Merge PR #103: Reduce memory usage and calculate statistics online/in…

…crementally

.travis.yml

@@ -3,13 +3,6 @@ language: python
 matrix:
   include:
     - python: "2.7"
-      virtualenv:
-        system_site_packages: true
-      addons:
-        apt:
-          packages:
-            - python-scipy
-
     - python: "3.6"
     # PyPy versions
     - python: pypy

rebench/executor.py

@@ -35,7 +35,6 @@
 
 from . import subprocess_with_timeout as subprocess_timeout
 from .interop.adapter import ExecutionDeliveredNoResults
-from .statistics import StatisticProperties, mean
 
 
 class FailedBuilding(Exception):
@@ -83,14 +82,10 @@ def _indicate_progress(self, completed_task, run):
         if not self._ui.spinner_initialized():
             return
 
-        totals = run.get_total_values()
         if completed_task:
             self._runs_completed += 1
 
-        if totals:
-            art_mean = mean(run.get_total_values())
-        else:
-            art_mean = 0
+        art_mean = run.get_mean_of_totals()
 
         hour, minute, sec = self._estimate_time_left()
 
@@ -419,15 +414,15 @@ def execute_run(self, run_id):
             terminate = self._generate_data_point(cmdline, gauge_adapter,
                                                   run_id, termination_check)
 
-        stats = StatisticProperties(run_id.get_total_values())
+        mean_of_totals = run_id.get_mean_of_totals()
         if terminate:
-            run_id.report_run_completed(stats, cmdline)
+            run_id.report_run_completed(cmdline)
             if (not run_id.is_failed() and run_id.min_iteration_time
-                    and stats.mean < run_id.min_iteration_time):
+                    and mean_of_totals < run_id.min_iteration_time):
                 self._ui.warning(
                     ("{ind}Warning: Low mean run time.\n"
                      + "{ind}{ind}The mean (%.1f) is lower than min_iteration_time (%d)\n")
-                    % (stats.mean, run_id.min_iteration_time), run_id, cmdline)
+                    % (mean_of_totals, run_id.min_iteration_time), run_id, cmdline)
 
         return terminate
 

rebench/model/data_point.py

@@ -27,6 +27,10 @@ def __init__(self, run_id):
         self._total = None
         self._invocation = -1
 
+    @property
+    def run_id(self):
+        return self._run_id
+
     @property
     def invocation(self):
         return self._invocation

rebench/model/run_id.py

@@ -21,6 +21,7 @@
 
 from .benchmark import Benchmark
 from .termination_check import TerminationCheck
+from ..statistics import StatisticProperties
 from ..ui import UIError
 
 
@@ -34,7 +35,8 @@ def __init__(self, benchmark, cores, input_size, var_value):
 
         self._reporters = set()
         self._persistence = set()
-        self._data_points = []
+        self._statistics = StatisticProperties()
+        self._total_unit = None
 
         self._termination_check = None
         self._cmdline = None
@@ -136,9 +138,9 @@ def report_run_failed(self, cmdline, return_code, output):
         for reporter in self._reporters:
             reporter.run_failed(self, cmdline, return_code, output)
 
-    def report_run_completed(self, statistics, cmdline):
+    def report_run_completed(self, cmdline):
         for reporter in self._reporters:
-            reporter.run_completed(self, statistics, cmdline)
+            reporter.run_completed(self, self._statistics, cmdline)
 
     def report_job_completed(self, run_ids):
         for reporter in self._reporters:
@@ -152,42 +154,44 @@ def report_start_run(self):
         for reporter in self._reporters:
             reporter.start_run(self)
 
+    def is_persisted_by(self, persistence):
+        return persistence in self._persistence
+
     def add_persistence(self, persistence):
         self._persistence.add(persistence)
 
     def close_files(self):
         for persistence in self._persistence:
             persistence.close()
 
-    def loaded_data_point(self, data_point):
+    def _new_data_point(self, data_point):
         self._max_invocation = max(self._max_invocation, data_point.invocation)
-        self._data_points.append(data_point)
+        if self._total_unit is None:
+            self._total_unit = data_point.get_total_unit()
+
+    def loaded_data_point(self, data_point):
+        self._new_data_point(data_point)
+        self._statistics.add_sample(data_point.get_total_value())
 
     def add_data_point(self, data_point, warmup):
-        self._max_invocation = max(self._max_invocation, data_point.invocation)
+        self._new_data_point(data_point)
 
         if not warmup:
-            self._data_points.append(data_point)
+            self._statistics.add_sample(data_point.get_total_value())
         for persistence in self._persistence:
             persistence.persist_data_point(data_point)
 
     def get_number_of_data_points(self):
-        return len(self._data_points)
+        return self._statistics.num_samples
 
