fstests: add fio perf results support
authorJosef Bacik <jbacik@fb.com>
Tue, 7 Nov 2017 21:53:32 +0000 (16:53 -0500)
committerEryu Guan <eguan@redhat.com>
Sun, 26 Nov 2017 07:37:22 +0000 (15:37 +0800)
This patch does the nuts and bolts of grabbing fio results and
storing them in a database in order to check against for future
runs.  This works by storing the results in resuts/fio-results.db as
a sqlite database.  The src/perf directory has all the supporting
python code for parsing the fio json results, storing it in the
database, and loading previous results from the database to compare
with the current results.

This also adds a PERF_CONFIGNAME option that must be set for this to
work.  Since we all have various ways we run fstests it doesn't make
sense to compare different configurations with each other (unless
specifically desired).  The PERF_CONFIGNAME will allow us to
separate out results for different test run configurations to make
sure we're comparing results correctly.

Currently we only check against the last perf result.  In the future
I will flesh this out to compare against the average of N number of
runs to be a little more complete, and hopefully that will allow us
to also watch latencies as well.

[eguan: add required Makefile updates]

Signed-off-by: Josef Bacik <jbacik@fb.com>
Reviewed-by: Eryu Guan <eguan@redhat.com>
Signed-off-by: Eryu Guan <eguan@redhat.com>
13 files changed:
.gitignore
README
common/config
common/perf [new file with mode: 0644]
doc/requirement-checking.txt
src/Makefile
src/perf/FioCompare.py [new file with mode: 0644]
src/perf/FioResultDecoder.py [new file with mode: 0644]
src/perf/Makefile [new file with mode: 0644]
src/perf/ResultData.py [new file with mode: 0644]
src/perf/fio-insert-and-compare.py [new file with mode: 0644]
src/perf/fio-results.sql [new file with mode: 0644]
src/perf/generate-schema.py [new file with mode: 0644]

index 2014c08..f27c30a 100644 (file)
 /src/aio-dio-regress/aiocp
 /src/aio-dio-regress/aiodio_sparse2
 /src/log-writes/replay-log
+/src/perf/*.pyc
 
 # dmapi/ binaries
 /dmapi/src/common/cmd/read_invis
diff --git a/README b/README
index a9da4f0..ed69332 100644 (file)
--- a/README
+++ b/README
@@ -8,13 +8,13 @@ _______________________
        sudo apt-get install xfslibs-dev uuid-dev libtool-bin \
        e2fsprogs automake gcc libuuid1 quota attr libattr1-dev make \
        libacl1-dev libaio-dev xfsprogs libgdbm-dev gawk fio dbench \
-       uuid-runtime
+       uuid-runtime python sqlite3
   For Fedora, RHEL, or CentOS:
        yum install acl attr automake bc dbench dump e2fsprogs fio \
        gawk gcc indent libtool lvm2 make psmisc quota sed \
        xfsdump xfsprogs \
        libacl-devel libattr-devel libaio-devel libuuid-devel \
-       xfsprogs-devel btrfs-progs-devel
+       xfsprogs-devel btrfs-progs-devel python sqlite
        (Older distributions may require xfsprogs-qa-devel as well.)
        (Note that for RHEL and CentOS, you may need the EPEL repo.)
 - run make
@@ -93,6 +93,12 @@ Preparing system for tests:
                It is no longer necessary to set TEST_XFS_SCRUB.
              - setenv LOGWRITES_DEV to a block device to use for power fail
                testing.
+             - setenv PERF_CONFIGNAME to a arbitrary string to be used for
+               identifying the test setup for running perf tests.  This should
+               be different for each type of performance test you wish to run so
+               that relevant results are compared.  For example 'spinningrust'
+               for configurations that use spinning disks and 'nvme' for tests
+               using nvme drives.
 
