diff --git a/docker/test/performance-comparison/compare.sh b/docker/test/performance-comparison/compare.sh
new file mode 100755
index 00000000000..4cd3e39522e
--- /dev/null
+++ b/docker/test/performance-comparison/compare.sh
@@ -0,0 +1,97 @@
+#!/bin/bash
+set -ex
+set -o pipefail
+trap "exit" INT TERM
+trap "kill 0" EXIT
+
+mkdir left ||:
+mkdir right ||:
+mkdir db0 ||:
+
+left_pr=$1
+left_sha=$2
+
+right_pr=$3
+right_sha=$4
+
+function download
+{
+    la="$left_pr-$left_sha.tgz"
+    ra="$right_pr-$right_sha.tgz"
+    wget -nd -c "https://clickhouse-builds.s3.yandex.net/$left_pr/$left_sha/performance/performance.tgz" -O "$la" && tar -C left --strip-components=1 -zxvf "$la" &
+    wget -nd -c "https://clickhouse-builds.s3.yandex.net/$right_pr/$right_sha/performance/performance.tgz" -O "$ra" && tar -C right --strip-components=1 -zxvf "$ra" &
+    cd db0 && wget -nd -c "https://s3.mds.yandex.net/clickhouse-private-datasets/hits_10m_single/partitions/hits_10m_single.tar" && tar -xvf hits_10m_single.tar &
+    cd db0 && wget -nd -c "https://s3.mds.yandex.net/clickhouse-private-datasets/hits_100m_single/partitions/hits_100m_single.tar" && tar -xvf hits_100m_single.tar &
+    #cd db0 && wget -nd -c "https://clickhouse-datasets.s3.yandex.net/hits/partitions/hits_v1.tar" && tar -xvf hits_v1.tar &
+    wait
+
+    # Use hardlinks instead of copying
+    rm -r left/db ||:
+    rm -r right/db ||:
+    cp -al db0/ left/db/
+    cp -al db0/ right/db/
+}
+
+#download
+
+function configure
+{
+    sed -i 's/<tcp_port>9000/<tcp_port>9001/g' right/config/config.xml
+
+    cat > right/config/config.d/perf-test-tweaks.xml <<EOF
+    <yandex>
+        <logger>
+            <console>true</console>
+        </logger>
+        <text_log remove="remove"/>
+    </yandex>
+EOF
+
+    cp right/config/config.d/perf-test-tweaks.xml left/config/config.d/perf-test-tweaks.xml
+}
+
+configure
+
+
+function restart
+{
+    while killall clickhouse ; do echo . ; sleep 1 ; done
+    echo all killed
+
+    # Spawn servers in their own process groups
+    set -m
+
+    left/clickhouse server --config-file=left/config/config.xml -- --path left/db &> left/log.txt &
+    left_pid=$!
+    kill -0 $left_pid
+    disown $left_pid
+
+    right/clickhouse server --config-file=right/config/config.xml -- --path right/db &> right/log.txt &
+    right_pid=$!
+    kill -0 $right_pid
+    disown $right_pid
+
+    set +m
+
+    while ! left/clickhouse client --query "select 1" ; do kill -0 $left_pid ; echo . ; sleep 1 ; done
+    echo left ok
+
+    while ! right/clickhouse client --port 9001 --query "select 1" ; do kill -0 $right_pid ; echo . ; sleep 1 ; done
+    echo right ok
+}
+
+restart
+
+for test in ch/dbms/tests/performance/*.xml
+do
+    test_name=$(basename $test ".xml")
+    ./perf.py "$test" > "$test_name-raw.tsv" || continue
+    right/clickhouse local --file "$test_name-raw.tsv" --structure 'query text, run int, version UInt32, time float' --query "$(cat eqmed.sql)" > "$test_name-report.tsv"
+done
+
+#while killall clickhouse ; do echo . ; sleep 1 ; done
+#echo ok
+
+
+
+
diff --git a/docker/test/performance-comparison/eqmed.sql b/docker/test/performance-comparison/eqmed.sql
new file mode 100644
index 00000000000..dd677b750a0
--- /dev/null
+++ b/docker/test/performance-comparison/eqmed.sql
@@ -0,0 +1,41 @@
+-- input is table(query text, run UInt32, version int, time float)
+select
+   abs(diff_percent) > rd_quantiles_percent[3] fail, 
+   floor(original_medians_array.time_by_version[1], 4) m1,
+   floor(original_medians_array.time_by_version[2], 4) m2,
+   floor((m1 - m2) / m1, 3) diff_percent,
+   arrayMap(x -> floor(x / m1, 3), rd.rd_quantiles) rd_quantiles_percent,
+   query 
+from
+   (
+      select query, quantiles(0.05, 0.5, 0.95)(abs(time_by_label[1] - time_by_label[2])) rd_quantiles -- quantiles of randomization distribution
+      from
+         (
+            select query, virtual_run, groupArrayInsertAt(median_time, random_label) time_by_label -- make array 'random label' -> 'median time'
+            from (
+                  select query, medianExact(time) median_time, virtual_run, random_label -- get median times, grouping by random label
+                  from (
+                        select *, toUInt32(rowNumberInBlock() % 2) random_label -- randomly relabel measurements
+                        from (
+                              select query, time, number virtual_run 
+                              from table, numbers(1, 10000) -- duplicate input measurements into many virtual runs
