blob: 3115539fdb183ad57240f1b2ea8fe9cf27ba526e [file] [log] [blame]
Jens Axboea64e88d2011-10-03 14:20:01 +02001#ifndef FIO_STAT_H
2#define FIO_STAT_H
3
4struct group_run_stats {
5 uint64_t max_run[2], min_run[2];
6 uint64_t max_bw[2], min_bw[2];
7 uint64_t io_kb[2];
8 uint64_t agg[2];
9 uint32_t kb_base;
10 uint32_t groupid;
11};
12
13/*
14 * How many depth levels to log
15 */
16#define FIO_IO_U_MAP_NR 7
17#define FIO_IO_U_LAT_U_NR 10
18#define FIO_IO_U_LAT_M_NR 12
19
20/*
21 * Aggregate clat samples to report percentile(s) of them.
22 *
23 * EXECUTIVE SUMMARY
24 *
25 * FIO_IO_U_PLAT_BITS determines the maximum statistical error on the
26 * value of resulting percentiles. The error will be approximately
27 * 1/2^(FIO_IO_U_PLAT_BITS+1) of the value.
28 *
29 * FIO_IO_U_PLAT_GROUP_NR and FIO_IO_U_PLAT_BITS determine the maximum
30 * range being tracked for latency samples. The maximum value tracked
31 * accurately will be 2^(GROUP_NR + PLAT_BITS -1) microseconds.
32 *
33 * FIO_IO_U_PLAT_GROUP_NR and FIO_IO_U_PLAT_BITS determine the memory
34 * requirement of storing those aggregate counts. The memory used will
35 * be (FIO_IO_U_PLAT_GROUP_NR * 2^FIO_IO_U_PLAT_BITS) * sizeof(int)
36 * bytes.
37 *
38 * FIO_IO_U_PLAT_NR is the total number of buckets.
39 *
40 * DETAILS
41 *
42 * Suppose the clat varies from 0 to 999 (usec), the straightforward
43 * method is to keep an array of (999 + 1) buckets, in which a counter
44 * keeps the count of samples which fall in the bucket, e.g.,
45 * {[0],[1],...,[999]}. However this consumes a huge amount of space,
46 * and can be avoided if an approximation is acceptable.
47 *
48 * One such method is to let the range of the bucket to be greater
49 * than one. This method has low accuracy when the value is small. For
50 * example, let the buckets be {[0,99],[100,199],...,[900,999]}, and
51 * the represented value of each bucket be the mean of the range. Then
52 * a value 0 has an round-off error of 49.5. To improve on this, we
53 * use buckets with non-uniform ranges, while bounding the error of
54 * each bucket within a ratio of the sample value. A simple example
55 * would be when error_bound = 0.005, buckets are {
56 * {[0],[1],...,[99]}, {[100,101],[102,103],...,[198,199]},..,
57 * {[900,909],[910,919]...} }. The total range is partitioned into
58 * groups with different ranges, then buckets with uniform ranges. An
59 * upper bound of the error is (range_of_bucket/2)/value_of_bucket
60 *
61 * For better efficiency, we implement this using base two. We group
62 * samples by their Most Significant Bit (MSB), extract the next M bit
63 * of them as an index within the group, and discard the rest of the
64 * bits.
65 *
66 * E.g., assume a sample 'x' whose MSB is bit n (starting from bit 0),
67 * and use M bit for indexing
68 *
69 * | n | M bits | bit (n-M-1) ... bit 0 |
70 *
71 * Because x is at least 2^n, and bit 0 to bit (n-M-1) is at most
72 * (2^(n-M) - 1), discarding bit 0 to (n-M-1) makes the round-off
73 * error
74 *
75 * 2^(n-M)-1 2^(n-M) 1
76 * e <= --------- <= ------- = ---
77 * 2^n 2^n 2^M
78 *
79 * Furthermore, we use "mean" of the range to represent the bucket,
80 * the error e can be lowered by half to 1 / 2^(M+1). By using M bits
81 * as the index, each group must contains 2^M buckets.