         - or add a case to the switch in common/config assigning
           these variables based on the hostname of your test
index 4b3402e..d0fbfe5 100644 (file)
@@ -197,6 +197,8 @@ export NFS4_SETFACL_PROG="`set_prog_path nfs4_setfacl`"
 export NFS4_GETFACL_PROG="`set_prog_path nfs4_getfacl`"
 export UBIUPDATEVOL_PROG="`set_prog_path ubiupdatevol`"
 export THIN_CHECK_PROG="$(set_prog_path thin_check)"
+export PYTHON2_PROG="`set_prog_path python2`"
+export SQLITE3_PROG="`set_prog_path sqlite3`"
 
 # use 'udevadm settle' or 'udevsettle' to wait for lv to be settled.
 # newer systems have udevadm command but older systems like RHEL5 don't.
diff --git a/common/perf b/common/perf
new file mode 100644 (file)
index 0000000..8b4c9be
--- /dev/null
@@ -0,0 +1,40 @@
+#
+# Common perf specific functions
+#
+
+
+_require_fio_results()
+{
+       if [ -z "$PERF_CONFIGNAME" ]
+       then
+               _notrun "this test requires \$PERF_CONFIGNAME to be set"
+       fi
+       _require_command $PYTHON2_PROG python2
+
+       $PYTHON2_PROG -c "import sqlite3" >/dev/null 2>&1
+       [ $? -ne 0 ] && _notrun "this test requires python sqlite support"
+
+       $PYTHON2_PROG -c "import json" >/dev/null 2>&1
+       [ $? -ne 0 ] && _notrun "this test requires python json support"
+
+       _require_command $SQLITE3_PROG sqlite3
+}
+
+_fio_results_init()
+{
+       cat $here/src/perf/fio-results.sql | \
+               $SQLITE3_PROG $RESULT_BASE/fio-results.db
+       [ $? -ne 0 ] && _fail "failed to create results database"
+       [ ! -e $RESULT_BASE/fio-results.db ] && \
+               _fail "failed to create results database"
+}
+
+_fio_results_compare()
+{
+       _testname=$1
+       _resultfile=$2
+
+       $PYTHON2_PROG $here/src/perf/fio-insert-and-compare.py \
+               -c $PERF_CONFIGNAME -d $RESULT_BASE/fio-results.db \
+               -n $_testname $_resultfile
+}
index 4e01b1f..1ec04d4 100644 (file)
@@ -122,3 +122,12 @@ _require_log_writes
      The test requires the use of the device mapper target log-writes.
      The test also requires the test program log-writes/replay-log is built
      and will be skipped if either isn't available.
+
+======================
+PERF TEST REQUIREMENTS
+======================
+
+_require_fio_results
+
+     This test requires the supporting tools for saving and comparing fio based
+     perf test results.
index 3eb25b1..b101217 100644 (file)
@@ -25,7 +25,7 @@ LINUX_TARGETS = xfsctl bstat t_mtab getdevicesize preallo_rw_pattern_reader \
        attr-list-by-handle-cursor-test listxattr dio-interleaved t_dir_type \
        dio-invalidate-cache stat_test t_encrypted_d_revalidate
 
