blob: 90a7fb31a1bc2bba6fc43af8b509481918ab1a0c [file] [log] [blame]
Jens Axboea64e88d2011-10-03 14:20:01 +02001#ifndef FIO_STAT_H
2#define FIO_STAT_H
3
Jens Axboeec412652012-03-08 12:37:31 +01004#include "iolog.h"
5
Jens Axboea64e88d2011-10-03 14:20:01 +02006struct group_run_stats {
Shaohua Li6eaf09d2012-09-14 08:49:43 +02007 uint64_t max_run[DDIR_RWDIR_CNT], min_run[DDIR_RWDIR_CNT];
8 uint64_t max_bw[DDIR_RWDIR_CNT], min_bw[DDIR_RWDIR_CNT];
9 uint64_t io_kb[DDIR_RWDIR_CNT];
10 uint64_t agg[DDIR_RWDIR_CNT];
Jens Axboea64e88d2011-10-03 14:20:01 +020011 uint32_t kb_base;
Steven Noonanad705bc2013-04-08 15:05:25 -070012 uint32_t unit_base;
Jens Axboea64e88d2011-10-03 14:20:01 +020013 uint32_t groupid;
Jens Axboe771e58b2013-01-30 12:56:23 +010014 uint32_t unified_rw_rep;
Jens Axboeeb663202014-06-30 08:51:33 -060015} __attribute__((packed));
Jens Axboea64e88d2011-10-03 14:20:01 +020016
17/*
18 * How many depth levels to log
19 */
20#define FIO_IO_U_MAP_NR 7
21#define FIO_IO_U_LAT_U_NR 10
22#define FIO_IO_U_LAT_M_NR 12
23
24/*
25 * Aggregate clat samples to report percentile(s) of them.
26 *
27 * EXECUTIVE SUMMARY
28 *
29 * FIO_IO_U_PLAT_BITS determines the maximum statistical error on the
30 * value of resulting percentiles. The error will be approximately
31 * 1/2^(FIO_IO_U_PLAT_BITS+1) of the value.
32 *
33 * FIO_IO_U_PLAT_GROUP_NR and FIO_IO_U_PLAT_BITS determine the maximum
34 * range being tracked for latency samples. The maximum value tracked
35 * accurately will be 2^(GROUP_NR + PLAT_BITS -1) microseconds.
36 *
37 * FIO_IO_U_PLAT_GROUP_NR and FIO_IO_U_PLAT_BITS determine the memory
38 * requirement of storing those aggregate counts. The memory used will
39 * be (FIO_IO_U_PLAT_GROUP_NR * 2^FIO_IO_U_PLAT_BITS) * sizeof(int)
40 * bytes.
41 *
42 * FIO_IO_U_PLAT_NR is the total number of buckets.
43 *
44 * DETAILS
45 *
46 * Suppose the clat varies from 0 to 999 (usec), the straightforward
47 * method is to keep an array of (999 + 1) buckets, in which a counter
48 * keeps the count of samples which fall in the bucket, e.g.,
49 * {[0],[1],...,[999]}. However this consumes a huge amount of space,
50 * and can be avoided if an approximation is acceptable.
51 *
52 * One such method is to let the range of the bucket to be greater
53 * than one. This method has low accuracy when the value is small. For
54 * example, let the buckets be {[0,99],[100,199],...,[900,999]}, and
55 * the represented value of each bucket be the mean of the range. Then
56 * a value 0 has an round-off error of 49.5. To improve on this, we
57 * use buckets with non-uniform ranges, while bounding the error of
58 * each bucket within a ratio of the sample value. A simple example
59 * would be when error_bound = 0.005, buckets are {
60 * {[0],[1],...,[99]}, {[100,101],[102,103],...,[198,199]},..,
61 * {[900,909],[910,919]...} }. The total range is partitioned into
62 * groups with different ranges, then buckets with uniform ranges. An
63 * upper bound of the error is (range_of_bucket/2)/value_of_bucket
64 *
65 * For better efficiency, we implement this using base two. We group
66 * samples by their Most Significant Bit (MSB), extract the next M bit
67 * of them as an index within the group, and discard the rest of the
68 * bits.