+                              order by query, virtual_run, rand() -- for each virtual run, randomly reorder measurements
+                           ) virtual_runs
+                     ) relabeled 
+                  group by query, virtual_run, random_label
+               ) virtual_medians
+            group by query, virtual_run -- aggregate by random_label
+         ) virtual_medians_array
+      group by query -- aggregate by virtual_run
+   ) rd,
+   (
+        select groupArrayInsertAt(median_time, version) time_by_version, query
+        from
+           (
+                select medianExact(time) median_time, query, version
+                from table group by query, version
+           ) original_medians
+        group by query
+   ) original_medians_array
+where rd.query = original_medians_array.query
+order by fail desc, rd_quantiles_percent[3] asc;
diff --git a/docker/test/performance-comparison/perf.py b/docker/test/performance-comparison/perf.py
new file mode 100755
index 00000000000..508d46b6261
--- /dev/null
+++ b/docker/test/performance-comparison/perf.py
@@ -0,0 +1,87 @@
+#!/usr/bin/python3
+
+import itertools
+import clickhouse_driver
+import xml.etree.ElementTree as et
+import argparse
+import pprint
+
+parser = argparse.ArgumentParser(description='Run performance test.')
+parser.add_argument('file', metavar='FILE', type=argparse.FileType('r'), nargs=1, help='test description file')
+args = parser.parse_args()
+
+tree = et.parse(args.file[0])
+root = tree.getroot()
+
+# Check main metric
+main_metric = root.find('main_metric/*').tag
+if main_metric != 'min_time':
+    raise Exception('Only the min_time main metric is supported. This test uses \'{}\''.format(main_metric))
+
+# Open connections
+servers = [{'host': 'localhost', 'port': 9000, 'client_name': 'left'}, {'host': 'localhost', 'port': 9001, 'client_name': 'right'}]
+connections = [clickhouse_driver.Client(**server) for server in servers]
+
+# Check tables that should exist
+tables = [e.text for e in root.findall('preconditions/table_exists')]
+if tables:
+    for c in connections:
+        tables_list = ", ".join("'{}'".format(t) for t in tables)
+        res = c.execute("select t from values('t text', {}) anti join system.tables on database = currentDatabase() and name = t".format(tables_list))
+        if res:
+            raise Exception('Some tables are not found: {}'.format(res))
+
+# Process substitutions
+subst_elems = root.findall('substitutions/substitution')
+
+parameter_keys = []         # ['table',                   'limit'    ]
+parameter_value_arrays = [] # [['hits_100m', 'hits_10m'], ['1', '10']]
+parameter_combinations = [] # [{table: hits_100m, limit: 1}, ...]
+for se in subst_elems:
+    parameter_keys.append(se.find('name').text)
+    parameter_value_arrays.append([v.text for v in se.findall('values/value')])
+parameter_combinations = [dict(zip(parameter_keys, parameter_combination)) for parameter_combination in itertools.product(*parameter_value_arrays)]
+
+def substitute_parameters(query_templates, parameter_combinations):
+    return list(set([template.format(**parameters) for template, parameters in itertools.product(query_templates, parameter_combinations)]))
+
+# Run drop queries, ignoring errors
+drop_query_templates = [q.text for q in root.findall('drop_query')]
+drop_queries = substitute_parameters(drop_query_templates, parameter_combinations)
+for c in connections:
+    for q in drop_queries:
+        try:
+            c.execute(q)
+        except:
+            print("Error:", sys.exc_info()[0], file=sys.stderr)
+
+# Run create queries
+create_query_templates = [q.text for q in root.findall('create_query')]
+create_queries = substitute_parameters(create_query_templates, parameter_combinations)
+for c in connections:
+    for q in create_queries:
+        c.execute(q)
+
+# Run fill queries
+fill_query_templates = [q.text for q in root.findall('fill_query')]
+fill_queries = substitute_parameters(fill_query_templates, parameter_combinations)
+for c in connections:
+    for q in fill_queries:
+        c.execute(q)
+
+# Run test queries
+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:
+    for run in range(0, 7):
+        for conn_index, c in enumerate(connections):
+            res = c.execute(q)
+            print(q + '\t' + str(run) + '\t' + str(conn_index) + '\t' + str(c.last_query.elapsed))
+
+# Run drop queries
+drop_query_templates = [q.text for q in root.findall('drop_query')]
+drop_queries = substitute_parameters(drop_query_templates, parameter_combinations)
+for c in connections:
+    for q in drop_queries:
+        c.execute(q)