# if you do this, don't pass normal CLI parameters to it
# otherwise it runs the CLI
-import sys, os, math, copy
+import sys, os, math, copy, time
from copy import deepcopy
-
+import argparse
import unittest2
msec_per_sec = 1000
nsec_per_usec = 1000
+direction_read = 0
+direction_write = 1
class FioHistoLogExc(Exception):
pass
-# if there is an error, print message, print syntax, and exit with error status
+# if there is an error, print message, and exit with error status
-def usage(msg):
+def myabort(msg):
print('ERROR: ' + msg)
- print('usage: fio-histo-log-pctiles.py ')
- print(' [ --fio-version 2|3 (default 3) ]')
- print(' [ --bucket-groups positive-int (default 29) ]')
- print(' [ --bucket-bits small-positive-int (default 6) ]')
- print(' [ --percentiles p1,p2,...,pN ] (default 0,50,95,99,100)')
- print(' [ --time-quantum positive-int (default 1 sec) ]')
- print(' [ --output-unit msec|usec|nsec (default msec) ]')
- print(' log-file1 log-file2 ...')
sys.exit(1)
-
# convert histogram log file into a list of
# (time_ms, direction, bsz, buckets) tuples where
# - time_ms is the time in msec at which the log record was written
# log file parser raises FioHistoLogExc exceptions
# it returns histogram buckets in whatever unit fio uses
-
-def parse_hist_file(logfn, buckets_per_interval):
- max_timestamp_ms = 0.0
-
+# inputs:
+# logfn: pathname to histogram log file
+# buckets_per_interval - how many histogram buckets to expect
+# log_hist_msec - if not None, expected time interval between histogram records
+
+def parse_hist_file(logfn, buckets_per_interval, log_hist_msec):
+ previous_ts_ms_read = -1
+ previous_ts_ms_write = -1
+
with open(logfn, 'r') as f:
records = [ l.strip() for l in f.readlines() ]
intervals = []
if len(int_tokens) < 3:
raise FioHistoLogExc('too few numbers %s' % exception_suffix(k+1, logfn))
- time_ms = int_tokens[0]
- if time_ms > max_timestamp_ms:
- max_timestamp_ms = time_ms
-
direction = int_tokens[1]
- if direction != 0 and direction != 1:
+ if direction != direction_read and direction != direction_write:
raise FioHistoLogExc('invalid I/O direction %s' % exception_suffix(k+1, logfn))
+ time_ms = int_tokens[0]
+ if direction == direction_read:
+ if time_ms < previous_ts_ms_read:
+ raise FioHistoLogExc('read timestamp in column 1 decreased %s' % exception_suffix(k+1, logfn))
+ previous_ts_ms_read = time_ms
+ elif direction == direction_write:
+ if time_ms < previous_ts_ms_write:
+ raise FioHistoLogExc('write timestamp in column 1 decreased %s' % exception_suffix(k+1, logfn))
+ previous_ts_ms_write = time_ms
+
bsz = int_tokens[2]
if bsz > (1 << 24):
raise FioHistoLogExc('block size too large %s' % exception_suffix(k+1, logfn))
intervals.append((time_ms, direction, bsz, buckets))
if len(intervals) == 0:
raise FioHistoLogExc('no records in %s' % logfn)
- return (intervals, max_timestamp_ms)
+ (first_timestamp, _, _, _) = intervals[0]
+ if first_timestamp < 1000000:
+ start_time = 0 # assume log_unix_epoch = 0
+ elif log_hist_msec != None:
+ start_time = first_timestamp - log_hist_msec
+ elif len(intervals) > 1:
+ (second_timestamp, _, _, _) = intervals[1]
+ start_time = first_timestamp - (second_timestamp - first_timestamp)
+ (end_timestamp, _, _, _) = intervals[-1]
+
+ return (intervals, start_time, end_timestamp)