-SUBDIRS = log-writes
+SUBDIRS = log-writes perf
 
 LLDLIBS = $(LIBATTR) $(LIBHANDLE) $(LIBACL) -lpthread
 
diff --git a/src/perf/FioCompare.py b/src/perf/FioCompare.py
new file mode 100644 (file)
index 0000000..6b56bd4
--- /dev/null
@@ -0,0 +1,113 @@
+default_keys = [ 'iops', 'io_bytes', 'bw' ]
+latency_keys = [ 'lat_ns_min', 'lat_ns_max' ]
+main_job_keys = [ 'sys_cpu', 'elapsed' ]
+io_ops = ['read', 'write', 'trim' ]
+
+def _fuzzy_compare(a, b, fuzzy):
+    if a == b:
+        return 0
+    if a == 0:
+        return 100
+    a = float(a)
+    b = float(b)
+    fuzzy = float(fuzzy)
+    val = ((b - a) / a) * 100
+    if val > fuzzy or val < -fuzzy:
+        return val;
+    return 0
+
+def _compare_jobs(ijob, njob, latency, fuzz, failures_only):
+    failed = 0
+    for k in default_keys:
+        for io in io_ops:
+            key = "{}_{}".format(io, k)
+            comp = _fuzzy_compare(ijob[key], njob[key], fuzz)
+            if comp < 0:
+                print("    {} regressed: old {} new {} {}%".format(key,
+                      ijob[key], njob[key], comp))
+                failed += 1
+            elif not failures_only and comp > 0:
+                print("    {} improved: old {} new {} {}%".format(key,
+                      ijob[key], njob[key], comp))
+            elif not failures_only:
+                print("{} is a-ok {} {}".format(key, ijob[key], njob[key]))
+    for k in latency_keys:
+        if not latency:
+            break
+        for io in io_ops:
+            key = "{}_{}".format(io, k)
+            comp = _fuzzy_compare(ijob[key], njob[key], fuzz)
+            if comp > 0:
+                print("    {} regressed: old {} new {} {}%".format(key,
+                      ijob[key], njob[key], comp))
+                failed += 1
+            elif not failures_only and comp < 0:
+                print("    {} improved: old {} new {} {}%".format(key,
+                      ijob[key], njob[key], comp))
+            elif not failures_only:
+                print("{} is a-ok {} {}".format(key, ijob[key], njob[key]))
+    for k in main_job_keys:
+        comp = _fuzzy_compare(ijob[k], njob[k], fuzz)
+        if comp > 0:
+            print("    {} regressed: old {} new {} {}%".format(k, ijob[k],
+                  njob[k], comp))
+            failed += 1
+        elif not failures_only and comp < 0:
+            print("    {} improved: old {} new {} {}%".format(k, ijob[k],
+                  njob[k], comp))
+        elif not failures_only:
+                print("{} is a-ok {} {}".format(k, ijob[k], njob[k]))
+    return failed
+
+def compare_individual_jobs(initial, data, fuzz, failures_only):
+    failed = 0;
+    initial_jobs = initial['jobs'][:]
+    for njob in data['jobs']:
+        for ijob in initial_jobs:
+            if njob['jobname'] == ijob['jobname']:
+                print("  Checking results for {}".format(njob['jobname']))
+                failed += _compare_jobs(ijob, njob, fuzz, failures_only)
+                initial_jobs.remove(ijob)
+                break
+    return failed
+
+def default_merge(data):
+    '''Default merge function for multiple jobs in one run
+
+    For runs that include multiple threads we will have a lot of variation
+    between the different threads, which makes comparing them to eachother
+    across multiple runs less that useful.  Instead merge the jobs into a single
+    job.  This function does that by adding up 'iops', 'io_kbytes', and 'bw' for
+    read/write/trim in the merged job, and then taking the maximal values of the
+    latency numbers.
+    '''
+    merge_job = {}
+    for job in data['jobs']:
+        for k in main_job_keys:
+            if k not in merge_job:
+                merge_job[k] = job[k]
