Commit | Line | Data |
---|---|---|
d8296fdd BE |
1 | #!/usr/bin/env python |
2 | ||
3 | # module to parse fio histogram log files, not using pandas | |
4 | # runs in python v2 or v3 | |
5 | # to get help with the CLI: $ python fio-histo-log-pctiles.py -h | |
6 | # this can be run standalone as a script but is callable | |
7 | # assumes all threads run for same time duration | |
8 | # assumes all threads are doing the same thing for the entire run | |
9 | ||
10 | # percentiles: | |
11 | # 0 - min latency | |
12 | # 50 - median | |
13 | # 100 - max latency | |
14 | ||
15 | # TO-DO: | |
16 | # separate read and write stats for randrw mixed workload | |
17 | # report average latency if needed | |
18 | # prove that it works (partially done with unit tests) | |
19 | ||
20 | # to run unit tests, set UNITTEST environment variable to anything | |
21 | # if you do this, don't pass normal CLI parameters to it | |
22 | # otherwise it runs the CLI | |
23 | ||
24 | import sys, os, math, copy | |
25 | from copy import deepcopy | |
c670cb44 | 26 | import argparse |
088a092e BE |
27 | |
28 | unittest2_imported = True | |
29 | try: | |
30 | import unittest2 | |
31 | except ImportError: | |
32 | unittest2_imported = False | |
d8296fdd BE |
33 | |
34 | msec_per_sec = 1000 | |
35 | nsec_per_usec = 1000 | |
36 | ||
37 | class FioHistoLogExc(Exception): | |
38 | pass | |
39 | ||
c670cb44 | 40 | # if there is an error, print message, and exit with error status |
d8296fdd | 41 | |
c670cb44 | 42 | def myabort(msg): |
d8296fdd | 43 | print('ERROR: ' + msg) |
d8296fdd BE |
44 | sys.exit(1) |
45 | ||
d8296fdd BE |
46 | # convert histogram log file into a list of |
47 | # (time_ms, direction, bsz, buckets) tuples where | |
48 | # - time_ms is the time in msec at which the log record was written | |
49 | # - direction is 0 (read) or 1 (write) | |
50 | # - bsz is block size (not used) | |
51 | # - buckets is a CSV list of counters that make up the histogram | |
52 | # caller decides if the expected number of counters are present | |
53 | ||
54 | ||
55 | def exception_suffix( record_num, pathname ): | |
56 | return 'in histogram record %d file %s' % (record_num+1, pathname) | |
57 | ||
58 | # log file parser raises FioHistoLogExc exceptions | |
59 | # it returns histogram buckets in whatever unit fio uses | |
60 | ||
61 | def parse_hist_file(logfn, buckets_per_interval): | |
62 | max_timestamp_ms = 0.0 | |
63 | ||
64 | with open(logfn, 'r') as f: | |
65 | records = [ l.strip() for l in f.readlines() ] | |
66 | intervals = [] | |
67 | for k, r in enumerate(records): | |
68 | if r == '': | |
69 | continue | |
70 | tokens = r.split(',') | |
71 | try: | |
72 | int_tokens = [ int(t) for t in tokens ] | |
73 | except ValueError as e: | |
74 | raise FioHistoLogExc('non-integer value %s' % exception_suffix(k+1, logfn)) | |
75 | ||
76 | neg_ints = list(filter( lambda tk : tk < 0, int_tokens )) | |
77 | if len(neg_ints) > 0: | |
78 | raise FioHistoLogExc('negative integer value %s' % exception_suffix(k+1, logfn)) | |
79 | ||
80 | if len(int_tokens) < 3: | |
81 | raise FioHistoLogExc('too few numbers %s' % exception_suffix(k+1, logfn)) | |
82 | ||
83 | time_ms = int_tokens[0] | |
84 | if time_ms > max_timestamp_ms: | |
85 | max_timestamp_ms = time_ms | |
86 | ||
87 | direction = int_tokens[1] | |
88 | if direction != 0 and direction != 1: | |
89 | raise FioHistoLogExc('invalid I/O direction %s' % exception_suffix(k+1, logfn)) | |
90 | ||
91 | bsz = int_tokens[2] | |
92 | if bsz > (1 << 24): | |
93 | raise FioHistoLogExc('block size too large %s' % exception_suffix(k+1, logfn)) | |
94 | ||
95 | buckets = int_tokens[3:] | |
96 | if len(buckets) != buckets_per_interval: | |