# compute time range for each bucket index in histogram record
# compute number of time quantum intervals in the test
-def get_time_intervals(time_quantum, max_timestamp_ms):
+def get_time_intervals(time_quantum, min_timestamp_ms, max_timestamp_ms):
# round down to nearest second
max_timestamp = max_timestamp_ms // msec_per_sec
+ min_timestamp = min_timestamp_ms // msec_per_sec
# round up to nearest whole multiple of time_quantum
- time_interval_count = (max_timestamp + time_quantum) // time_quantum
- end_time = time_interval_count * time_quantum
+ time_interval_count = ((max_timestamp - min_timestamp) + time_quantum) // time_quantum
+ end_time = min_timestamp + (time_interval_count * time_quantum)
return (end_time, time_interval_count)
# align raw histogram log data to time quantum so
# so the contribution of this bucket to this time quantum is
# 515 x 0.99 = 509.85
-def align_histo_log(raw_histogram_log, time_quantum, bucket_count, max_timestamp_ms):
+def align_histo_log(raw_histogram_log, time_quantum, bucket_count, min_timestamp_ms, max_timestamp_ms):
# slice up test time int intervals of time_quantum seconds
- (end_time, time_interval_count) = get_time_intervals(time_quantum, max_timestamp_ms)
+ (end_time, time_interval_count) = get_time_intervals(time_quantum, min_timestamp_ms, max_timestamp_ms)
time_qtm_ms = time_quantum * msec_per_sec
end_time_ms = end_time * msec_per_sec
aligned_intervals = []
for j in range(0, time_interval_count):
aligned_intervals.append((
- j * time_qtm_ms,
+ min_timestamp_ms + (j * time_qtm_ms),
[ 0.0 for j in range(0, bucket_count) ] ))
log_record_count = len(raw_histogram_log)
# calculate first quantum that overlaps this histogram record
- qtm_start_ms = (time_msec // time_qtm_ms) * time_qtm_ms
- qtm_end_ms = ((time_msec + time_qtm_ms) // time_qtm_ms) * time_qtm_ms
- qtm_index = qtm_start_ms // time_qtm_ms
+ offset_from_min_ts = time_msec - min_timestamp_ms
+ qtm_start_ms = min_timestamp_ms + (offset_from_min_ts // time_qtm_ms) * time_qtm_ms
+ qtm_end_ms = min_timestamp_ms + ((offset_from_min_ts + time_qtm_ms) // time_qtm_ms) * time_qtm_ms
+ qtm_index = offset_from_min_ts // time_qtm_ms
# for each quantum that overlaps this histogram record's time interval
while qtm_start_ms < time_msec_end: # while quantum overlaps record
+ # some histogram logs may be longer than others
+
+ if len(aligned_intervals) <= qtm_index:
+ break
+
# calculate fraction of time that this quantum
# overlaps histogram record's time interval
# this prevents floating-point error from preventing loop exit
almost_100 = 99.9999
+ # pct is the percentile corresponding to
+ # all I/O requests up through bucket b
+ pct = 0.0
total_so_far = 0
for b, io_count in enumerate(buckets):
+ if io_count == 0:
+ continue
total_so_far += io_count
- pct_lt = 100.0 * float(total_so_far) / total_ios
+ # last_pct_lt is the percentile corresponding to
+ # all I/O requests up to, but not including, bucket b
+ last_pct = pct
+ pct = 100.0 * float(total_so_far) / total_ios
# a single bucket could satisfy multiple pctiles
# so this must be a while loop
- # consider both the 0-percentile (min latency)
- # and 100-percentile (max latency) case here
- while ((next_pctile == 100.0 and pct_lt >= almost_100) or
- (next_pctile < 100.0 and pct_lt > next_pctile)):
- # FIXME: interpolate between these fractions
+ # for 100-percentile (max latency) case, no bucket exceeds it
+ # so we must stop there.