+            else:
+                merge_job[k] += job[k]
+        for io in io_ops:
+            for k in default_keys:
+                key = "{}_{}".format(io, k)
+                if key not in merge_job:
+                    merge_job[key] = job[key]
+                else:
+                    merge_job[key] += job[key]
+            for k in latency_keys:
+                key = "{}_{}".format(io, k)
+                if key not in merge_job:
+                    merge_job[key] = job[key]
+                elif merge_job[key] < job[key]:
+                    merge_job[key] = job[key]
+    return merge_job
+
+def compare_fiodata(initial, data, latency, merge_func=default_merge, fuzz=5,
+                    failures_only=True):
+    failed  = 0
+    if merge_func is None:
+        return compare_individual_jobs(initial, data, fuzz, failures_only)
+    ijob = merge_func(initial)
+    njob = merge_func(data)
+    return _compare_jobs(ijob, njob, latency, fuzz, failures_only)
diff --git a/src/perf/FioResultDecoder.py b/src/perf/FioResultDecoder.py
new file mode 100644 (file)
index 0000000..e15406f
--- /dev/null
@@ -0,0 +1,62 @@
+import json
+
+class FioResultDecoder(json.JSONDecoder):
+    """Decoder for decoding fio result json to an object for our database
+
+    This decodes the json output from fio into an object that can be directly
+    inserted into our database.  This just strips out the fields we don't care
+    about and collapses the read/write/trim classes into a flat value structure
+    inside of the jobs object.
+
+    For example
+        "write" : {
+            "io_bytes" : 313360384,
+            "bw" : 1016,
+        }
+
+    Get's collapsed to
+
+        "write_io_bytes" : 313360384,
+        "write_bw": 1016,
+
+    Currently any dict under 'jobs' get's dropped, with the exception of 'read',
+    'write', and 'trim'.  For those sub sections we drop any dict's under those.
+
+    Attempt to keep this as generic as possible, we don't want to break every
+    time fio changes it's json output format.
+    """
+    _ignore_types = ['dict', 'list']
+    _override_keys = ['lat_ns', 'lat']
+    _io_ops = ['read', 'write', 'trim']
+
+    _transform_keys = { 'lat': 'lat_ns' }
+
+    def decode(self, json_string):
+        """This does the dirty work of converting everything"""
+        default_obj = super(FioResultDecoder, self).decode(json_string)
+        obj = {}
+        obj['global'] = {}
+        obj['global']['time'] = default_obj['time']
+        obj['jobs'] = []
+        for job in default_obj['jobs']:
+            new_job = {}
+            for key,value in job.iteritems():
+                if key not in self._io_ops:
+                    if value.__class__.__name__ in self._ignore_types:
+                        continue
+                    new_job[key] = value
+                    continue
+                for k,v in value.iteritems():
+                    if k in self._override_keys:
+                        if k in self._transform_keys:
+                            k = self._transform_keys[k]
+                        for subk,subv in v.iteritems():
+                            collapsed_key = "{}_{}_{}".format(key, k, subk)
+                            new_job[collapsed_key] = subv
+                        continue
+                    if v.__class__.__name__ in self._ignore_types:
+                        continue
+                    collapsed_key = "{}_{}".format(key, k)
+                    new_job[collapsed_key] = v
+            obj['jobs'].append(new_job)
+        return obj
diff --git a/src/perf/Makefile b/src/perf/Makefile
new file mode 100644 (file)
index 0000000..819e767
--- /dev/null
@@ -0,0 +1,14 @@
+TOPDIR = ../..
+include $(TOPDIR)/include/builddefs
