Increase reliability of per patch perf job

CSIT-1485

- Measurement results are inconsistent, creating false positives.
- To avoid false positives, sensitivity has to be lowered.
+ Measurements are no more interleaved.
+ Measurement result are no longer filtered.
+ The job now runs faster.
- Ip6base regressions smaller than 4% are tolerated.
+ Ip6base false positives can happen, but less than 10% of the time.
- Part of job parameters is set on ci-management level.
+ Voter script output is improved.

Change-Id: I0ec7d5df8b397daf7ff0277a1137ee9f36d8d866
Signed-off-by: Vratko Polak <vrpolak@cisco.com>

2 files changed, 49 insertions, 30 deletions

diff --git a/resources/libraries/bash/entry/per_patch_perf.sh b/resources/libraries/bash/entry/per_patch_perf.sh
index d3aabfae17..82dadc2ed6 100644
--- a/resources/libraries/bash/entry/per_patch_perf.sh
+++ b/resources/libraries/bash/entry/per_patch_perf.sh
@@ -58,17 +58,21 @@ select_tags || die
 compose_pybot_arguments || die
 generate_tests || die
 archive_tests || die
-iterations=8
+# Support for interleaved measurements is kept for future.
+iterations=1 # 8
 for ((iter=0; iter<iterations; iter++)); do
-    # TODO: Use less heavy way to avoid apt remove falilures.
-    # Also, reserve_testbed has already called cleanup once.
-    cleanup_topo
+    if ((iter)); then
+        # Reserve testbed has already cleaned it once,
+        # but we need to clean it explicitly on subsequent iterations.
+        cleanup_topo
+    fi
     select_build "build_parent" || die
     check_download_dir || die
     run_pybot || die
     copy_archives || die
     archive_parse_test_results "csit_parent/${iter}" || die
     die_on_pybot_error || die
+    # TODO: Use less heavy way to avoid apt remove failures.
     cleanup_topo
     select_build "build_current" || die
     check_download_dir || die
diff --git a/resources/tools/scripts/compare_perpatch.py b/resources/tools/scripts/compare_perpatch.py
index 3bda338d49..9c9bd20023 100644
--- a/resources/tools/scripts/compare_perpatch.py
+++ b/resources/tools/scripts/compare_perpatch.py
@@ -32,6 +32,8 @@ from jumpavg.BitCountingClassifier import BitCountingClassifier
 def hack(value_list):
     """Return middle two quartiles, hoping to reduce influence of outliers.
 
+    Currently "middle two" is "all", but that can change in future.
+
     :param value_list: List to pick subset from.
     :type value_list: list of float
     :returns: New list containing middle values.
@@ -40,16 +42,16 @@ def hack(value_list):
     tmp = sorted(value_list)
     eight = len(tmp) / 8
     ret = tmp[3*eight:-eight]
-    return ret
+    return tmp # ret
 
 iteration = -1
 parent_iterations = list()
-new_iterations = list()
+current_iterations = list()
 num_tests = None
 while 1:
     iteration += 1
     parent_lines = list()
-    new_lines = list()
+    current_lines = list()
     filename = "csit_parent/{iter}/results.txt".format(iter=iteration)
     try:
         with open(filename) as parent_file:
@@ -58,9 +60,9 @@ while 1:
         break
     num_lines = len(parent_lines)
     filename = "csit_current/{iter}/results.txt".format(iter=iteration)
-    with open(filename) as new_file:
-        new_lines = new_file.readlines()
-    if num_lines != len(new_lines):
+    with open(filename) as current_file:
+        current_lines = current_file.readlines()
+    if num_lines != len(current_lines):
         print "Number of tests does not match within iteration", iteration
         sys.exit(1)
     if num_tests is None:
@@ -69,37 +71,50 @@ while 1:
         print "Number of tests does not match previous at iteration", iteration
         sys.exit(1)
     parent_iterations.append(parent_lines)
-    new_iterations.append(new_lines)
+    current_iterations.append(current_lines)
 classifier = BitCountingClassifier()
 exit_code = 0
 for test_index in range(num_tests):
     val_max = 1.0
     parent_values = list()
-    new_values = list()
+    current_values = list()
     for iteration_index in range(len(parent_iterations)):
         parent_values.extend(
             json.loads(parent_iterations[iteration_index][test_index]))
-        new_values.extend(
-            json.loads(new_iterations[iteration_index][test_index]))
-    print "TRACE pre-hack parent: {p}".format(p=parent_values)
-    print "TRACE pre-hack current: {n}".format(n=new_values)
+        current_values.extend(
+            json.loads(current_iterations[iteration_index][test_index]))
+    print "Time-ordered MRR values for parent build: {p}".format(
+        p=parent_values)
+    print "Time-ordered MRR values for current build: {c}".format(
+        c=current_values)
     parent_values = hack(parent_values)
-    new_values = hack(new_values)
+    current_values = hack(current_values)
     parent_max = BitCountingMetadataFactory.find_max_value(parent_values)
-    new_max = BitCountingMetadataFactory.find_max_value(new_values)
-    val_max = max(val_max, parent_max, new_max)
+    current_max = BitCountingMetadataFactory.find_max_value(current_values)
+    val_max = max(val_max, parent_max, current_max)
     factory = BitCountingMetadataFactory(val_max)
     parent_stats = factory.from_data(parent_values)
-    new_factory = BitCountingMetadataFactory(val_max, parent_stats.avg)
-    new_stats = new_factory.from_data(new_values)
-    print "TRACE parent: {p}".format(p=parent_values)
-    print "TRACE current: {n}".format(n=new_values)
-    print "DEBUG parent: {p}".format(p=parent_stats)
-    print "DEBUG current: {n}".format(n=new_stats)
-    common_max = max(parent_stats.avg, new_stats.avg)
-    difference = (new_stats.avg - parent_stats.avg) / common_max
-    print "DEBUG difference: {d}%".format(d=100 * difference)
-    classified_list = classifier.classify([parent_stats, new_stats])
+    current_factory = BitCountingMetadataFactory(val_max, parent_stats.avg)
+    current_stats = current_factory.from_data(current_values)
+    both_stats = factory.from_data(parent_values + current_values)
+    print "Value-ordered MRR values for parent build: {p}".format(
+        p=parent_values)
+    print "Value-ordered MRR values for current build: {c}".format(
+        c=current_values)
+    difference = (current_stats.avg - parent_stats.avg) / parent_stats.avg
+    print "Difference of averages relative to parent: {d}%".format(
+        d=100 * difference)
+    print "Jumpavg representation of parent group: {p}".format(
+        p=parent_stats)
+    print "Jumpavg representation of current group: {c}".format(
+        c=current_stats)
+    print "Jumpavg representation of both as one group: {b}".format(
+        b=both_stats)
+    bits = parent_stats.bits + current_stats.bits - both_stats.bits
+    compared = "longer" if bits >= 0 else "shorter"
+    print "Separate groups are {cmp} than single group by {bit} bits".format(
+        cmp=compared, bit=abs(bits))
+    classified_list = classifier.classify([parent_stats, current_stats])
     if len(classified_list) < 2:
         print "Test test_index {test_index}: normal (no anomaly)".format(
             test_index=test_index)
@@ -112,5 +127,5 @@ for test_index in range(num_tests):
         continue
     print "Test test_index {test_index}: anomaly {anomaly}".format(
         test_index=test_index, anomaly=anomaly)
-print "DEBUG exit code {code}".format(code=exit_code)
+print "Exit code {code}".format(code=exit_code)
 sys.exit(exit_code)