97 | raise FioHistoLogExc('%d buckets per interval but %d expected in %s' % | |
98 | (len(buckets), buckets_per_interval, exception_suffix(k+1, logfn))) | |
99 | intervals.append((time_ms, direction, bsz, buckets)) | |
100 | if len(intervals) == 0: | |
101 | raise FioHistoLogExc('no records in %s' % logfn) | |
102 | return (intervals, max_timestamp_ms) | |
103 | ||
104 | ||
105 | # compute time range for each bucket index in histogram record | |
106 | # see comments in https://github.com/axboe/fio/blob/master/stat.h | |
107 | # for description of bucket groups and buckets | |
108 | # fio v3 bucket ranges are in nanosec (since response times are measured in nanosec) | |
109 | # but we convert fio v3 nanosecs to floating-point microseconds | |
110 | ||
111 | def time_ranges(groups, counters_per_group, fio_version=3): | |
112 | bucket_width = 1 | |
113 | bucket_base = 0 | |
114 | bucket_intervals = [] | |
115 | for g in range(0, groups): | |
116 | for b in range(0, counters_per_group): | |
117 | rmin = float(bucket_base) | |
118 | rmax = rmin + bucket_width | |
119 | if fio_version == 3: | |
120 | rmin /= nsec_per_usec | |
121 | rmax /= nsec_per_usec | |
122 | bucket_intervals.append( [rmin, rmax] ) | |
123 | bucket_base += bucket_width | |
124 | if g != 0: | |
125 | bucket_width *= 2 | |
126 | return bucket_intervals | |
127 | ||
128 | ||
129 | # compute number of time quantum intervals in the test | |
130 | ||
131 | def get_time_intervals(time_quantum, max_timestamp_ms): | |
132 | # round down to nearest second | |
133 | max_timestamp = max_timestamp_ms // msec_per_sec | |
134 | # round up to nearest whole multiple of time_quantum | |
135 | time_interval_count = (max_timestamp + time_quantum) // time_quantum | |
136 | end_time = time_interval_count * time_quantum | |
137 | return (end_time, time_interval_count) | |
138 | ||
139 | # align raw histogram log data to time quantum so | |
140 | # we can then combine histograms from different threads with addition | |
141 | # for randrw workload we count both reads and writes in same output bucket | |
142 | # but we separate reads and writes for purposes of calculating | |
143 | # end time for histogram record. | |
144 | # this requires us to weight a raw histogram bucket by the | |
145 | # fraction of time quantum that the bucket overlaps the current | |
146 | # time quantum interval | |
147 | # for example, if we have a bucket with 515 samples for time interval | |
148 | # [ 1010, 2014 ] msec since start of test, and time quantum is 1 sec, then | |
149 | # for time quantum interval [ 1000, 2000 ] msec, the overlap is | |
150 | # (2000 - 1010) / (2000 - 1000) = 0.99 | |
151 | # so the contribution of this bucket to this time quantum is | |
152 | # 515 x 0.99 = 509.85 | |
153 | ||
154 | def align_histo_log(raw_histogram_log, time_quantum, bucket_count, max_timestamp_ms): | |
155 | ||
156 | # slice up test time int intervals of time_quantum seconds | |
157 | ||
158 | (end_time, time_interval_count) = get_time_intervals(time_quantum, max_timestamp_ms) | |
159 | time_qtm_ms = time_quantum * msec_per_sec | |
160 | end_time_ms = end_time * msec_per_sec | |
161 | aligned_intervals = [] | |
162 | for j in range(0, time_interval_count): | |
163 | aligned_intervals.append(( | |
164 | j * time_qtm_ms, | |
165 | [ 0.0 for j in range(0, bucket_count) ] )) | |
166 | ||
167 | log_record_count = len(raw_histogram_log) | |
168 | for k, record in enumerate(raw_histogram_log): | |
169 | ||
170 | # find next record with same direction to get end-time | |
171 | # have to avoid going past end of array | |
172 | # for fio randrw workload, | |
173 | # we have read and write records on same time interval | |
174 | # sometimes read and write records are in opposite order | |