+
+TARGETS = $(wildcard *.py) $(wildcard *.sql)
+
+depend: .dep
+
+include $(BUILDRULES)
+
+install:
+       $(INSTALL) -m 755 -d $(PKG_LIB_DIR)/src/perf
+       $(INSTALL) -m 755 $(TARGETS) $(PKG_LIB_DIR)/src/perf
+
+-include .dep
diff --git a/src/perf/ResultData.py b/src/perf/ResultData.py
new file mode 100644 (file)
index 0000000..f0c7eac
--- /dev/null
@@ -0,0 +1,43 @@
+import sqlite3
+
+def _dict_factory(cursor, row):
+    d = {}
+    for idx,col in enumerate(cursor.description):
+        d[col[0]] = row[idx]
+    return d
+
+class ResultData:
+    def __init__(self, filename):
+        self.db = sqlite3.connect(filename)
+        self.db.row_factory = _dict_factory
+
+    def load_last(self, testname, config):
+        d = {}
+        cur = self.db.cursor()
+        cur.execute("SELECT * FROM fio_runs WHERE config = ? AND name = ?ORDER BY time DESC LIMIT 1",
+                    (config,testname))
+        d['global'] = cur.fetchone()
+        if d['global'] is None:
+            return None
+        cur.execute("SELECT * FROM fio_jobs WHERE run_id = ?",
+                    (d['global']['id'],))
+        d['jobs'] = cur.fetchall()
+        return d
+
+    def _insert_obj(self, tablename, obj):
+        keys = obj.keys()
+        values = obj.values()
+        cur = self.db.cursor()
+        cmd = "INSERT INTO {} ({}) VALUES ({}".format(tablename,
+                                                       ",".join(keys),
+                                                       '?,' * len(values))
+        cmd = cmd[:-1] + ')'
+        cur.execute(cmd, tuple(values))
+        self.db.commit()
+        return cur.lastrowid
+
+    def insert_result(self, result):
+        row_id = self._insert_obj('fio_runs', result['global'])
+        for job in result['jobs']:
+            job['run_id'] = row_id
+            self._insert_obj('fio_jobs', job)
diff --git a/src/perf/fio-insert-and-compare.py b/src/perf/fio-insert-and-compare.py
new file mode 100644 (file)
index 0000000..064af6d
--- /dev/null
@@ -0,0 +1,35 @@
+import FioResultDecoder
+import ResultData
+import FioCompare
+import json
+import argparse
+import sys
+import platform
+
+parser = argparse.ArgumentParser()
+parser.add_argument('-c', '--configname', type=str,
+                    help="The config name to save the results under.",
+                    required=True)
+parser.add_argument('-d', '--db', type=str,
+                    help="The db that is being used", required=True)
+parser.add_argument('-n', '--testname', type=str,
+                    help="The testname for the result", required=True)
+parser.add_argument('result', type=str,
+                    help="The result file to compare and insert")
+args = parser.parse_args()
+
+result_data = ResultData.ResultData(args.db)
+compare = result_data.load_last(args.testname, args.configname)
+
+json_data = open(args.result)
+data = json.load(json_data, cls=FioResultDecoder.FioResultDecoder)
+data['global']['name'] = args.testname
+data['global']['config'] = args.configname
+data['global']['kernel'] = platform.release()
+result_data.insert_result(data)
+
+if compare is None:
+    sys.exit(0)
+
+if FioCompare.compare_fiodata(compare, data, False):
+    sys.exit(1)
diff --git a/src/perf/fio-results.sql b/src/perf/fio-results.sql
new file mode 100644 (file)
index 0000000..62e1464
--- /dev/null
@@ -0,0 +1,94 @@
+CREATE TABLE IF NOT EXISTS `fio_runs` (
+  `id` INTEGER PRIMARY KEY AUTOINCREMENT,
+  `kernel` datetime NOT NULL,
+  `config` varchar(256) NOT NULL,
+  `name` varchar(256) NOT NULL,
+  `time` datetime NOT NULL
+);
+CREATE TABLE IF NOT EXISTS `fio_jobs` (
+  `id` INTEGER PRIMARY KEY AUTOINCREMENT,
+  `run_id` int NOT NULL,
+  `read_bw_dev` float,
+  `trim_lat_ns_mean` float,
