public
/
ottertune
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

When ranking knobs in the background tasks use the target objective(s) by default. Added an option that when enabled also uses the pruned metrics.

This commit is contained in:
dvanaken 2020-04-13 23:54:44 -04:00 committed by Dana Van Aken
parent aeaae7d9f4
commit efa02899b5
3 changed files with 34 additions and 12 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
server/website/website/admin.py
View File

@ -146,7 +146,7 @@ class TaskMetaAdmin(admin.ModelAdmin):
fields = readonly_fields fields = readonly_fields
list_filter = ('status',) list_filter = ('status',)
list_per_page = 10 list_per_page = 10
ordering = ('date_done',) ordering = ('-date_done',)
max_field_length = 1000 max_field_length = 1000
@staticmethod @staticmethod

13
server/website/website/settings/constants.py
View File

@ -11,3 +11,16 @@ CHECK_CELERY = True
# address categorical knobs (enum, boolean) # address categorical knobs (enum, boolean)
ENABLE_DUMMY_ENCODER = False ENABLE_DUMMY_ENCODER = False
# Whether to include the pruned metrics from the workload characterization subtask in
# the output (y) when ranking the knobs for a given workload in the knob identification
# subtask.
# When computing the ranked knobs in the knob identification subtask, the output (y) is
# the set of target objectives used to tune the given workload. If this flag is enabled
# then the pruned metrics from the workload characterization subtask are also included
# in the output. (See website/tasks/periodic_tasks.py)
KNOB_IDENT_USE_PRUNED_METRICS = False
# The background tasks only process workloads containing this minimum amount of results
MIN_WORKLOAD_RESULTS_COUNT = 5

31
server/website/website/tasks/periodic_tasks.py
View File

@ -21,14 +21,13 @@ from analysis.preprocessing import (Bin, get_shuffle_indices,
DummyEncoder, DummyEncoder,
consolidate_columnlabels) consolidate_columnlabels)
from website.models import PipelineData, PipelineRun, Result, Workload, ExecutionTime from website.models import PipelineData, PipelineRun, Result, Workload, ExecutionTime
from website.settings import ENABLE_DUMMY_ENCODER, TIME_ZONE from website.settings import (ENABLE_DUMMY_ENCODER, KNOB_IDENT_USE_PRUNED_METRICS,
MIN_WORKLOAD_RESULTS_COUNT, TIME_ZONE)
from website.types import PipelineTaskType, WorkloadStatusType from website.types import PipelineTaskType, WorkloadStatusType
from website.utils import DataUtil, JSONUtil from website.utils import DataUtil, JSONUtil
# Log debug messages # Log debug messages
LOG = get_task_logger(__name__) LOG = get_task_logger(__name__)
# Only process workload containing this minimum amount of results
MIN_WORKLOAD_RESULTS_COUNT = 5
def save_execution_time(start_ts, fn): def save_execution_time(start_ts, fn):
@ -134,21 +133,31 @@ def run_background_tasks():
creation_time=now()) creation_time=now())
pruned_metrics_entry.save() pruned_metrics_entry.save()
# Use the pruned metrics to filter the metric_data # Workload target objective data
pruned_metric_idxs = [i for i, metric_name in enumerate(metric_data['columnlabels']) ranked_knob_metrics = sorted(wkld_results.distinct('session').values_list(
if metric_name in pruned_metrics] 'session__target_objective', flat=True).distinct())
pruned_metric_data = { LOG.debug("Target objectives for workload %s: %s", workload_name,
'data': metric_data['data'][:, pruned_metric_idxs], ', '.join(ranked_knob_metrics))
if KNOB_IDENT_USE_PRUNED_METRICS:
ranked_knob_metrics = sorted(set(ranked_knob_metrics) + set(pruned_metrics))
# Use the set of metrics to filter the metric_data
metric_idxs = [i for i, metric_name in enumerate(metric_data['columnlabels'])
if metric_name in ranked_knob_metrics]
ranked_metric_data = {
'data': metric_data['data'][:, metric_idxs],
'rowlabels': copy.deepcopy(metric_data['rowlabels']), 'rowlabels': copy.deepcopy(metric_data['rowlabels']),
'columnlabels': [metric_data['columnlabels'][i] for i in pruned_metric_idxs] 'columnlabels': [metric_data['columnlabels'][i] for i in metric_idxs]
} }
# Execute the Knob Identification task to compute an ordered list of knobs # Execute the Knob Identification task to compute an ordered list of knobs
# ranked by their impact on the DBMS's performance. Save them in a new # ranked by their impact on the DBMS's performance. Save them in a new
# PipelineData object. # PipelineData object.
LOG.info("Ranking knobs for workload %s...", workload_name) LOG.info("Ranking knobs for workload %s (use pruned metric data: %s)...",
workload_name, KNOB_IDENT_USE_PRUNED_METRICS)
ranked_knobs = run_knob_identification(knob_data=knob_data, ranked_knobs = run_knob_identification(knob_data=knob_data,
metric_data=pruned_metric_data, metric_data=ranked_metric_data,
dbms=workload.dbms) dbms=workload.dbms)
LOG.info("Done ranking knobs for workload %s (# ranked knobs: %s).\n\n" LOG.info("Done ranking knobs for workload %s (# ranked knobs: %s).\n\n"
"Ranked knobs: %s\n", workload_name, len(ranked_knobs), ranked_knobs) "Ranked knobs: %s\n", workload_name, len(ranked_knobs), ranked_knobs)