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 # Recommend memory constraints that are this multiple of an integral # number of GiB. (Actual nodes tend to be sold in sizes like 8 GiB # that have amounts like 7.5 GiB according to the kernel.) 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 self.series = collections.defaultdict(list) class Summarizer(object): def __init__(self, logdata, label=None, skip_child_jobs=False): self._logdata = logdata self.label = label self.starttime = None self.finishtime = None self._skip_child_jobs = skip_child_jobs # stats_max: {category: {stat: val}} self.stats_max = collections.defaultdict( 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) # 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 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: 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\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')) 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\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.starttime = self.starttime child_summarizer.run() logger.debug('%s: done %s', self.label, 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): 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 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: 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 logger.debug('%s: done parsing', self.label) self.job_tot = collections.defaultdict( functools.partial(collections.defaultdict, int)) for task_id, task_stat in self.task_stats.iteritems(): for category, stat_last in task_stat.iteritems(): for stat, val in stat_last.iteritems(): if stat in ['cpus', 'cache', 'swap', 'rss']: # meaningless stats like 16 cpu cores x 5 tasks = 80 continue self.job_tot[category][stat] += val logger.debug('%s: done totals', self.label) def long_label(self): label = self.label if self.finishtime: label += ' -- elapsed time ' s = (self.finishtime - self.starttime).total_seconds() if s > 86400: label += '{}d'.format(int(s/86400)) if s > 3600: label += '{}h'.format(int(s/3600) % 24) if s > 60: label += '{}m'.format(int(s/60) % 60) label += '{}s'.format(int(s) % 60) 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() def _text_report_gen(self): yield "\t".join(['category', 'metric', 'task_max', 'task_max_rate', 'job_total']) for category, stat_max in sorted(self.stats_max.iteritems()): for stat, val in sorted(stat_max.iteritems()): if stat.endswith('__rate'): continue max_rate = self._format(stat_max.get(stat+'__rate', '-')) val = self._format(val) tot = self._format(self.job_tot[category].get(stat, '-')) yield "\t".join([category, stat, str(val), max_rate, tot]) for args in ( ('Number of tasks: {}', len(self.tasks), None), ('Max CPU time spent by a single task: {}s', self.stats_max['cpu']['user+sys'], None), ('Max CPU usage in a single interval: {}%', self.stats_max['cpu']['user+sys__rate'], lambda x: x * 100), ('Overall CPU usage: {}%', self.job_tot['cpu']['user+sys'] / 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:keep0']['tx+rx'], lambda x: x / 1e9), ('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), ('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 if transform: val = transform(val) yield "# "+format_string.format(self._format(val)) def _recommend_gen(self): return itertools.chain( self._recommend_cpu(), self._recommend_ram(), self._recommend_keep_cache()) def _recommend_cpu(self): """Recommend asking for 4 cores if max CPU usage was 333%""" cpu_max_rate = self.stats_max['cpu']['user+sys__rate'] if cpu_max_rate == float('-Inf'): logger.warning('%s: no CPU usage data', self.label) return 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 {}% -- ' 'try runtime_constraints "min_cores_per_node":{}' ).format( self.label, int(math.ceil(cpu_max_rate*100)), int(used_cores)) def _recommend_ram(self): """Recommend an economical RAM constraint for this job. Nodes that are advertised as "8 gibibytes" actually have what we might call "8 nearlygibs" of memory available for jobs. Here, we calculate a whole number of nearlygibs that would have sufficed to run the job, then recommend requesting a node with that number of nearlygibs (expressed as mebibytes). Requesting a node with "nearly 8 gibibytes" is our best hope of getting a node that actually has nearly 8 gibibytes available. If the node manager is smart enough to account for the discrepancy itself when choosing/creating a node, we'll get an 8 GiB node with nearly 8 GiB available. Otherwise, the advertised size of the next-size-smaller node (say, 6 GiB) will be too low to satisfy our request, so we will effectively get rounded up to 8 GiB. For example, if we need 7500 MiB, we can ask for 7500 MiB, and we will generally get a node that is advertised as "8 GiB" and has at least 7500 MiB available. However, asking for 8192 MiB would either result in an unnecessarily expensive 12 GiB node (if node manager knows about the discrepancy), or an 8 GiB node which has less than 8192 MiB available and is therefore considered by crunch-dispatch to be too small to meet our constraint. When node manager learns how to predict the available memory for each node type such that crunch-dispatch always agrees that a node is big enough to run the job it was brought up for, all this will be unnecessary. We'll just ask for exactly the memory we want -- even if that happens to be 8192 MiB. """ used_bytes = self.stats_max['mem']['rss'] if used_bytes == float('-Inf'): logger.warning('%s: no memory usage data', self.label) return used_mib = math.ceil(float(used_bytes) / 1048576) asked_mib = self.existing_constraints.get('min_ram_mb_per_node') nearlygibs = lambda mebibytes: mebibytes/AVAILABLE_RAM_RATIO/1024 if asked_mib is None or ( math.ceil(nearlygibs(used_mib)) < nearlygibs(asked_mib)): yield ( '#!! {} max RSS was {} MiB -- ' 'try runtime_constraints "min_ram_mb_per_node":{}' ).format( self.label, 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. {:.2f} for floats, default format for everything else.""" if isinstance(val, float): return '{:.2f}'.format(val) else: return '{}'.format(val) class CollectionSummarizer(Summarizer): def __init__(self, collection_id, **kwargs): super(CollectionSummarizer, self).__init__( crunchstat_summary.reader.CollectionReader(collection_id), **kwargs) 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 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', {}) 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() for cname, component in instance['components'].iteritems(): if 'job' not in component: logger.warning( "%s: skipping component with no job assigned", cname) else: logger.info( "%s: job %s", cname, component['job']['uuid']) summarizer = JobSummarizer(component['job'], **kwargs) 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(): 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 = '' for cname, summarizer in self.summarizers.iteritems(): txt += '### Summary for {} ({})\n'.format( cname, summarizer.job['uuid']) txt += summarizer.text_report() txt += '\n' return txt def html_report(self): return crunchstat_summary.chartjs.ChartJS( self.label, self.summarizers.itervalues()).html()