05.03.2021 04:22, John Snow wrote:
On 3/4/21 5:17 AM, Vladimir Sementsov-Ogievskiy wrote:
Sometimes one of cells in a testing table runs too slow. And we really
don't want to wait so long. Limit number of runs in this case.
Signed-off-by: Vladimir Sementsov-Ogievskiy <vsement...@virtuozzo.com>
---
scripts/simplebench/simplebench.py | 29 +++++++++++++++++++++++++----
1 file changed, 25 insertions(+), 4 deletions(-)
diff --git a/scripts/simplebench/simplebench.py
b/scripts/simplebench/simplebench.py
index f61513af90..b153cae274 100644
--- a/scripts/simplebench/simplebench.py
+++ b/scripts/simplebench/simplebench.py
@@ -19,9 +19,11 @@
#
import statistics
+import time
-def bench_one(test_func, test_env, test_case, count=5, initial_run=True):
+def bench_one(test_func, test_env, test_case, count=5, initial_run=True,
+ slow_limit=100):
"""Benchmark one test-case
test_func -- benchmarking function with prototype
@@ -36,6 +38,8 @@ def bench_one(test_func, test_env, test_case, count=5,
initial_run=True):
test_case -- test case - opaque second argument for test_func
count -- how many times to call test_func, to calculate average
initial_run -- do initial run of test_func, which don't get into result
+ slow_limit -- reduce test runs to 2, if current run exceedes the limit
+ (in seconds)
s/exceedes/exceeds, and you need to mention that if the initial run exceeds the
limit, it will change the behavior to count that result.
It is also possible (conceivably) that the initial run exceeds the limit, but
subsequent runs don't, so it might be hard to predict how many tests it'll
actually run.
If you're OK with that behavior, maybe:
"Consider a test run 'slow' once it exceeds this limit, in seconds.
Stop early once there are two 'slow' runs, including the initial run.
Slow initial runs will be included in the results."
Lastly, this will change existing behavior -- do we care? Should it default to
None instead? Should we be able to pass None or 0 to disable this behavior?
For sure I don't care about changing the behavior. Consider simplebench in a
version 0.0.1 :). Maybe, I should make a comment somewhere, but nobody will
read it anyway.
The aim of the patch is to minimize waiting for too long cells of the table,
which are obviously too much longer then the others. Probably the logic should
be improved a bit about ignoring or using initial-run result..
Like this:
If both initial and first run are slow, count both and stop here.
Otherwise, stop at first slow normal run and don't count initial run.
Or may be even
If both initial and first run are slow, count both and stop here.
Otherwise, behave the common way.
Returns dict with the following fields:
'runs': list of test_func results
@@ -47,17 +51,34 @@ def bench_one(test_func, test_env, test_case, count=5,
initial_run=True):
'n-failed': number of failed runs (exists only if at least one run
failed)
"""
+ runs = []
+ i = 0
if initial_run:
+ t = time.time()
+
print(' #initial run:')
- print(' ', test_func(test_env, test_case))
+ res = test_func(test_env, test_case)
+ print(' ', res)
+
+ if time.time() - t > slow_limit:
+ print(' - initial run is too slow, so it counts')
+ runs.append(res)
+ i = 1
+
+ for i in range(i, count):
+ t = time.time()
- runs = []
- for i in range(count):
print(' #run {}'.format(i+1))
res = test_func(test_env, test_case)
print(' ', res)
runs.append(res)
+ if time.time() - t > slow_limit and len(runs) >= 2:
+ print(' - run is too slow, and we have enough runs, stop here')
+ break
+
+ count = len(runs)
+
result = {'runs': runs}
succeeded = [r for r in runs if ('seconds' in r or 'iops' in r)]
--
Best regards,
Vladimir