69 *
70 * E.g., assume a sample 'x' whose MSB is bit n (starting from bit 0),
71 * and use M bit for indexing
72 *
73 * | n | M bits | bit (n-M-1) ... bit 0 |
74 *
75 * Because x is at least 2^n, and bit 0 to bit (n-M-1) is at most
76 * (2^(n-M) - 1), discarding bit 0 to (n-M-1) makes the round-off
77 * error
78 *
79 * 2^(n-M)-1 2^(n-M) 1
80 * e <= --------- <= ------- = ---
81 * 2^n 2^n 2^M
82 *
83 * Furthermore, we use "mean" of the range to represent the bucket,
84 * the error e can be lowered by half to 1 / 2^(M+1). By using M bits
85 * as the index, each group must contains 2^M buckets.
86 *
87 * E.g. Let M (FIO_IO_U_PLAT_BITS) be 6
88 * Error bound is 1/2^(6+1) = 0.0078125 (< 1%)
89 *
90 * Group MSB #discarded range of #buckets
91 * error_bits value
92 * ----------------------------------------------------------------
93 * 0* 0~5 0 [0,63] 64
94 * 1* 6 0 [64,127] 64
95 * 2 7 1 [128,255] 64
96 * 3 8 2 [256,511] 64
97 * 4 9 3 [512,1023] 64
98 * ... ... ... [...,...] ...
99 * 18 23 17 [8838608,+inf]** 64
100 *
101 * * Special cases: when n < (M-1) or when n == (M-1), in both cases,
102 * the value cannot be rounded off. Use all bits of the sample as
103 * index.
104 *
105 * ** If a sample's MSB is greater than 23, it will be counted as 23.
106 */
107
108#define FIO_IO_U_PLAT_BITS 6
109#define FIO_IO_U_PLAT_VAL (1 << FIO_IO_U_PLAT_BITS)
110#define FIO_IO_U_PLAT_GROUP_NR 19
111#define FIO_IO_U_PLAT_NR (FIO_IO_U_PLAT_GROUP_NR * FIO_IO_U_PLAT_VAL)
112#define FIO_IO_U_LIST_MAX_LEN 20 /* The size of the default and user-specified
113 list of percentiles */
114
115#define MAX_PATTERN_SIZE 512
116#define FIO_JOBNAME_SIZE 128
Jens Axboe4e59d0f2014-03-14 08:41:39 -0600117#define FIO_JOBDESC_SIZE 256
Jens Axboea64e88d2011-10-03 14:20:01 +0200118#define FIO_VERROR_SIZE 128
119
120struct thread_stat {
121 char name[FIO_JOBNAME_SIZE];
122 char verror[FIO_VERROR_SIZE];
Jens Axboeddcc0b62011-10-03 14:45:27 +0200123 uint32_t error;
Jens Axboe2f122b12012-03-15 13:10:19 +0100124 uint32_t thread_number;
Jens Axboeddcc0b62011-10-03 14:45:27 +0200125 uint32_t groupid;
Jens Axboea64e88d2011-10-03 14:20:01 +0200126 uint32_t pid;
Jens Axboe4e59d0f2014-03-14 08:41:39 -0600127 char description[FIO_JOBDESC_SIZE];
Jens Axboea64e88d2011-10-03 14:20:01 +0200128 uint32_t members;
Jens Axboe771e58b2013-01-30 12:56:23 +0100129 uint32_t unified_rw_rep;
Jens Axboea64e88d2011-10-03 14:20:01 +0200130
131 /*
132 * bandwidth and latency stats
133 */
Shaohua Li6eaf09d2012-09-14 08:49:43 +0200134 struct io_stat clat_stat[DDIR_RWDIR_CNT]; /* completion latency */
135 struct io_stat slat_stat[DDIR_RWDIR_CNT]; /* submission latency */
136 struct io_stat lat_stat[DDIR_RWDIR_CNT]; /* total latency */
137 struct io_stat bw_stat[DDIR_RWDIR_CNT]; /* bandwidth stats */
138 struct io_stat iops_stat[DDIR_RWDIR_CNT]; /* IOPS stats */
Jens Axboea64e88d2011-10-03 14:20:01 +0200139
140 /*
141 * fio system usage accounting
142 */
143 uint64_t usr_time;
