X-Git-Url: https://git.arvados.org/arvados.git/blobdiff_plain/0b2a6fedb1a3e03a7c8173331f8ebcca1ca4a125..689901263ebfdd996da3711236615038e6245db3:/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 543011ea2c..884f16b4a7 100644 --- a/tools/crunchstat-summary/crunchstat_summary/summarizer.py +++ b/tools/crunchstat-summary/crunchstat_summary/summarizer.py @@ -22,7 +22,7 @@ from crunchstat_summary import logger # 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 - +MB=2**20 # Workaround datetime.datetime.strptime() thread-safety bug by calling # it once before starting threads. https://bugs.python.org/issue7980 @@ -167,9 +167,9 @@ class Summarizer(object): if task.finishtime is None or timestamp > task.finishtime: task.finishtime = timestamp - if self.starttime is None or timestamp < task.starttime: + if self.starttime is None or timestamp < self.starttime: self.starttime = timestamp - if self.finishtime is None or timestamp < task.finishtime: + if self.finishtime is None or timestamp > self.finishtime: self.finishtime = timestamp if (not self.detected_crunch1) and task.starttime is not None and task.finishtime is not None: @@ -207,17 +207,18 @@ class Summarizer(object): 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.items(): - if group == 'interval': - if stat == 'seconds': - this_interval_s = val - continue - elif not (this_interval_s > 0): + if group == 'interval': + if 'seconds' in stats: + this_interval_s = stats.get('seconds',0) + del stats['seconds'] + if this_interval_s <= 0: logger.error( "BUG? interval stat given with duration {!r}". format(this_interval_s)) - continue - else: + else: + logger.error('BUG? interval stat missing duration') + for stat, val in stats.items(): + if group == 'interval' and this_interval_s: stat = stat + '__rate' val = val / this_interval_s if stat in ['user+sys__rate', 'tx+rx__rate']: @@ -290,7 +291,7 @@ class Summarizer(object): self.stats_max['cpu']['user+sys__rate'], lambda x: x * 100), ('Overall CPU usage: {}%', - self.job_tot['cpu']['user+sys'] / + float(self.job_tot['cpu']['user+sys']) / self.job_tot['time']['elapsed'] if self.job_tot['time']['elapsed'] > 0 else 0, lambda x: x * 100), @@ -323,6 +324,7 @@ class Summarizer(object): yield "# "+format_string.format(self._format(val)) def _recommend_gen(self): + # TODO recommend fixing job granularity if elapsed time is too short return itertools.chain( self._recommend_cpu(), self._recommend_ram(), @@ -333,21 +335,27 @@ class Summarizer(object): constraint_key = self._map_runtime_constraint('vcpus') cpu_max_rate = self.stats_max['cpu']['user+sys__rate'] - if cpu_max_rate == float('-Inf'): + if cpu_max_rate == float('-Inf') or cpu_max_rate == 0.0: logger.warning('%s: no CPU usage data', self.label) return + # TODO Don't necessarily want to recommend on isolated max peak + # take average CPU usage into account as well or % time at max used_cores = max(1, int(math.ceil(cpu_max_rate))) asked_cores = self.existing_constraints.get(constraint_key) - if asked_cores is None or used_cores < asked_cores: + if asked_cores is None: + asked_cores = 1 + # TODO: This should be more nuanced in cases where max >> avg + if used_cores < asked_cores: yield ( '#!! {} max CPU usage was {}% -- ' - 'try runtime_constraints "{}":{}' + 'try reducing runtime_constraints to "{}":{}' ).format( self.label, - int(math.ceil(cpu_max_rate*100)), + math.ceil(cpu_max_rate*100), constraint_key, int(used_cores)) + # FIXME: This needs to be updated to account for current nodemanager algorithms def _recommend_ram(self): """Recommend an economical RAM constraint for this job. @@ -387,20 +395,20 @@ class Summarizer(object): if used_bytes == float('-Inf'): logger.warning('%s: no memory usage data', self.label) return - used_mib = math.ceil(float(used_bytes) / 1048576) + used_mib = math.ceil(float(used_bytes) / MB) asked_mib = self.existing_constraints.get(constraint_key) nearlygibs = lambda mebibytes: mebibytes/AVAILABLE_RAM_RATIO/1024 - if asked_mib is None or ( - math.ceil(nearlygibs(used_mib)) < nearlygibs(asked_mib)): + if used_mib > 0 and (asked_mib is None or ( + math.ceil(nearlygibs(used_mib)) < nearlygibs(asked_mib))): yield ( '#!! {} max RSS was {} MiB -- ' - 'try runtime_constraints "{}":{}' + 'try reducing runtime_constraints to "{}":{}' ).format( self.label, int(used_mib), constraint_key, - int(math.ceil(nearlygibs(used_mib))*AVAILABLE_RAM_RATIO*1024*(2**20)/self._runtime_constraint_mem_unit())) + int(math.ceil(nearlygibs(used_mib))*AVAILABLE_RAM_RATIO*1024*(MB)/self._runtime_constraint_mem_unit())) def _recommend_keep_cache(self): """Recommend increasing keep cache if utilization < 80%""" @@ -409,17 +417,18 @@ class Summarizer(object): return utilization = (float(self.job_tot['blkio:0:0']['read']) / float(self.job_tot['net:keep0']['rx'])) - asked_mib = self.existing_constraints.get(constraint_key, 256) + # FIXME: the default on this get won't work correctly + asked_cache = self.existing_constraints.get(constraint_key, 256) * self._runtime_constraint_mem_unit() if utilization < 0.8: yield ( '#!! {} Keep cache utilization was {:.2f}% -- ' - 'try runtime_constraints "{}":{} (or more)' + 'try doubling runtime_constraints to "{}":{} (or more)' ).format( self.label, utilization * 100.0, constraint_key, - asked_mib*2*(2**20)/self._runtime_constraint_mem_unit()) + math.ceil(asked_cache * 2 / self._runtime_constraint_mem_unit())) def _format(self, val): @@ -511,7 +520,7 @@ class ProcessSummarizer(Summarizer): class JobSummarizer(ProcessSummarizer): - runtime_constraint_mem_unit = 1048576 + runtime_constraint_mem_unit = MB map_runtime_constraint = { 'keep_cache_ram': 'keep_cache_mb_per_task', 'ram': 'min_ram_mb_per_node',