import sys, os, math, copy
from copy import deepcopy
-
+import argparse
import unittest2
msec_per_sec = 1000
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
# 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("--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':
+ 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 = 1000.0
- elif output_unit == 'usec':
+ 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)
+ bucket_times = time_ranges(args.bucket_groups, buckets_per_group, fio_version=args.fio_version)
# construct template for each histogram bucket array with buckets all zeroes
# we just copy this for each new histogram
# print CSV header just like fiologparser_hist does
header = 'msec, '
- for p in pctiles_wanted:
+ for p in args.pctiles_wanted:
header += '%3.1f, ' % p
- print('time (millisec), percentiles in increasing order with values in ' + output_unit)
+ print('time (millisec), percentiles in increasing order with values in ' + args.output_unit)
print(header)
# parse the histogram logs
max_timestamp_all_logs = 0
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)
except FioHistoLogExc as e:
- usage(str(e))
+ myabort(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))
+ (end_time, time_interval_count) = get_time_intervals(args.time_quantum, max_timestamp_all_logs)
+ 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)
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
+
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)
+ 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() ]
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]))
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, max_timestamp_ms) = parse_hist_file(self.fn, fio_v3_bucket_count)
+ self.A(max_timestamp_ms == 9000)
+ aligned_log = align_histo_log(raw_histo_log, 5, fio_v3_bucket_count, 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__':
projects / fio.git / blobdiff