144 uint64_t sys_time;
145 uint64_t ctx;
146 uint64_t minf, majf;
147
148 /*
149 * IO depth and latency stats
150 */
151 uint64_t clat_percentiles;
Vincent Kang Fu435d1952013-02-06 08:43:40 +0100152 uint64_t percentile_precision;
Jens Axboe802ad4a2011-10-05 09:51:58 +0200153 fio_fp64_t percentile_list[FIO_IO_U_LIST_MAX_LEN];
Jens Axboea64e88d2011-10-03 14:20:01 +0200154
155 uint32_t io_u_map[FIO_IO_U_MAP_NR];
156 uint32_t io_u_submit[FIO_IO_U_MAP_NR];
157 uint32_t io_u_complete[FIO_IO_U_MAP_NR];
158 uint32_t io_u_lat_u[FIO_IO_U_LAT_U_NR];
159 uint32_t io_u_lat_m[FIO_IO_U_LAT_M_NR];
Shaohua Li6eaf09d2012-09-14 08:49:43 +0200160 uint32_t io_u_plat[DDIR_RWDIR_CNT][FIO_IO_U_PLAT_NR];
Jens Axboea64e88d2011-10-03 14:20:01 +0200161 uint64_t total_io_u[3];
162 uint64_t short_io_u[3];
163 uint64_t total_submit;
164 uint64_t total_complete;
165
Shaohua Li6eaf09d2012-09-14 08:49:43 +0200166 uint64_t io_bytes[DDIR_RWDIR_CNT];
167 uint64_t runtime[DDIR_RWDIR_CNT];
Jens Axboea64e88d2011-10-03 14:20:01 +0200168 uint64_t total_run_time;
169
170 /*
171 * IO Error related stats
172 */
173 uint16_t continue_on_error;
174 uint64_t total_err_count;
Jens Axboeddcc0b62011-10-03 14:45:27 +0200175 uint32_t first_error;
Jens Axboea64e88d2011-10-03 14:20:01 +0200176
177 uint32_t kb_base;
Steven Noonanad705bc2013-04-08 15:05:25 -0700178 uint32_t unit_base;
Jens Axboe3e260a42013-12-09 12:38:53 -0700179
180 uint32_t latency_depth;
181 uint64_t latency_target;
182 fio_fp64_t latency_percentile;
183 uint64_t latency_window;
Jens Axboeeb663202014-06-30 08:51:33 -0600184} __attribute__((packed));
Jens Axboea64e88d2011-10-03 14:20:01 +0200185
Jens Axboeb75a3942011-10-03 16:03:43 +0200186struct jobs_eta {
187 uint32_t nr_running;
188 uint32_t nr_ramp;
Jens Axboece8ea6e2014-06-27 15:01:06 -0600189
Jens Axboeb75a3942011-10-03 16:03:43 +0200190 uint32_t nr_pending;
Jens Axboe714e85f2013-04-15 10:21:56 +0200191 uint32_t nr_setting_up;
Jens Axboece8ea6e2014-06-27 15:01:06 -0600192
Jens Axboeb75a3942011-10-03 16:03:43 +0200193 uint32_t files_open;
Jens Axboece8ea6e2014-06-27 15:01:06 -0600194
Jens Axboed79db122012-09-24 08:51:24 +0200195 uint32_t m_rate[DDIR_RWDIR_CNT], t_rate[DDIR_RWDIR_CNT];
196 uint32_t m_iops[DDIR_RWDIR_CNT], t_iops[DDIR_RWDIR_CNT];
Shaohua Li6eaf09d2012-09-14 08:49:43 +0200197 uint32_t rate[DDIR_RWDIR_CNT];
198 uint32_t iops[DDIR_RWDIR_CNT];
Jens Axboeb75a3942011-10-03 16:03:43 +0200199 uint64_t elapsed_sec;
200 uint64_t eta_sec;
Jens Axboeb7f05eb2012-05-11 20:33:02 +0200201 uint32_t is_pow2;
Steven Noonanad705bc2013-04-08 15:05:25 -0700202 uint32_t unit_base;
Jens Axboe1d1f45a2011-10-03 19:44:41 +0200203
204 /*
205 * Network 'copy' of run_str[]
206 */
207 uint32_t nr_threads;
Jens Axboe372aecb2013-01-08 13:42:41 +0100208 uint8_t run_str[];
Jens Axboeeb663202014-06-30 08:51:33 -0600209} __attribute__((packed));
Jens Axboeb75a3942011-10-03 16:03:43 +0200210
Jens Axboe61f6cce2014-06-24 08:40:21 -0600211extern struct jobs_eta *get_jobs_eta(int force, size_t *size);