175 | # assertion checks that next read/write record | |
176 | # can be separated by at most 2 other records | |
177 | ||
178 | (time_msec, direction, sz, interval_buckets) = record | |
179 | if k+1 < log_record_count: | |
180 | (time_msec_end, direction2, _, _) = raw_histogram_log[k+1] | |
181 | if direction2 != direction: | |
182 | if k+2 < log_record_count: | |
183 | (time_msec_end, direction2, _, _) = raw_histogram_log[k+2] | |
184 | if direction2 != direction: | |
185 | if k+3 < log_record_count: | |
186 | (time_msec_end, direction2, _, _) = raw_histogram_log[k+3] | |
187 | assert direction2 == direction | |
188 | else: | |
189 | time_msec_end = end_time_ms | |
190 | else: | |
191 | time_msec_end = end_time_ms | |
192 | else: | |
193 | time_msec_end = end_time_ms | |
194 | ||
195 | # calculate first quantum that overlaps this histogram record | |
196 | ||
197 | qtm_start_ms = (time_msec // time_qtm_ms) * time_qtm_ms | |
198 | qtm_end_ms = ((time_msec + time_qtm_ms) // time_qtm_ms) * time_qtm_ms | |
199 | qtm_index = qtm_start_ms // time_qtm_ms | |
200 | ||
201 | # for each quantum that overlaps this histogram record's time interval | |
202 | ||
203 | while qtm_start_ms < time_msec_end: # while quantum overlaps record | |
204 | ||
205 | # calculate fraction of time that this quantum | |
206 | # overlaps histogram record's time interval | |
207 | ||
208 | overlap_start = max(qtm_start_ms, time_msec) | |
209 | overlap_end = min(qtm_end_ms, time_msec_end) | |
210 | weight = float(overlap_end - overlap_start) | |
211 | weight /= (time_msec_end - time_msec) | |
212 | (_,aligned_histogram) = aligned_intervals[qtm_index] | |
213 | for bx, b in enumerate(interval_buckets): | |
214 | weighted_bucket = weight * b | |
215 | aligned_histogram[bx] += weighted_bucket | |
216 | ||
217 | # advance to the next time quantum | |
218 | ||
219 | qtm_start_ms += time_qtm_ms | |
220 | qtm_end_ms += time_qtm_ms | |
221 | qtm_index += 1 | |
222 | ||
223 | return aligned_intervals | |
224 | ||
225 | # add histogram in "source" to histogram in "target" | |
226 | # it is assumed that the 2 histograms are precisely time-aligned | |
227 | ||
228 | def add_to_histo_from( target, source ): | |
229 | for b in range(0, len(source)): | |
230 | target[b] += source[b] | |
231 | ||
232 | # compute percentiles | |
233 | # inputs: | |
234 | # buckets: histogram bucket array | |
235 | # wanted: list of floating-pt percentiles to calculate | |
236 | # time_ranges: [tmin,tmax) time interval for each bucket | |
237 | # returns None if no I/O reported. | |
238 | # otherwise we would be dividing by zero | |
239 | # think of buckets as probability distribution function | |
240 | # and this loop is integrating to get cumulative distribution function | |
241 | ||
242 | def get_pctiles(buckets, wanted, time_ranges): | |
243 | ||
244 | # get total of IO requests done | |
245 | total_ios = 0 | |
246 | for io_count in buckets: | |
247 | total_ios += io_count | |
248 | ||
249 | # don't return percentiles if no I/O was done during interval | |
250 | if total_ios == 0.0: | |
251 | return None | |
252 | ||
253 | pctile_count = len(wanted) | |
254 | ||
255 | # results returned as dictionary keyed by percentile | |
256 | pctile_result = {} | |
257 | ||
258 | # index of next percentile in list | |
259 | pctile_index = 0 | |
260 | ||
261 | # next percentile | |
262 | next_pctile = wanted[pctile_index] | |
263 | ||
264 | # no one is interested in percentiles bigger than this but not 100.0 | |
265 | # this prevents floating-point error from preventing loop exit | |
266 | almost_100 = 99.9999 | |
267 | ||
0456267b BE |
268 | # pct is the percentile corresponding to |
269 | # all I/O requests up through bucket b | |
270 | pct = 0.0 | |