+ while ((next_pctile == 100.0 and pct >= almost_100) or
+ (next_pctile < 100.0 and pct > next_pctile)):
+ # interpolate between min and max time for bucket time interval
+ # we keep the time_ranges access inside this loop,
+ # even though it could be above the loop,
+ # because in many cases we will not be even entering
+ # the loop so we optimize out these accesses
range_max_time = time_ranges[b][1]
- pctile_result[next_pctile] = range_max_time
+ range_min_time = time_ranges[b][0]
+ offset_frac = (next_pctile - last_pct)/(pct - last_pct)
+ interpolation = range_min_time + (offset_frac*(range_max_time - range_min_time))
+ pctile_result[next_pctile] = interpolation
pctile_index += 1
if pctile_index == pctile_count:
break
return pctile_result
-# parse parameters
-# returns a tuple of command line parameters
-# parameters have default values unless otherwise shown
-
-def parse_cli_params():
-
- # default values for input parameters
-
- fio_version = 3 # we are using fio 3.x now
- bucket_groups = None # defaulting comes later
- bucket_bits = 6 # default in fio 3.x
- pctiles_wanted = [ 0, 50, 90, 95, 99, 100 ]
- time_quantum = 1
- output_unit = 'usec'
-
- # parse command line parameters and display them
-
- argindex = 1
- argct = len(sys.argv)
- if argct < 2:
- usage('must supply at least one histogram log file')
- while argindex < argct:
- if argct < argindex + 2:
- break
- pname = sys.argv[argindex]
- pval = sys.argv[argindex+1]
- if not pname.startswith('--'):
- break
- argindex += 2
- pname = pname[2:]
-
- if pname == 'bucket-groups':
- bucket_groups = int(pval)
- elif pname == 'bucket-bits':
- bucket_bits = int(pval)
- elif pname == 'time-quantum':
- time_quantum = int(pval)
- elif pname == 'percentiles':
- pctiles_wanted = [ float(p) for p in pval.split(',') ]
- elif pname == 'output-unit':
- if pval == 'msec' or pval == 'usec':
- output_unit = pval
- else:
- usage('output-unit must be usec (microseconds) or msec (milliseconds)')
- elif pname == 'fio-version':
- if pval != '2' and pval != '3':
- usage('invalid fio version, must be 2 or 3')
- fio_version = int(pval)
- else:
- usage('invalid parameter name --%s' % pname)
-
- if not bucket_groups:
- # default changes based on fio version
- if fio_version == 2:
- bucket_groups = 19
- else:
- # default in fio 3.x
- bucket_groups = 29
-
- filename_list = sys.argv[argindex:]
- for f in filename_list:
- if not os.path.exists(f):
- usage('file %s does not exist' % f)
- return (bucket_groups, bucket_bits, fio_version, pctiles_wanted,
- filename_list, time_quantum, output_unit)
-
-
# this is really the main program
def compute_percentiles_from_logs():
- (bucket_groups, bucket_bits, fio_version, pctiles_wanted,
- file_list, time_quantum, output_unit) = parse_cli_params()
-
- print('bucket groups = %d' % bucket_groups)
- print('bucket bits = %d' % bucket_bits)
- print('time quantum = %d sec' % time_quantum)
- print('percentiles = %s' % ','.join([ str(p) for p in pctiles_wanted ]))
- buckets_per_group = 1 << bucket_bits
+ parser = argparse.ArgumentParser()
+ parser.add_argument("--fio-version", dest="fio_version",
+ default="3", choices=[2,3], type=int,
+ help="fio version (default=3)")
+ parser.add_argument("--bucket-groups", dest="bucket_groups", default="29", type=int,
+ help="fio histogram bucket groups (default=29)")
+ parser.add_argument("--bucket-bits", dest="bucket_bits",
+ default="6", type=int,
+ help="fio histogram buckets-per-group bits (default=6 means 64 buckets/group)")
+ parser.add_argument("--percentiles", dest="pctiles_wanted",
+ default=[ 0., 50., 95., 99., 100.], type=float, nargs='+',
+ help="fio histogram buckets-per-group bits (default=6 means 64 buckets/group)")
