Summarizer(object): 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 @@ -159,11 +161,11 @@ class Summarizer(object): 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 @@ -195,6 +197,8 @@ 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 +228,30 @@ 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 +262,8 @@ 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 +272,7 @@ 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 +335,24 @@ 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 +363,10 @@ 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 +377,17 @@ 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 @@ -419,21 +404,24 @@ 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 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 = ''