d8296fdd BE |
271 | total_so_far = 0 |
272 | for b, io_count in enumerate(buckets): | |
0456267b BE |
273 | if io_count == 0: |
274 | continue | |
d8296fdd | 275 | total_so_far += io_count |
0456267b BE |
276 | # last_pct_lt is the percentile corresponding to |
277 | # all I/O requests up to, but not including, bucket b | |
278 | last_pct = pct | |
279 | pct = 100.0 * float(total_so_far) / total_ios | |
d8296fdd BE |
280 | # a single bucket could satisfy multiple pctiles |
281 | # so this must be a while loop | |
0456267b BE |
282 | # for 100-percentile (max latency) case, no bucket exceeds it |
283 | # so we must stop there. | |
284 | while ((next_pctile == 100.0 and pct >= almost_100) or | |
285 | (next_pctile < 100.0 and pct > next_pctile)): | |
286 | # interpolate between min and max time for bucket time interval | |
287 | # we keep the time_ranges access inside this loop, | |
288 | # even though it could be above the loop, | |
289 | # because in many cases we will not be even entering | |
290 | # the loop so we optimize out these accesses | |
d8296fdd | 291 | range_max_time = time_ranges[b][1] |
0456267b BE |
292 | range_min_time = time_ranges[b][0] |
293 | offset_frac = (next_pctile - last_pct)/(pct - last_pct) | |
294 | interpolation = range_min_time + (offset_frac*(range_max_time - range_min_time)) | |
295 | pctile_result[next_pctile] = interpolation | |
d8296fdd BE |
296 | pctile_index += 1 |
297 | if pctile_index == pctile_count: | |
298 | break | |
299 | next_pctile = wanted[pctile_index] | |
300 | if pctile_index == pctile_count: | |
301 | break | |
302 | assert pctile_index == pctile_count | |
303 | return pctile_result | |
304 | ||
305 | ||
c670cb44 | 306 | # this is really the main program |
d8296fdd | 307 | |
c670cb44 BE |
308 | def compute_percentiles_from_logs(): |
309 | parser = argparse.ArgumentParser() | |
310 | parser.add_argument("--fio-version", dest="fio_version", | |
311 | default="3", choices=[2,3], type=int, | |
312 | help="fio version (default=3)") | |
313 | parser.add_argument("--bucket-groups", dest="bucket_groups", default="29", type=int, | |
314 | help="fio histogram bucket groups (default=29)") | |
315 | parser.add_argument("--bucket-bits", dest="bucket_bits", | |
316 | default="6", type=int, | |
317 | help="fio histogram buckets-per-group bits (default=6 means 64 buckets/group)") | |
318 | parser.add_argument("--percentiles", dest="pctiles_wanted", | |
4b34a0eb | 319 | default=[ 0., 50., 95., 99., 100.], type=float, nargs='+', |
c670cb44 BE |
320 | help="fio histogram buckets-per-group bits (default=6 means 64 buckets/group)") |
321 | parser.add_argument("--time-quantum", dest="time_quantum", | |
322 | default="1", type=int, | |
323 | help="time quantum in seconds (default=1)") | |
324 | parser.add_argument("--output-unit", dest="output_unit", | |
325 | default="usec", type=str, | |
326 | help="Latency percentile output unit: msec|usec|nsec (default usec)") | |
4b34a0eb BE |
327 | parser.add_argument("file_list", nargs='+', |
328 | help='list of files, preceded by " -- " if necessary') | |
c670cb44 | 329 | args = parser.parse_args() |
c670cb44 | 330 | |
4b34a0eb BE |
331 | # default changes based on fio version |
332 | if args.fio_version == 2: | |
333 | args.bucket_groups = 19 | |
d8296fdd | 334 | |
c670cb44 | 335 | # print parameters |
d8296fdd | 336 | |
4b34a0eb | 337 | print('fio version = %d' % args.fio_version) |
c670cb44 BE |
338 | print('bucket groups = %d' % args.bucket_groups) |
339 | print('bucket bits = %d' % args.bucket_bits) | |
340 | print('time quantum = %d sec' % args.time_quantum) | |
341 | print('percentiles = %s' % ','.join([ str(p) for p in args.pctiles_wanted ])) | |
342 | buckets_per_group = 1 << args.bucket_bits | |