+ parser.add_argument("--time-quantum", dest="time_quantum",
+ default="1", type=int,
+ help="time quantum in seconds (default=1)")
+ parser.add_argument("--log-hist-msec", dest="log_hist_msec",
+ type=int, default=None,
+ help="log_hist_msec value in fio job file")
+ parser.add_argument("--output-unit", dest="output_unit",
+ default="usec", type=str,
+ help="Latency percentile output unit: msec|usec|nsec (default usec)")
+ parser.add_argument("file_list", nargs='+',
+ help='list of files, preceded by " -- " if necessary')
+ args = parser.parse_args()
+
+ # default changes based on fio version
+ if args.fio_version == 2:
+ args.bucket_groups = 19
+
+ # print parameters
+
+ print('fio version = %d' % args.fio_version)
+ print('bucket groups = %d' % args.bucket_groups)
+ print('bucket bits = %d' % args.bucket_bits)
+ print('time quantum = %d sec' % args.time_quantum)
+ print('percentiles = %s' % ','.join([ str(p) for p in args.pctiles_wanted ]))
+ buckets_per_group = 1 << args.bucket_bits
print('buckets per group = %d' % buckets_per_group)
- buckets_per_interval = buckets_per_group * bucket_groups
+ buckets_per_interval = buckets_per_group * args.bucket_groups
print('buckets per interval = %d ' % buckets_per_interval)
bucket_index_range = range(0, buckets_per_interval)
- if time_quantum == 0:
- usage('time-quantum must be a positive number of seconds')
- print('output unit = ' + output_unit)
- if output_unit == 'msec':
- time_divisor = 1000.0
- elif output_unit == 'usec':
+ if args.log_hist_msec != None:
+ print('log_hist_msec = %d' % args.log_hist_msec)
+ if args.time_quantum == 0:
+ print('ERROR: time-quantum must be a positive number of seconds')
+ print('output unit = ' + args.output_unit)
+ if args.output_unit == 'msec':
+ time_divisor = float(msec_per_sec)
+ elif args.output_unit == 'usec':
time_divisor = 1.0
- # calculate response time interval associated with each histogram bucket
-
- bucket_times = time_ranges(bucket_groups, buckets_per_group, fio_version=fio_version)
-
# construct template for each histogram bucket array with buckets all zeroes
# we just copy this for each new histogram
zeroed_buckets = [ 0.0 for r in bucket_index_range ]
- # print CSV header just like fiologparser_hist does
+ # calculate response time interval associated with each histogram bucket
- header = 'msec, '
- for p in pctiles_wanted:
- header += '%3.1f, ' % p
- print('time (millisec), percentiles in increasing order with values in ' + output_unit)
- print(header)
+ bucket_times = time_ranges(args.bucket_groups, buckets_per_group, fio_version=args.fio_version)
# parse the histogram logs
# assumption: each bucket has a monotonically increasing time
# (exception: if randrw workload, then there is a read and a write
# record for the same time interval)
- max_timestamp_all_logs = 0
+ test_start_time = 0
+ test_end_time = 1.0e18
hist_files = {}
- for fn in file_list:
+ for fn in args.file_list:
try:
- (hist_files[fn], max_timestamp_ms) = parse_hist_file(fn, buckets_per_interval)
+ (hist_files[fn], log_start_time, log_end_time) = parse_hist_file(fn, buckets_per_interval, args.log_hist_msec)
except FioHistoLogExc as e:
- usage(str(e))
- max_timestamp_all_logs = max(max_timestamp_all_logs, max_timestamp_ms)
-
- (end_time, time_interval_count) = get_time_intervals(time_quantum, max_timestamp_all_logs)
- all_threads_histograms = [ ((j*time_quantum*msec_per_sec), deepcopy(zeroed_buckets))
- for j in range(0, time_interval_count) ]
+ myabort(str(e))
+ # we consider the test started when all threads have started logging
+ test_start_time = max(test_start_time, log_start_time)
+ # we consider the test over when one of the logs has ended
+ test_end_time = min(test_end_time, log_end_time)