212
Jens Axboecef91752013-04-26 17:05:57 -0600213extern void stat_init(void);
214extern void stat_exit(void);
215
Castor Fu952b05e2013-10-31 11:00:34 -0600216extern struct json_object * show_thread_status(struct thread_stat *ts, struct group_run_stats *rs);
Jens Axboea64e88d2011-10-03 14:20:01 +0200217extern void show_group_stats(struct group_run_stats *rs);
Jens Axboeaf9c9fb2011-10-09 21:54:10 +0200218extern int calc_thread_status(struct jobs_eta *je, int force);
Jens Axboecf451d12011-10-03 16:48:30 +0200219extern void display_thread_status(struct jobs_eta *je);
Jens Axboe5b9babb2011-10-10 12:14:30 +0200220extern void show_run_stats(void);
Jens Axboe2e627242014-07-25 09:51:56 +0200221extern void __show_run_stats(void);
Jens Axboeb852e7c2012-03-30 10:30:35 +0200222extern void show_running_run_stats(void);
Jens Axboe06464902013-04-24 20:38:54 -0600223extern void check_for_running_stats(void);
Jens Axboe5b9babb2011-10-10 12:14:30 +0200224extern void sum_thread_stats(struct thread_stat *dst, struct thread_stat *src, int nr);
Jens Axboe37f0c1a2011-10-11 14:08:33 +0200225extern void sum_group_stats(struct group_run_stats *dst, struct group_run_stats *src);
226extern void init_thread_stat(struct thread_stat *ts);
227extern void init_group_run_stat(struct group_run_stats *gs);
Jens Axboe3e47bd22012-02-29 13:45:02 +0100228extern void eta_to_str(char *str, unsigned long eta_sec);
Jens Axboeb29ad562012-03-05 13:08:51 +0100229extern int calc_lat(struct io_stat *is, unsigned long *min, unsigned long *max, double *mean, double *dev);
Jens Axboea2697902012-03-05 16:43:49 +0100230extern unsigned int calc_clat_percentiles(unsigned int *io_u_plat, unsigned long nr, fio_fp64_t *plist, unsigned int **output, unsigned int *maxv, unsigned int *minv);
Jens Axboee5bd1342012-03-05 21:38:12 +0100231extern void stat_calc_lat_m(struct thread_stat *ts, double *io_u_lat);
232extern void stat_calc_lat_u(struct thread_stat *ts, double *io_u_lat);
Jens Axboe2e331012012-03-05 22:07:54 +0100233extern void stat_calc_dist(unsigned int *map, unsigned long total, double *io_u_dist);
Jens Axboe6bb58212014-02-21 13:55:31 -0800234extern void reset_io_stats(struct thread_data *);
Jens Axboe2e331012012-03-05 22:07:54 +0100235
Jens Axboeb29ad562012-03-05 13:08:51 +0100236static inline int usec_to_msec(unsigned long *min, unsigned long *max,
237 double *mean, double *dev)
238{
239 if (*min > 1000 && *max > 1000 && *mean > 1000.0 && *dev > 1000.0) {
240 *min /= 1000;
241 *max /= 1000;
242 *mean /= 1000.0;
243 *dev /= 1000.0;
244 return 0;
245 }
246
247 return 1;
248}
Jens Axboe61f6cce2014-06-24 08:40:21 -0600249/*
250 * Worst level condensing would be 1:5, so allow enough room for that
251 */
252#define __THREAD_RUNSTR_SZ(nr) ((nr) * 5)
Jens Axboe18148762014-06-23 19:07:12 -0600253#define THREAD_RUNSTR_SZ __THREAD_RUNSTR_SZ(thread_number)
254
Jens Axboea64e88d2011-10-03 14:20:01 +0200255#endif