-    def get_data_points(self):
-        return self._data_points
+    def get_mean_of_totals(self):
+        return self._statistics.mean
 
-    def discard_data_points(self):
-        self._data_points = []
-        self._max_invocation = 0
-
-    def get_total_values(self):
-        return [dp.get_total_value() for dp in self._data_points]
+    def get_statistics(self):
+        return self._statistics
 
     def get_total_unit(self):
-        if not self._data_points:
-            return None
-        return self._data_points[0].get_total_unit()
+        return self._total_unit
 
     def get_termination_check(self, ui):
         if self._termination_check is None:
@@ -202,7 +206,7 @@ def is_completed(self, ui):
     def run_failed(self):
         return (self._termination_check.fails_consecutively() or
                 self._termination_check.has_too_many_failures(
-                    len(self._data_points)))
+                    self.get_number_of_data_points()))
 
     def __hash__(self):
         return hash(self.cmdline())
@@ -236,18 +240,22 @@ def _expand_vars(self, string):
     def cmdline(self):
         if self._cmdline:
             return self._cmdline
+        return self._construct_cmdline()
 
+    def _construct_cmdline(self):
         cmdline = ""
         if self._benchmark.suite.executor.path:
-            cmdline = "%s/" % (self._benchmark.suite.executor.path, )
+            cmdline = self._benchmark.suite.executor.path + "/"
+
+        cmdline += self._benchmark.suite.executor.executable
 
-        cmdline += "%s %s" % (self._benchmark.suite.executor.executable,
-                              self._benchmark.suite.executor.args or '')
+        if self._benchmark.suite.executor.args:
+            cmdline += " " + str(self._benchmark.suite.executor.args)
 
-        cmdline += self._benchmark.suite.command
+        cmdline += " " + self._benchmark.suite.command
 
-        if self._benchmark.extra_args is not None:
-            cmdline += " %s" % self._benchmark.extra_args
+        if self._benchmark.extra_args:
+            cmdline += " " + str(self._benchmark.extra_args)
 
         cmdline = self._expand_vars(cmdline)
 

rebench/persistence.py

@@ -18,9 +18,11 @@
 # FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
 # IN THE SOFTWARE.
 import os
+import shutil
 import subprocess
 import sys
 from datetime import datetime
+from tempfile import NamedTemporaryFile
 from threading import Lock
 
 from .model.data_point  import DataPoint
@@ -36,9 +38,9 @@ def __init__(self, ui):
         self._bench_cfgs = {}
         self._ui = ui
 
-    def load_data(self):
+    def load_data(self, runs, discard_run_data):
         for persistence in list(self._files.values()):
-            persistence._load_data()
+            persistence.load_data(runs, discard_run_data)
 
     def get(self, filename, discard_old_data):
         if filename not in self._files:
@@ -81,41 +83,6 @@ def register_config(self, cfg):
             self._bench_cfgs[key] = cfg
         return cfg
 
-    @classmethod
-    def get_by_file(cls, runs):
-        by_file = {}
-        for run in runs:
-            points = run.get_data_points()
-            run.discard_data_points()
-            for point in points:
-                measurements = point.get_measurements()
-                for measure in measurements:
-                    if measure.filename in by_file:
-                        by_file[measure.filename].append(measure)
-                    else:
-                        by_file[measure.filename] = [measure]
-        return by_file
-
-    @classmethod
-    def discard_data_of_runs(cls, runs, ui):
-        by_file = cls.get_by_file(runs)
-        for filename, measures in by_file.items():
-            try:
-                with open(filename, 'r') as data_file:
-                    lines = data_file.readlines()
-            except IOError:
-                ui.debug_error_info(
-                    "Tried to discard old data, but file does not seem to exist: %s\n" % filename)
-                continue
-
-            for measure in measures:
-                lines[measure.line_number] = None
-
-            lines = filter(None, lines)
-
-            with open(filename, 'w') as data_file:
-                data_file.writelines(lines)
-
 
 class _DataPointPersistence(object):
 
@@ -141,18 +108,30 @@ def _truncate_file(filename):
         with open(filename, 'w'):
             pass
 