+
+ if test_start_time >= test_end_time:
+ raise FioHistoLogExc('no time interval when all threads logs overlapped')
+ if test_start_time > 0:
+ print('all threads running as of unix epoch time %d = %s' % (
+ test_start_time/float(msec_per_sec),
+ time.ctime(test_start_time/1000.0)))
+
+ (end_time, time_interval_count) = get_time_intervals(args.time_quantum, test_start_time, test_end_time)
+ all_threads_histograms = [ ((j*args.time_quantum*msec_per_sec), deepcopy(zeroed_buckets))
+ for j in range(0, time_interval_count) ]
for logfn in hist_files.keys():
aligned_per_thread = align_histo_log(hist_files[logfn],
- time_quantum,
+ args.time_quantum,
buckets_per_interval,
- max_timestamp_all_logs)
+ test_start_time,
+ test_end_time)
for t in range(0, time_interval_count):
(_, all_threads_histo_t) = all_threads_histograms[t]
(_, log_histo_t) = aligned_per_thread[t]
- pct = get_pctiles(log_histo_t, pctiles_wanted, bucket_times)
add_to_histo_from( all_threads_histo_t, log_histo_t )
- print('percentiles for entire set of threads')
+ # calculate percentiles across aggregate histogram for all threads
+ # print CSV header just like fiologparser_hist does
+
+ header = 'msec-since-start, '
+ for p in args.pctiles_wanted:
+ header += '%3.1f, ' % p
+ print('time (millisec), percentiles in increasing order with values in ' + args.output_unit)
+ print(header)
+
for (t_msec, all_threads_histo_t) in all_threads_histograms:
- record = '%d, ' % t_msec
- pct = get_pctiles(all_threads_histo_t, pctiles_wanted, bucket_times)
+ record = '%8d, ' % t_msec
+ pct = get_pctiles(all_threads_histo_t, args.pctiles_wanted, bucket_times)
if not pct:
- for w in pctiles_wanted:
+ for w in args.pctiles_wanted:
record += ', '
else:
pct_keys = [ k for k in pct.keys() ]
with open(self.fn, 'w') as f:
f.write('1234, 0, 4096, 1, 2, 3, 4\n')
f.write('5678,1,16384,5,6,7,8 \n')
- (raw_histo_log, max_timestamp) = parse_hist_file(self.fn, 4) # 4 buckets per interval
- self.A(len(raw_histo_log) == 2 and max_timestamp == 5678)
+ (raw_histo_log, min_timestamp, max_timestamp) = parse_hist_file(self.fn, 4, None) # 4 buckets per interval
+ # if not log_unix_epoch=1, then min_timestamp will always be set to zero
+ self.A(len(raw_histo_log) == 2 and min_timestamp == 0 and max_timestamp == 5678)
(time_ms, direction, bsz, histo) = raw_histo_log[0]
self.A(time_ms == 1234 and direction == 0 and bsz == 4096 and histo == [ 1, 2, 3, 4 ])
(time_ms, direction, bsz, histo) = raw_histo_log[1]
with open(self.fn, 'w') as f:
pass
try:
- (raw_histo_log, max_timestamp_ms) = parse_hist_file(self.fn, 4)
+ (raw_histo_log, _, _) = parse_hist_file(self.fn, 4, None)
self.A(should_not_get_here)
except FioHistoLogExc as e:
self.A(str(e).startswith('no records'))
f.write('1234, 0, 4096, 1, 2, 3, 4\n')
f.write('5678,1,16384,5,6,7,8 \n')
f.write('\n')
- (raw_histo_log, max_timestamp_ms) = parse_hist_file(self.fn, 4)
+ (raw_histo_log, _, max_timestamp_ms) = parse_hist_file(self.fn, 4, None)
self.A(len(raw_histo_log) == 2 and max_timestamp_ms == 5678)
(time_ms, direction, bsz, histo) = raw_histo_log[0]
self.A(time_ms == 1234 and direction == 0 and bsz == 4096 and histo == [ 1, 2, 3, 4 ])
with open(self.fn, 'w') as f:
f.write('12, 0, 4096, 1a, 2, 3, 4\n')
try:
- (raw_histo_log, _) = parse_hist_file(self.fn, 4)
+ (raw_histo_log, _, _) = parse_hist_file(self.fn, 4, None)
self.A(False)
except FioHistoLogExc as e:
self.A(str(e).startswith('non-integer'))
with open(self.fn, 'w') as f:
f.write('-12, 0, 4096, 1, 2, 3, 4\n')
try:
- (raw_histo_log, _) = parse_hist_file(self.fn, 4)
+ (raw_histo_log, _, _) = parse_hist_file(self.fn, 4, None)
self.A(False)
except FioHistoLogExc as e:
self.A(str(e).startswith('negative integer'))
