performance comparison

docker/test/performance-comparison/compare.sh

@@ -129,7 +129,7 @@ function run_tests
     rm right/performance/{IPv4,IPv6,modulo,parse_engine_file,number_formatting_formats,select_format}.xml ||:
 
     # Run the tests
-    for test in right/performance/${CHPC_TEST_GLOB:-*.xml}
+    for test in right/performance/${CHPC_TEST_GLOB:-*}.xml
     do
         test_name=$(basename $test ".xml")
         echo test $test_name
@@ -150,22 +150,24 @@ rm test-times.tsv test-dump.tsv unstable.tsv changed-perf.tsv unstable-tests.tsv
 right/clickhouse local --query "
 create table queries engine Memory as select
         replaceAll(_file, '-report.tsv', '') test,
-        if(abs(diff) < 0.05 and rd[3] > 0.05,      1, 0) unstable,
+        left + right < 0.01 as short,
-        if(abs(diff) > 0.05 and abs(diff) > rd[3], 1, 0) changed,
-        *
-    from file('*-report.tsv', TSV, 'left float, right float, diff float, rd Array(float), query text')
         -- FIXME Comparison mode doesn't make sense for queries that complete
         -- immediately, so for now we pretend they don't exist. We don't want to
         -- remove them altogether because we want to be able to detect regressions,
         -- but the right way to do this is not yet clear.
-    where left + right > 0.01;
+        not short and abs(diff) < 0.05 and rd[3] > 0.05 as unstable,
+        not short and abs(diff) > 0.05 and abs(diff) > rd[3] as changed,
+        *
+    from file('*-report.tsv', TSV, 'left float, right float, diff float, rd Array(float), query text');
 
 create table changed_perf_tsv engine File(TSV, 'changed-perf.tsv') as
     select left, right, diff, rd, test, query from queries where changed
     order by rd[3] desc;
+
 create table unstable_queries_tsv engine File(TSV, 'unstable-queries.tsv') as
     select left, right, diff, rd, test, query from queries where unstable
     order by rd[3] desc;
+
 create table unstable_tests_tsv engine File(TSV, 'bad-tests.tsv') as
     select test, sum(unstable) u, sum(changed) c, u + c s from queries
     group by test having s > 0 order by s desc;
@@ -182,20 +184,23 @@ create table slow_on_client_tsv engine File(TSV, 'slow-on-client.tsv') as
 
 create table test_time engine Memory as
     select test, sum(client) total_client_time,
-        max(client) query_max, min(client) query_min, count(*) queries
+        maxIf(client, not short) query_max,
-    from query_time
+        minIf(client, not short) query_min,
-    -- for consistency, filter out everything we filtered out of queries table
+        count(*) queries,
-    semi join queries using query
+        sum(short) short_queries
+    from query_time, queries
+    where query_time.query = queries.query
     group by test;
 
 create table test_times_tsv engine File(TSV, 'test-times.tsv') as
     select wall_clock.test, real,
         floor(total_client_time, 3),
         queries,
+        short_queries,
         floor(query_max, 3),
         floor(real / queries, 3) avg_real_per_query,
         floor(query_min, 3)
-    from test_time right join wall_clock using test
+    from test_time join wall_clock using test
     order by query_max / query_min desc;
 
 create table all_queries_tsv engine File(TSV, 'all-queries.tsv') as

docker/test/performance-comparison/perf.py

@@ -94,10 +94,11 @@ test_query_templates = [q.text for q in root.findall('query')]
 test_queries = substitute_parameters(test_query_templates, parameter_combinations)
 
 for q in test_queries:
-    # Warmup: run once on both servers. Helps to bring the data into memory,
+    # Prewarm: run once on both servers. Helps to bring the data into memory,
     # precompile the queries, etc.
-    for c in connections:
+    for conn_index, c in enumerate(connections):
-        c.execute(q)
+        res = c.execute(q)
+        print('prewarm\t' + tsv_escape(q) + '\t' + str(conn_index) + '\t' + str(c.last_query.elapsed))
 
     # Now, perform measured runs.
     # Track the time spent by the client to process this query, so that we can notice

docker/test/performance-comparison/report.py

@@ -81,7 +81,8 @@ def nextRowAnchor():
 
 def tr(x):
     a = nextRowAnchor()
-    return '<tr onclick="location.href=\'#{a}\'" id={a}>{x}</tr>'.format(a=a, x=str(x))
+    #return '<tr onclick="location.href=\'#{a}\'" id={a}>{x}</tr>'.format(a=a, x=str(x))
+    return '<tr id={a}>{x}</tr>'.format(a=a, x=str(x))
 
 def td(x):
     return '<td>' + str(x) + '</td>'
@@ -140,16 +141,17 @@ params['test_part'] = (
     table_template.format(
         anchor = nextTableAnchor(),
         caption = 'Tests with most unstable queries',
-        header = table_header(['Test', 'Unstable', 'Changed perf', 'Total']),
+        header = table_header(['Test', 'Unstable', 'Changed perf', 'Total not OK']),
         rows = tsv_rows('bad-tests.tsv')) +
     table_template.format(
         anchor = nextTableAnchor(),
         caption = 'Tests times',
         header = table_header(['Test', 'Wall clock time, s', 'Total client time, s',
             'Number of queries',
-            'Max client time<br>(sum for all runs), s',
+            'Number of short queries',
+            'Longest query<br>(sum for all runs), s',
             'Avg wall clock time<br>(sum for all runs), s',
-            'Min client time<br>(sum for all runs), s']),
+            'Shortest query<br>(sum for all runs), s']),
         rows = tsv_rows('test-times.tsv'))
 )
 print(doc_template.format_map(params))