+  `read_runtime` int,
+  `trim_runtime` int,
+  `read_io_bytes` int,
+  `read_short_ios` int,
+  `write_lat_ns_stddev` float,
+  `minf` int,
+  `read_drop_ios` int,
+  `trim_iops_samples` int,
+  `trim_iops_max` int,
+  `trim_bw_agg` float,
+  `write_bw_min` int,
+  `latency_percentile` float,
+  `read_bw_max` int,
+  `write_bw` int,
+  `read_bw_min` int,
+  `trim_bw_dev` float,
+  `read_iops_max` int,
+  `read_lat_ns_mean` float,
+  `write_iops` float,
+  `latency_target` int,
+  `trim_bw` int,
+  `write_iops_samples` int,
+  `read_bw_samples` int,
+  `trim_io_kbytes` int,
+  `read_iops_samples` int,
+  `write_drop_ios` int,
+  `trim_iops_min` int,
+  `write_bw_samples` int,
+  `read_iops_stddev` float,
+  `write_io_kbytes` int,
+  `groupid` int,
+  `trim_bw_mean` float,
+  `write_bw_agg` float,
+  `write_bw_dev` float,
+  `read_bw` int,
+  `trim_lat_ns_stddev` float,
+  `read_bw_mean` float,
+  `latency_depth` int,
+  `trim_short_ios` int,
+  `read_lat_ns_stddev` float,
+  `read_io_kbytes` int,
+  `latency_window` int,
+  `write_iops_stddev` float,
+  `trim_bw_samples` int,
+  `trim_lat_ns_min` int,
+  `error` int,
+  `trim_iops_mean` float,
+  `elapsed` int,
+  `write_iops_mean` float,
+  `write_bw_mean` float,
+  `write_short_ios` int,
+  `write_io_bytes` int,
+  `usr_cpu` float,
+  `trim_drop_ios` int,
+  `read_iops_min` int,
+  `jobname` varchar(256),
+  `write_iops_min` int,
+  `trim_bw_min` int,
+  `read_bw_agg` float,
+  `trim_lat_ns_max` int,
+  `write_lat_ns_min` int,
+  `read_iops_mean` float,
+  `trim_iops_stddev` float,
+  `write_lat_ns_max` int,
+  `majf` int,
+  `write_total_ios` int,
+  `ctx` int,
+  `read_lat_ns_min` int,
+  `trim_bw_max` int,
+  `read_total_ios` int,
+  `write_runtime` int,
+  `trim_io_bytes` int,
+  `eta` int,
+  `read_iops` float,
+  `trim_total_ios` int,
+  `write_lat_ns_mean` float,
+  `write_iops_max` int,
+  `write_bw_max` int,
+  `sys_cpu` float,
+  `read_lat_ns_max` int,
+  `trim_iops` float
+);
diff --git a/src/perf/generate-schema.py b/src/perf/generate-schema.py
new file mode 100644 (file)
index 0000000..b61504b
--- /dev/null
@@ -0,0 +1,55 @@
+import json
+import argparse
+import FioResultDecoder
+from dateutil.parser import parse
+
+def is_date(string):
+    try:
+        parse(string)
+        return True
+    except ValueError:
+        return False
+
+def print_schema_def(key, value, required):
+    typestr = value.__class__.__name__
+    if typestr == 'str' or typestr == 'unicode':
+        if (is_date(value)):
+            typestr = "datetime"
+        else:
+            typestr = "varchar(256)"
+    requiredstr = ""
+    if required:
+        requiredstr = " NOT NULL"
+    return ",\n  `{}` {}{}".format(key, typestr, requiredstr)
+
+parser = argparse.ArgumentParser()
+parser.add_argument('infile', help="The json file to strip")
+args = parser.parse_args()
+
+json_data = open(args.infile)
+data = json.load(json_data, cls=FioResultDecoder.FioResultDecoder)
+
+# These get populated by the test runner, not fio, so add them so their
+# definitions get populated in the schema properly
+data['global']['config'] = 'default'
+data['global']['kernel'] = '4.14'
+data['global']['name'] = 'alrightalrightalright'
+
+print("CREATE TABLE IF NOT EXISTS `fio_runs` (")
+outstr = "  `id` INTEGER PRIMARY KEY AUTOINCREMENT"
+for key,value in data['global'].iteritems():
+    outstr += print_schema_def(key, value, True)
+print(outstr)
+print(");")
+
+required_fields = ['run_id']
+
+job = data['jobs'][0]
+job['run_id'] = 0
+
+print("CREATE TABLE IF NOT EXISTS `fio_jobs` (")
+outstr = "  `id` INTEGER PRIMARY KEY AUTOINCREMENT"
+for key,value in job.iteritems():
+    outstr += print_schema_def(key, value, key in required_fields)
+print(outstr)
+print(");")