with open(self.fn, 'w') as f:
f.write('0, 0\n')
try:
- (raw_histo_log, _) = parse_hist_file(self.fn, 4)
+ (raw_histo_log, _, _) = parse_hist_file(self.fn, 4, None)
self.A(False)
except FioHistoLogExc as e:
self.A(str(e).startswith('too few numbers'))
with open(self.fn, 'w') as f:
f.write('100, 2, 4096, 1, 2, 3, 4\n')
try:
- (raw_histo_log, _) = parse_hist_file(self.fn, 4)
+ (raw_histo_log, _, _) = parse_hist_file(self.fn, 4, None)
self.A(False)
except FioHistoLogExc as e:
self.A(str(e).startswith('invalid I/O direction'))
def test_b8_parse_bsz_too_big(self):
with open(self.fn+'_good', 'w') as f:
f.write('100, 1, %d, 1, 2, 3, 4\n' % (1<<24))
- (raw_histo_log, max_timestamp_ms) = parse_hist_file(self.fn+'_good', 4)
+ (raw_histo_log, _, _) = parse_hist_file(self.fn+'_good', 4, None)
with open(self.fn+'_bad', 'w') as f:
f.write('100, 1, 20000000, 1, 2, 3, 4\n')
try:
- (raw_histo_log, _) = parse_hist_file(self.fn+'_bad', 4)
+ (raw_histo_log, _, _) = parse_hist_file(self.fn+'_bad', 4, None)
self.A(False)
except FioHistoLogExc as e:
self.A(str(e).startswith('block size too large'))
with open(self.fn, 'w') as f:
f.write('100, 1, %d, 1, 2, 3, 4, 5\n' % (1<<24))
try:
- (raw_histo_log, _) = parse_hist_file(self.fn, 4)
+ (raw_histo_log, _, _) = parse_hist_file(self.fn, 4, None)
self.A(False)
except FioHistoLogExc as e:
self.A(str(e).__contains__('buckets per interval'))
def test_d1_align_histo_log_1_quantum(self):
with open(self.fn, 'w') as f:
f.write('100, 1, 4096, 1, 2, 3, 4')
- (raw_histo_log, max_timestamp_ms) = parse_hist_file(self.fn, 4)
- self.A(max_timestamp_ms == 100)
- aligned_log = align_histo_log(raw_histo_log, 5, 4, max_timestamp_ms)
+ (raw_histo_log, min_timestamp_ms, max_timestamp_ms) = parse_hist_file(self.fn, 4, None)
+ self.A(min_timestamp_ms == 0 and max_timestamp_ms == 100)
+ aligned_log = align_histo_log(raw_histo_log, 5, 4, min_timestamp_ms, max_timestamp_ms)
self.A(len(aligned_log) == 1)
(time_ms0, h) = aligned_log[0]
- self.A(time_ms0 == 0 and h == [1.0, 2.0, 3.0, 4.0])
+ self.A(time_ms0 == 0 and h == [1., 2., 3., 4.])
+
+ # handle case with log_unix_epoch=1 timestamps, 1-second time quantum
+ # here both records will be separated into 2 aligned intervals
+
+ def test_d1a_align_2rec_histo_log_epoch_1_quantum_1sec(self):
+ with open(self.fn, 'w') as f:
+ f.write('1536504002123, 1, 4096, 1, 2, 3, 4\n')
+ f.write('1536504003123, 1, 4096, 4, 3, 2, 1\n')
+ (raw_histo_log, min_timestamp_ms, max_timestamp_ms) = parse_hist_file(self.fn, 4, None)
+ self.A(min_timestamp_ms == 1536504001123 and max_timestamp_ms == 1536504003123)
+ aligned_log = align_histo_log(raw_histo_log, 1, 4, min_timestamp_ms, max_timestamp_ms)
+ self.A(len(aligned_log) == 3)
+ (time_ms0, h) = aligned_log[0]
+ self.A(time_ms0 == 1536504001123 and h == [0., 0., 0., 0.])
+ (time_ms1, h) = aligned_log[1]
+ self.A(time_ms1 == 1536504002123 and h == [1., 2., 3., 4.])
+ (time_ms2, h) = aligned_log[2]
+ self.A(time_ms2 == 1536504003123 and h == [4., 3., 2., 1.])
+
+ # handle case with log_unix_epoch=1 timestamps, 5-second time quantum
+ # here both records will be merged into a single aligned time interval
+
+ def test_d1b_align_2rec_histo_log_epoch_1_quantum_5sec(self):
+ with open(self.fn, 'w') as f:
+ f.write('1536504002123, 1, 4096, 1, 2, 3, 4\n')
+ f.write('1536504003123, 1, 4096, 4, 3, 2, 1\n')
+ (raw_histo_log, min_timestamp_ms, max_timestamp_ms) = parse_hist_file(self.fn, 4, None)
+ self.A(min_timestamp_ms == 1536504001123 and max_timestamp_ms == 1536504003123)
+ aligned_log = align_histo_log(raw_histo_log, 5, 4, min_timestamp_ms, max_timestamp_ms)
+ self.A(len(aligned_log) == 1)
+ (time_ms0, h) = aligned_log[0]
+ self.A(time_ms0 == 1536504001123 and h == [5., 5., 5., 5.])