-    def _load_data(self):
+    def load_data(self, runs, discard_run_data):
         """
         Loads the data from the configured data file
         """
+        if discard_run_data:
+            current_runs = {run for run in runs if run.is_persisted_by(self)}
+        else:
+            current_runs = None
+
         try:
-            with open(self._data_filename, 'r') as data_file:
-                self._process_lines(data_file)
+            if current_runs:
+                with NamedTemporaryFile('w', delete=False) as target:
+                    with open(self._data_filename, 'r') as data_file:
+                        self._process_lines(data_file, current_runs, target)
+                    os.unlink(self._data_filename)
+                    shutil.move(target.name, self._data_filename)
+            else:
+                with open(self._data_filename, 'r') as data_file:
+                    self._process_lines(data_file, current_runs, None)
         except IOError:
             self._ui.debug_error_info("No data loaded, since %s does not exist.\n"
                                       % self._data_filename)
 
-    def _process_lines(self, data_file):
+    def _process_lines(self, data_file, runs, filtered_data_file):
         """
          The most important assumptions we make here is that the total
          measurement is always the last one serialized for a data point.
@@ -165,6 +144,8 @@ def _process_lines(self, data_file):
         for line in data_file:
             if line.startswith('#'):  # skip comments, and shebang lines
                 line_number += 1
+                if filtered_data_file:
+                    filtered_data_file.write(line)
                 continue
 
             try:
@@ -173,6 +154,13 @@ def _process_lines(self, data_file):
                     line_number, self._data_filename)
 
                 run_id = measurement.run_id
+                if filtered_data_file and runs and run_id in runs:
+                    continue
+
+                # these are all the measurements that are not filtered out
+                if filtered_data_file:
+                    filtered_data_file.write(line)
+
                 if previous_run_id is not run_id:
                     data_point = DataPoint(run_id)
                     previous_run_id = run_id

rebench/rebench.py

@@ -203,22 +203,16 @@ def run(self, argv=None):
         except ConfigurationError as exc:
             raise UIError(exc.message + "\n", exc)
 
-        data_store.load_data()
-        return self.execute_experiment()
+        runs = self._config.get_runs()
+        data_store.load_data(runs, self._config.options.do_rerun)
+        return self.execute_experiment(runs)
 
-    def execute_experiment(self):
+    def execute_experiment(self, runs):
         self._ui.verbose_output_info("Execute experiment: " + self._config.experiment_name + "\n")
 
-        # first load old data if available
-        if self._config.options.clean:
-            pass
-
         scheduler_class = {'batch':       BatchScheduler,
                            'round-robin': RoundRobinScheduler,
                            'random':      RandomScheduler}.get(self._config.options.scheduler)
-        runs = self._config.get_runs()
-        if self._config.options.do_rerun:
-            DataStore.discard_data_of_runs(runs, self._ui)
 
         executor = Executor(runs, self._config.use_nice, self._config.do_builds,
                             self._ui,

rebench/reporter.py

@@ -36,8 +36,6 @@
     from urllib import urlencode # pylint: disable=ungrouped-imports
     from urllib2 import urlopen
 
-from .statistics import StatisticProperties
-
 
 class Reporter(object):
 
@@ -80,13 +78,14 @@ def _generate_all_output(run_ids):
         rows = []
 
         for run_id in run_ids:
-            stats = StatisticProperties(run_id.get_total_values())
+            mean = run_id.get_mean_of_totals()
+            num_samples = run_id.get_number_of_data_points()
             out = run_id.as_str_list()
-            out.append(stats.num_samples)
-            if stats.num_samples == 0:
+            out.append(num_samples)
+            if num_samples == 0:
                 out.append("Failed")
             else:
-                out.append(int(round(stats.mean, 0)))
+                out.append(int(round(mean, 0)))
             rows.append(out)
 
         return sorted(rows, key=itemgetter(2, 1, 3, 4, 5, 6, 7))
@@ -248,8 +247,7 @@ def _send_to_codespeed(self, results, run_id):
                     + "{ind}{ind}" + msg + "\n", run_id)
 
     def _prepare_result(self, run_id):
-        stats = StatisticProperties(run_id.get_total_values())
-        return self._format_for_codespeed(run_id, stats)
+        return self._format_for_codespeed(run_id, run_id.get_statistics())
 
     def report_job_completed(self, run_ids):
         if self._incremental_report: