diff --git a/MAINTAINERS b/MAINTAINERS
index 75e0f2d750..a77f246569 100644
--- a/MAINTAINERS
+++ b/MAINTAINERS
@@ -2525,6 +2525,7 @@ Benchmark util
 M: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
 S: Maintained
 F: scripts/simplebench/
+T: git https://src.openvz.org/scm/~vsementsov/qemu.git simplebench
 
 Transactions helper
 M: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
diff --git a/scripts/simplebench/bench-backup.py b/scripts/simplebench/bench-backup.py
index 33a1ecfefa..5a0675c593 100755
--- a/scripts/simplebench/bench-backup.py
+++ b/scripts/simplebench/bench-backup.py
@@ -23,7 +23,7 @@ import json
 
 import simplebench
 from results_to_text import results_to_text
-from bench_block_job import bench_block_copy, drv_file, drv_nbd
+from bench_block_job import bench_block_copy, drv_file, drv_nbd, drv_qcow2
 
 
 def bench_func(env, case):
@@ -37,29 +37,56 @@ def bench_func(env, case):
 def bench(args):
     test_cases = []
 
-    sources = {}
-    targets = {}
-    for d in args.dir:
-        label, path = d.split(':')  # paths with colon not supported
-        sources[label] = drv_file(path + '/test-source')
-        targets[label] = drv_file(path + '/test-target')
+    # paths with colon not supported, so we just split by ':'
+    dirs = dict(d.split(':') for d in args.dir)
 
+    nbd_drv = None
     if args.nbd:
         nbd = args.nbd.split(':')
         host = nbd[0]
         port = '10809' if len(nbd) == 1 else nbd[1]
-        drv = drv_nbd(host, port)
-        sources['nbd'] = drv
-        targets['nbd'] = drv
+        nbd_drv = drv_nbd(host, port)
 
     for t in args.test:
         src, dst = t.split(':')
 
-        test_cases.append({
-            'id': t,
-            'source': sources[src],
-            'target': targets[dst]
-        })
+        if src == 'nbd' and dst == 'nbd':
+            raise ValueError("Can't use 'nbd' label for both src and dst")
+
+        if (src == 'nbd' or dst == 'nbd') and not nbd_drv:
+            raise ValueError("'nbd' label used but --nbd is not given")
+
+        if src == 'nbd':
+            source = nbd_drv
+        elif args.qcow2_sources:
+            source = drv_qcow2(drv_file(dirs[src] + '/test-source.qcow2'))
+        else:
+            source = drv_file(dirs[src] + '/test-source')
+
+        if dst == 'nbd':
+            test_cases.append({'id': t, 'source': source, 'target': nbd_drv})
+            continue
+
+        if args.target_cache == 'both':
+            target_caches = ['direct', 'cached']
+        else:
+            target_caches = [args.target_cache]
+
+        for c in target_caches:
+            o_direct = c == 'direct'
+            fname = dirs[dst] + '/test-target'
+            if args.compressed:
+                fname += '.qcow2'
+            target = drv_file(fname, o_direct=o_direct)
+            if args.compressed:
+                target = drv_qcow2(target)
+
+            test_id = t
+            if args.target_cache == 'both':
+                test_id += f'({c})'
+
+            test_cases.append({'id': test_id, 'source': source,
+                               'target': target})
 
     binaries = []  # list of (<label>, <path>, [<options>])
     for i, q in enumerate(args.env):
@@ -106,6 +133,13 @@ def bench(args):
             elif opt.startswith('max-workers='):
                 x_perf['max-workers'] = int(opt.split('=')[1])
 
+        backup_options = {}
+        if x_perf:
+            backup_options['x-perf'] = x_perf
+
+        if args.compressed:
+            backup_options['compress'] = True
+
         if is_mirror:
             assert not x_perf
             test_envs.append({
@@ -117,11 +151,13 @@ def bench(args):
             test_envs.append({
                 'id': f'backup({label})\n' + '\n'.join(opts),
                 'cmd': 'blockdev-backup',
-                'cmd-options': {'x-perf': x_perf} if x_perf else {},
+                'cmd-options': backup_options,
                 'qemu-binary': path
             })
 
-    result = simplebench.bench(bench_func, test_envs, test_cases, count=3)
+    result = simplebench.bench(bench_func, test_envs, test_cases,
+                               count=args.count, initial_run=args.initial_run,
+                               drop_caches=args.drop_caches)
     with open('results.json', 'w') as f:
         json.dump(result, f, indent=4)
     print(results_to_text(result))
@@ -163,5 +199,30 @@ default port 10809). Use it in tests, label is "nbd"
     p.add_argument('--test', nargs='+', help='''\
 Tests, in form source-dir-label:target-dir-label''',
                    action=ExtendAction)
+    p.add_argument('--compressed', help='''\
+Use compressed backup. It automatically means
+automatically creating qcow2 target with
+lazy_refcounts for each test run''', action='store_true')
+    p.add_argument('--qcow2-sources', help='''\
+Use test-source.qcow2 images as sources instead of
+test-source raw images''', action='store_true')
+    p.add_argument('--target-cache', help='''\
+Setup cache for target nodes. Options:
+   direct: default, use O_DIRECT and aio=native
+   cached: use system cache (Qemu default) and aio=threads (Qemu default)
+   both: generate two test cases for each src:dst pair''',
+                   default='direct', choices=('direct', 'cached', 'both'))
+
+    p.add_argument('--count', type=int, default=3, help='''\
+Number of test runs per table cell''')
+
+    # BooleanOptionalAction helps to support --no-initial-run option
+    p.add_argument('--initial-run', action=argparse.BooleanOptionalAction,
+                   help='''\
+Do additional initial run per cell which doesn't count in result,
+default true''')
+
+    p.add_argument('--drop-caches', action='store_true', help='''\
+Do "sync; echo 3 > /proc/sys/vm/drop_caches" before each test run''')
 
     bench(p.parse_args())
diff --git a/scripts/simplebench/bench_block_job.py b/scripts/simplebench/bench_block_job.py
index 7332845c1c..4f03c12169 100755
--- a/scripts/simplebench/bench_block_job.py
+++ b/scripts/simplebench/bench_block_job.py
@@ -21,6 +21,7 @@
 
 import sys
 import os
+import subprocess
 import socket
 import json
 
@@ -69,6 +70,10 @@ def bench_block_job(cmd, cmd_args, qemu_args):
             vm.shutdown()
             return {'error': 'block-job failed: ' + str(e),
                     'vm-log': vm.get_log()}
+        if 'error' in e['data']:
+            vm.shutdown()
+            return {'error': 'block-job failed: ' + e['data']['error'],
+                    'vm-log': vm.get_log()}
         end_ms = e['timestamp']['seconds'] * 1000000 + \
             e['timestamp']['microseconds']
     finally:
@@ -77,11 +82,34 @@ def bench_block_job(cmd, cmd_args, qemu_args):
     return {'seconds': (end_ms - start_ms) / 1000000.0}
 
 
+def get_image_size(path):
+    out = subprocess.run(['qemu-img', 'info', '--out=json', path],
+                         stdout=subprocess.PIPE, check=True).stdout
+    return json.loads(out)['virtual-size']
+
+
+def get_blockdev_size(obj):
+    img = obj['filename'] if 'filename' in obj else obj['file']['filename']
+    return get_image_size(img)
+
+
 # Bench backup or mirror
 def bench_block_copy(qemu_binary, cmd, cmd_options, source, target):
     """Helper to run bench_block_job() for mirror or backup"""
     assert cmd in ('blockdev-backup', 'blockdev-mirror')
 
+    if target['driver'] == 'qcow2':
+        try:
+            os.remove(target['file']['filename'])
+        except OSError:
+            pass
+
+        subprocess.run(['qemu-img', 'create', '-f', 'qcow2',
+                        target['file']['filename'],
+                        str(get_blockdev_size(source))],
+                       stdout=subprocess.DEVNULL,
+                       stderr=subprocess.DEVNULL, check=True)
+
     source['node-name'] = 'source'
     target['node-name'] = 'target'
 
@@ -96,9 +124,13 @@ def bench_block_copy(qemu_binary, cmd, cmd_options, source, target):
                             '-blockdev', json.dumps(target)])
 
 
-def drv_file(filename):
-    return {'driver': 'file', 'filename': filename,
-            'cache': {'direct': True}, 'aio': 'native'}
+def drv_file(filename, o_direct=True):
+    node = {'driver': 'file', 'filename': filename}
+    if o_direct:
+        node['cache'] = {'direct': True}
+        node['aio'] = 'native'
+
+    return node
 
 
 def drv_nbd(host, port):
@@ -106,6 +138,10 @@ def drv_nbd(host, port):
             'server': {'type': 'inet', 'host': host, 'port': port}}
 
 
+def drv_qcow2(file):
+    return {'driver': 'qcow2', 'file': file}
+
+
 if __name__ == '__main__':
     import sys
 
diff --git a/scripts/simplebench/simplebench.py b/scripts/simplebench/simplebench.py
index f61513af90..8efca2af98 100644
--- a/scripts/simplebench/simplebench.py
+++ b/scripts/simplebench/simplebench.py
@@ -19,9 +19,17 @@
 #
 
 import statistics
+import subprocess
+import time
 
 
-def bench_one(test_func, test_env, test_case, count=5, initial_run=True):
+def do_drop_caches():
+    subprocess.run('sync; echo 3 > /proc/sys/vm/drop_caches', shell=True,
+                   check=True)
+
+
+def bench_one(test_func, test_env, test_case, count=5, initial_run=True,
+              slow_limit=100, drop_caches=False):
     """Benchmark one test-case
 
     test_func   -- benchmarking function with prototype
@@ -36,6 +44,9 @@ 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  -- stop at slow run (that exceedes the slow_limit by seconds).
+                   (initial run is not measured)
+    drop_caches -- drop caches before each run
 
     Returns dict with the following fields:
         'runs':     list of test_func results
@@ -49,15 +60,25 @@ def bench_one(test_func, test_env, test_case, count=5, initial_run=True):
     """
     if initial_run:
         print('  #initial run:')
+        do_drop_caches()
         print('   ', test_func(test_env, test_case))
 
     runs = []
     for i in range(count):
+        t = time.time()
+
         print('  #run {}'.format(i+1))
+        do_drop_caches()
         res = test_func(test_env, test_case)
         print('   ', res)
         runs.append(res)
 
+        if time.time() - t > slow_limit:
+            print('    - run is too slow, stop here')
+            break
+
+    count = len(runs)
+
     result = {'runs': runs}
 
     succeeded = [r for r in runs if ('seconds' in r or 'iops' in r)]
@@ -71,7 +92,10 @@ def bench_one(test_func, test_env, test_case, count=5, initial_run=True):
             dim = 'seconds'
         result['dimension'] = dim
         result['average'] = statistics.mean(r[dim] for r in succeeded)
-        result['stdev'] = statistics.stdev(r[dim] for r in succeeded)
+        if len(succeeded) == 1:
+            result['stdev'] = 0
+        else:
+            result['stdev'] = statistics.stdev(r[dim] for r in succeeded)
 
     if len(succeeded) < count:
         result['n-failed'] = count - len(succeeded)