d8296fdd | 343 | print('buckets per group = %d' % buckets_per_group) |
c670cb44 | 344 | buckets_per_interval = buckets_per_group * args.bucket_groups |
d8296fdd BE |
345 | print('buckets per interval = %d ' % buckets_per_interval) |
346 | bucket_index_range = range(0, buckets_per_interval) | |
c670cb44 BE |
347 | if args.time_quantum == 0: |
348 | print('ERROR: time-quantum must be a positive number of seconds') | |
349 | print('output unit = ' + args.output_unit) | |
350 | if args.output_unit == 'msec': | |
d8296fdd | 351 | time_divisor = 1000.0 |
c670cb44 | 352 | elif args.output_unit == 'usec': |
d8296fdd BE |
353 | time_divisor = 1.0 |
354 | ||
355 | # calculate response time interval associated with each histogram bucket | |
356 | ||
c670cb44 | 357 | bucket_times = time_ranges(args.bucket_groups, buckets_per_group, fio_version=args.fio_version) |
d8296fdd BE |
358 | |
359 | # construct template for each histogram bucket array with buckets all zeroes | |
360 | # we just copy this for each new histogram | |
361 | ||
362 | zeroed_buckets = [ 0.0 for r in bucket_index_range ] | |
363 | ||
364 | # print CSV header just like fiologparser_hist does | |
365 | ||
366 | header = 'msec, ' | |
c670cb44 | 367 | for p in args.pctiles_wanted: |
d8296fdd | 368 | header += '%3.1f, ' % p |
c670cb44 | 369 | print('time (millisec), percentiles in increasing order with values in ' + args.output_unit) |
d8296fdd BE |
370 | print(header) |
371 | ||
372 | # parse the histogram logs | |
373 | # assumption: each bucket has a monotonically increasing time | |
374 | # assumption: time ranges do not overlap for a single thread's records | |
375 | # (exception: if randrw workload, then there is a read and a write | |
376 | # record for the same time interval) | |
377 | ||
378 | max_timestamp_all_logs = 0 | |
379 | hist_files = {} | |
c670cb44 | 380 | for fn in args.file_list: |
d8296fdd BE |
381 | try: |
382 | (hist_files[fn], max_timestamp_ms) = parse_hist_file(fn, buckets_per_interval) | |
383 | except FioHistoLogExc as e: | |
c670cb44 | 384 | myabort(str(e)) |
d8296fdd BE |
385 | max_timestamp_all_logs = max(max_timestamp_all_logs, max_timestamp_ms) |
386 | ||
c670cb44 BE |
387 | (end_time, time_interval_count) = get_time_intervals(args.time_quantum, max_timestamp_all_logs) |
388 | all_threads_histograms = [ ((j*args.time_quantum*msec_per_sec), deepcopy(zeroed_buckets)) | |
d8296fdd BE |
389 | for j in range(0, time_interval_count) ] |
390 | ||
391 | for logfn in hist_files.keys(): | |
392 | aligned_per_thread = align_histo_log(hist_files[logfn], | |
c670cb44 | 393 | args.time_quantum, |
d8296fdd BE |
394 | buckets_per_interval, |
395 | max_timestamp_all_logs) | |
396 | for t in range(0, time_interval_count): | |
397 | (_, all_threads_histo_t) = all_threads_histograms[t] | |
398 | (_, log_histo_t) = aligned_per_thread[t] | |
d8296fdd BE |
399 | add_to_histo_from( all_threads_histo_t, log_histo_t ) |
400 | ||
c670cb44 BE |
401 | # calculate percentiles across aggregate histogram for all threads |
402 | ||
d8296fdd BE |
403 | for (t_msec, all_threads_histo_t) in all_threads_histograms: |
404 | record = '%d, ' % t_msec | |
c670cb44 | 405 | pct = get_pctiles(all_threads_histo_t, args.pctiles_wanted, bucket_times) |
d8296fdd | 406 | if not pct: |
c670cb44 | 407 | for w in args.pctiles_wanted: |
d8296fdd BE |
408 | record += ', ' |
409 | else: | |
410 | pct_keys = [ k for k in pct.keys() ] | |
411 | pct_values = [ str(pct[wanted]/time_divisor) for wanted in sorted(pct_keys) ] | |
412 | record += ', '.join(pct_values) | |
413 | print(record) | |
414 | ||
415 | ||
416 | ||
417 | #end of MAIN PROGRAM | |
418 | ||
419 | ||
d8296fdd BE |
420 | ##### below are unit tests ############## |
421 | ||
088a092e BE |
422 | if unittest2_imported: |
423 | import tempfile, shutil | |
424 | from os.path import join | |
425 | should_not_get_here = False | |
d8296fdd | 426 | |
088a092e | 427 | class Test(unittest2.TestCase): |
d8296fdd BE |
428 | tempdir = None |
429 | ||
430 | # a little less typing please | |
431 | def A(self, boolean_val): | |
432 | self.assertTrue(boolean_val) | |
433 | ||
434 | # initialize unit test environment | |
435 | ||
436 | @classmethod | |
437 | def setUpClass(cls): | |
438 | d = tempfile.mkdtemp() | |
439 | Test.tempdir = d | |
440 | ||
441 | # remove anything left by unit test environment | |
442 | # unless user sets UNITTEST_LEAVE_FILES environment variable | |
443 | ||
444 | @classmethod | |
445 | def tearDownClass(cls): | |
446 | if not os.getenv("UNITTEST_LEAVE_FILES"): | |
447 | shutil.rmtree(cls.tempdir) | |
448 | ||
449 | def setUp(self): | |
450 | self.fn = join(Test.tempdir, self.id()) | |
451 | ||
452 | def test_a_add_histos(self): | |
453 | a = [ 1.0, 2.0 ] | |
454 | b = [ 1.5, 2.5 ] | |
455 | add_to_histo_from( a, b ) | |
456 | self.A(a == [2.5, 4.5]) | |
457 | self.A(b == [1.5, 2.5]) | |
458 | ||
459 | def test_b1_parse_log(self): | |
460 | with open(self.fn, 'w') as f: | |
461 | f.write('1234, 0, 4096, 1, 2, 3, 4\n') | |
462 | f.write('5678,1,16384,5,6,7,8 \n') | |
463 | (raw_histo_log, max_timestamp) = parse_hist_file(self.fn, 4) # 4 buckets per interval | |
464 | self.A(len(raw_histo_log) == 2 and max_timestamp == 5678) | |
465 | (time_ms, direction, bsz, histo) = raw_histo_log[0] | |
466 | self.A(time_ms == 1234 and direction == 0 and bsz == 4096 and histo == [ 1, 2, 3, 4 ]) | |
467 | (time_ms, direction, bsz, histo) = raw_histo_log[1] | |
468 | self.A(time_ms == 5678 and direction == 1 and bsz == 16384 and histo == [ 5, 6, 7, 8 ]) | |
469 | ||
470 | def test_b2_parse_empty_log(self): | |
471 | with open(self.fn, 'w') as f: | |
472 | pass | |
473 | try: | |
474 | (raw_histo_log, max_timestamp_ms) = parse_hist_file(self.fn, 4) | |
475 | self.A(should_not_get_here) | |
476 | except FioHistoLogExc as e: | |
477 | self.A(str(e).startswith('no records')) | |
478 | ||
479 | def test_b3_parse_empty_records(self): | |
480 | with open(self.fn, 'w') as f: | |
481 | f.write('\n') | |
482 | f.write('1234, 0, 4096, 1, 2, 3, 4\n') | |
483 | f.write('5678,1,16384,5,6,7,8 \n') | |
484 | f.write('\n') | |
485 | (raw_histo_log, max_timestamp_ms) = parse_hist_file(self.fn, 4) | |
486 | self.A(len(raw_histo_log) == 2 and max_timestamp_ms == 5678) | |
487 | (time_ms, direction, bsz, histo) = raw_histo_log[0] | |
488 | self.A(time_ms == 1234 and direction == 0 and bsz == 4096 and histo == [ 1, 2, 3, 4 ]) | |
489 | (time_ms, direction, bsz, histo) = raw_histo_log[1] | |
490 | self.A(time_ms == 5678 and direction == 1 and bsz == 16384 and histo == [ 5, 6, 7, 8 ]) | |
491 | ||
492 | def test_b4_parse_non_int(self): | |
493 | with open(self.fn, 'w') as f: | |
494 | f.write('12, 0, 4096, 1a, 2, 3, 4\n') | |
495 | try: | |
496 | (raw_histo_log, _) = parse_hist_file(self.fn, 4) | |
497 | self.A(False) | |
498 | except FioHistoLogExc as e: | |
499 | self.A(str(e).startswith('non-integer')) | |
500 | ||
501 | def test_b5_parse_neg_int(self): | |
502 | with open(self.fn, 'w') as f: | |
503 | f.write('-12, 0, 4096, 1, 2, 3, 4\n') | |
504 | try: | |
505 | (raw_histo_log, _) = parse_hist_file(self.fn, 4) | |
506 | self.A(False) | |
507 | except FioHistoLogExc as e: | |
508 | self.A(str(e).startswith('negative integer')) | |
509 | ||
510 | def test_b6_parse_too_few_int(self): | |
511 | with open(self.fn, 'w') as f: | |
512 | f.write('0, 0\n') | |
513 | try: | |
514 | (raw_histo_log, _) = parse_hist_file(self.fn, 4) | |
515 | self.A(False) | |
516 | except FioHistoLogExc as e: | |
517 | self.A(str(e).startswith('too few numbers')) | |
518 | ||
519 | def test_b7_parse_invalid_direction(self): | |
520 | with open(self.fn, 'w') as f: | |
521 | f.write('100, 2, 4096, 1, 2, 3, 4\n') | |
522 | try: | |
523 | (raw_histo_log, _) = parse_hist_file(self.fn, 4) | |
524 | self.A(False) | |
525 | except FioHistoLogExc as e: | |
526 | self.A(str(e).startswith('invalid I/O direction')) | |
527 | ||
528 | def test_b8_parse_bsz_too_big(self): | |
529 | with open(self.fn+'_good', 'w') as f: | |
530 | f.write('100, 1, %d, 1, 2, 3, 4\n' % (1<<24)) | |
531 | (raw_histo_log, max_timestamp_ms) = parse_hist_file(self.fn+'_good', 4) | |
532 | with open(self.fn+'_bad', 'w') as f: | |
533 | f.write('100, 1, 20000000, 1, 2, 3, 4\n') | |
534 | try: | |
535 | (raw_histo_log, _) = parse_hist_file(self.fn+'_bad', 4) | |
536 | self.A(False) | |
537 | except FioHistoLogExc as e: | |
538 | self.A(str(e).startswith('block size too large')) | |
539 | ||
540 | def test_b9_parse_wrong_bucket_count(self): | |
541 | with open(self.fn, 'w') as f: | |
542 | f.write('100, 1, %d, 1, 2, 3, 4, 5\n' % (1<<24)) | |
543 | try: | |
544 | (raw_histo_log, _) = parse_hist_file(self.fn, 4) | |
545 | self.A(False) | |
546 | except FioHistoLogExc as e: | |
547 | self.A(str(e).__contains__('buckets per interval')) | |
548 | ||
549 | def test_c1_time_ranges(self): | |
550 | ranges = time_ranges(3, 2) # fio_version defaults to 3 | |
551 | expected_ranges = [ # fio_version 3 is in nanoseconds | |
552 | [0.000, 0.001], [0.001, 0.002], # first group | |
553 | [0.002, 0.003], [0.003, 0.004], # second group same width | |
554 | [0.004, 0.006], [0.006, 0.008]] # subsequent groups double width | |
555 | self.A(ranges == expected_ranges) | |
556 | ranges = time_ranges(3, 2, fio_version=3) | |
557 | self.A(ranges == expected_ranges) | |
558 | ranges = time_ranges(3, 2, fio_version=2) | |
559 | expected_ranges_v2 = [ [ 1000.0 * min_or_max for min_or_max in time_range ] | |
560 | for time_range in expected_ranges ] | |
561 | self.A(ranges == expected_ranges_v2) | |
562 | # see fio V3 stat.h for why 29 groups and 2^6 buckets/group | |
563 | normal_ranges_v3 = time_ranges(29, 64) | |
564 | # for v3, bucket time intervals are measured in nanoseconds | |
565 | self.A(len(normal_ranges_v3) == 29 * 64 and normal_ranges_v3[-1][1] == 64*(1<<(29-1))/1000.0) | |
566 | normal_ranges_v2 = time_ranges(19, 64, fio_version=2) | |
567 | # for v2, bucket time intervals are measured in microseconds so we have fewer buckets | |
568 | self.A(len(normal_ranges_v2) == 19 * 64 and normal_ranges_v2[-1][1] == 64*(1<<(19-1))) | |
569 | ||
570 | def test_d1_align_histo_log_1_quantum(self): | |
571 | with open(self.fn, 'w') as f: | |
572 | f.write('100, 1, 4096, 1, 2, 3, 4') | |
573 | (raw_histo_log, max_timestamp_ms) = parse_hist_file(self.fn, 4) | |
574 | self.A(max_timestamp_ms == 100) | |
575 | aligned_log = align_histo_log(raw_histo_log, 5, 4, max_timestamp_ms) | |
576 | self.A(len(aligned_log) == 1) | |
577 | (time_ms0, h) = aligned_log[0] | |
578 | self.A(time_ms0 == 0 and h == [1.0, 2.0, 3.0, 4.0]) | |
579 | ||
580 | # we need this to compare 2 lists of floating point numbers for equality | |
581 | # because of floating-point imprecision | |
582 | ||
583 | def compare_2_floats(self, x, y): | |
584 | if x == 0.0 or y == 0.0: | |
585 | return (x+y) < 0.0000001 | |
586 | else: | |
587 | return (math.fabs(x-y)/x) < 0.00001 | |
588 | ||
589 | def is_close(self, buckets, buckets_expected): | |
590 | if len(buckets) != len(buckets_expected): | |
591 | return False | |
592 | compare_buckets = lambda k: self.compare_2_floats(buckets[k], buckets_expected[k]) | |
593 | indices_close = list(filter(compare_buckets, range(0, len(buckets)))) | |
594 | return len(indices_close) == len(buckets) | |
595 | ||
596 | def test_d2_align_histo_log_2_quantum(self): | |
597 | with open(self.fn, 'w') as f: | |
598 | f.write('2000, 1, 4096, 1, 2, 3, 4\n') | |
599 | f.write('7000, 1, 4096, 1, 2, 3, 4\n') | |
600 | (raw_histo_log, max_timestamp_ms) = parse_hist_file(self.fn, 4) | |
601 | self.A(max_timestamp_ms == 7000) | |
602 | (_, _, _, raw_buckets1) = raw_histo_log[0] | |
603 | (_, _, _, raw_buckets2) = raw_histo_log[1] | |
604 | aligned_log = align_histo_log(raw_histo_log, 5, 4, max_timestamp_ms) | |
605 | self.A(len(aligned_log) == 2) | |
606 | (time_ms1, h1) = aligned_log[0] | |
607 | (time_ms2, h2) = aligned_log[1] | |
608 | # because first record is from time interval [2000, 7000] | |
609 | # we weight it according | |
610 | expect1 = [float(b) * 0.6 for b in raw_buckets1] | |
611 | expect2 = [float(b) * 0.4 for b in raw_buckets1] | |
612 | for e in range(0, len(expect2)): | |
613 | expect2[e] += raw_buckets2[e] | |
614 | self.A(time_ms1 == 0 and self.is_close(h1, expect1)) | |
615 | self.A(time_ms2 == 5000 and self.is_close(h2, expect2)) | |
616 | ||
0456267b BE |
617 | # what to expect if histogram buckets are all equal |
618 | def test_e1_get_pctiles_flat_histo(self): | |
d8296fdd BE |
619 | with open(self.fn, 'w') as f: |
620 | buckets = [ 100 for j in range(0, 128) ] | |
621 | f.write('9000, 1, 4096, %s\n' % ', '.join([str(b) for b in buckets])) | |
622 | (raw_histo_log, max_timestamp_ms) = parse_hist_file(self.fn, 128) | |
623 | self.A(max_timestamp_ms == 9000) | |
624 | aligned_log = align_histo_log(raw_histo_log, 5, 128, max_timestamp_ms) | |
625 | time_intervals = time_ranges(4, 32) | |
626 | # since buckets are all equal, then median is halfway through time_intervals | |
627 | # and max latency interval is at end of time_intervals | |
628 | self.A(time_intervals[64][1] == 0.066 and time_intervals[127][1] == 0.256) | |
629 | pctiles_wanted = [ 0, 50, 100 ] | |
630 | pct_vs_time = [] | |
631 | for (time_ms, histo) in aligned_log: | |
632 | pct_vs_time.append(get_pctiles(histo, pctiles_wanted, time_intervals)) | |
633 | self.A(pct_vs_time[0] == None) # no I/O in this time interval | |
0456267b | 634 | expected_pctiles = { 0:0.000, 50:0.064, 100:0.256 } |
d8296fdd BE |
635 | self.A(pct_vs_time[1] == expected_pctiles) |
636 | ||
0456267b BE |
637 | # what to expect if just the highest histogram bucket is used |
638 | def test_e2_get_pctiles_highest_pct(self): | |
639 | fio_v3_bucket_count = 29 * 64 | |
640 | with open(self.fn, 'w') as f: | |
641 | # make a empty fio v3 histogram | |
642 | buckets = [ 0 for j in range(0, fio_v3_bucket_count) ] | |
643 | # add one I/O request to last bucket | |
644 | buckets[-1] = 1 | |
645 | f.write('9000, 1, 4096, %s\n' % ', '.join([str(b) for b in buckets])) | |
646 | (raw_histo_log, max_timestamp_ms) = parse_hist_file(self.fn, fio_v3_bucket_count) | |
647 | self.A(max_timestamp_ms == 9000) | |
648 | aligned_log = align_histo_log(raw_histo_log, 5, fio_v3_bucket_count, max_timestamp_ms) | |
649 | (time_ms, histo) = aligned_log[1] | |
650 | time_intervals = time_ranges(29, 64) | |
651 | expected_pctiles = { 100.0:(64*(1<<28))/1000.0 } | |
652 | pct = get_pctiles( histo, [ 100.0 ], time_intervals ) | |
653 | self.A(pct == expected_pctiles) | |
654 | ||
d8296fdd BE |
655 | # we are using this module as a standalone program |
656 | ||
657 | if __name__ == '__main__': | |
658 | if os.getenv('UNITTEST'): | |
088a092e BE |
659 | if unittest2_imported: |
660 | sys.exit(unittest2.main()) | |
661 | else: | |
662 | raise Exception('you must install unittest2 module to run unit test') | |
d8296fdd BE |
663 | else: |
664 | compute_percentiles_from_logs() | |
665 |