# we need this to compare 2 lists of floating point numbers for equality
# because of floating-point imprecision
with open(self.fn, 'w') as f:
f.write('2000, 1, 4096, 1, 2, 3, 4\n')
f.write('7000, 1, 4096, 1, 2, 3, 4\n')
- (raw_histo_log, max_timestamp_ms) = parse_hist_file(self.fn, 4)
- self.A(max_timestamp_ms == 7000)
+ (raw_histo_log, min_timestamp_ms, max_timestamp_ms) = parse_hist_file(self.fn, 4, None)
+ self.A(min_timestamp_ms == 0 and max_timestamp_ms == 7000)
(_, _, _, raw_buckets1) = raw_histo_log[0]
(_, _, _, raw_buckets2) = raw_histo_log[1]
- aligned_log = align_histo_log(raw_histo_log, 5, 4, max_timestamp_ms)
+ aligned_log = align_histo_log(raw_histo_log, 5, 4, min_timestamp_ms, max_timestamp_ms)
self.A(len(aligned_log) == 2)
(time_ms1, h1) = aligned_log[0]
(time_ms2, h2) = aligned_log[1]
self.A(time_ms1 == 0 and self.is_close(h1, expect1))
self.A(time_ms2 == 5000 and self.is_close(h2, expect2))
- def test_e1_get_pctiles(self):
+ # what to expect if histogram buckets are all equal
+ def test_e1_get_pctiles_flat_histo(self):
with open(self.fn, 'w') as f:
buckets = [ 100 for j in range(0, 128) ]
f.write('9000, 1, 4096, %s\n' % ', '.join([str(b) for b in buckets]))
- (raw_histo_log, max_timestamp_ms) = parse_hist_file(self.fn, 128)
- self.A(max_timestamp_ms == 9000)
- aligned_log = align_histo_log(raw_histo_log, 5, 128, max_timestamp_ms)
+ (raw_histo_log, min_timestamp_ms, max_timestamp_ms) = parse_hist_file(self.fn, 128, None)
+ self.A(min_timestamp_ms == 0 and max_timestamp_ms == 9000)
+ aligned_log = align_histo_log(raw_histo_log, 5, 128, min_timestamp_ms, max_timestamp_ms)
time_intervals = time_ranges(4, 32)
# since buckets are all equal, then median is halfway through time_intervals
# and max latency interval is at end of time_intervals
for (time_ms, histo) in aligned_log:
pct_vs_time.append(get_pctiles(histo, pctiles_wanted, time_intervals))
self.A(pct_vs_time[0] == None) # no I/O in this time interval
- expected_pctiles = { 0:0.001, 50:0.066, 100:0.256 }
+ expected_pctiles = { 0:0.000, 50:0.064, 100:0.256 }
self.A(pct_vs_time[1] == expected_pctiles)
+ # what to expect if just the highest histogram bucket is used
+ def test_e2_get_pctiles_highest_pct(self):
+ fio_v3_bucket_count = 29 * 64
+ with open(self.fn, 'w') as f:
+ # make a empty fio v3 histogram
+ buckets = [ 0 for j in range(0, fio_v3_bucket_count) ]
+ # add one I/O request to last bucket
+ buckets[-1] = 1
+ f.write('9000, 1, 4096, %s\n' % ', '.join([str(b) for b in buckets]))
+ (raw_histo_log, min_timestamp_ms, max_timestamp_ms) = parse_hist_file(self.fn, fio_v3_bucket_count, None)
+ self.A(min_timestamp_ms == 0 and max_timestamp_ms == 9000)
+ aligned_log = align_histo_log(raw_histo_log, 5, fio_v3_bucket_count, min_timestamp_ms, max_timestamp_ms)
+ (time_ms, histo) = aligned_log[1]
+ time_intervals = time_ranges(29, 64)
+ expected_pctiles = { 100.0:(64*(1<<28))/1000.0 }
+ pct = get_pctiles( histo, [ 100.0 ], time_intervals )
+ self.A(pct == expected_pctiles)
+
# we are using this module as a standalone program
if __name__ == '__main__':