82 *
83 * E.g. Let M (FIO_IO_U_PLAT_BITS) be 6
84 * Error bound is 1/2^(6+1) = 0.0078125 (< 1%)
85 *
86 * Group MSB #discarded range of #buckets
87 * error_bits value
88 * ----------------------------------------------------------------
89 * 0* 0~5 0 [0,63] 64
90 * 1* 6 0 [64,127] 64
91 * 2 7 1 [128,255] 64
92 * 3 8 2 [256,511] 64
93 * 4 9 3 [512,1023] 64
94 * ... ... ... [...,...] ...
95 * 18 23 17 [8838608,+inf]** 64
96 *
97 * * Special cases: when n < (M-1) or when n == (M-1), in both cases,
98 * the value cannot be rounded off. Use all bits of the sample as
99 * index.
100 *
101 * ** If a sample's MSB is greater than 23, it will be counted as 23.
102 */
103
104#define FIO_IO_U_PLAT_BITS 6
105#define FIO_IO_U_PLAT_VAL (1 << FIO_IO_U_PLAT_BITS)
106#define FIO_IO_U_PLAT_GROUP_NR 19
107#define FIO_IO_U_PLAT_NR (FIO_IO_U_PLAT_GROUP_NR * FIO_IO_U_PLAT_VAL)
108#define FIO_IO_U_LIST_MAX_LEN 20 /* The size of the default and user-specified
109 list of percentiles */
110
111#define MAX_PATTERN_SIZE 512
112#define FIO_JOBNAME_SIZE 128
113#define FIO_VERROR_SIZE 128
114
115struct thread_stat {
116 char name[FIO_JOBNAME_SIZE];
117 char verror[FIO_VERROR_SIZE];
Jens Axboeddcc0b62011-10-03 14:45:27 +0200118 uint32_t error;
119 uint32_t groupid;
Jens Axboea64e88d2011-10-03 14:20:01 +0200120 uint32_t pid;
121 char description[FIO_JOBNAME_SIZE];
122 uint32_t members;
123
124 /*
125 * bandwidth and latency stats
126 */
127 struct io_stat clat_stat[2]; /* completion latency */
128 struct io_stat slat_stat[2]; /* submission latency */
129 struct io_stat lat_stat[2]; /* total latency */
130 struct io_stat bw_stat[2]; /* bandwidth stats */
Jens Axboec8eeb9d2011-10-05 14:02:22 +0200131 struct io_stat iops_stat[2]; /* IOPS stats */
Jens Axboea64e88d2011-10-03 14:20:01 +0200132
133 /*
134 * fio system usage accounting
135 */
136 uint64_t usr_time;
137 uint64_t sys_time;
138 uint64_t ctx;
139 uint64_t minf, majf;
140
141 /*
142 * IO depth and latency stats
143 */
144 uint64_t clat_percentiles;
Jens Axboe802ad4a2011-10-05 09:51:58 +0200145 fio_fp64_t percentile_list[FIO_IO_U_LIST_MAX_LEN];
Jens Axboea64e88d2011-10-03 14:20:01 +0200146
147 uint32_t io_u_map[FIO_IO_U_MAP_NR];
148 uint32_t io_u_submit[FIO_IO_U_MAP_NR];
149 uint32_t io_u_complete[FIO_IO_U_MAP_NR];
150 uint32_t io_u_lat_u[FIO_IO_U_LAT_U_NR];
151 uint32_t io_u_lat_m[FIO_IO_U_LAT_M_NR];
152 uint32_t io_u_plat[2][FIO_IO_U_PLAT_NR];
153 uint64_t total_io_u[3];
154 uint64_t short_io_u[3];
155 uint64_t total_submit;
156 uint64_t total_complete;
157
158 uint64_t io_bytes[2];
159 uint64_t runtime[2];
160 uint64_t total_run_time;
161
162 /*
163 * IO Error related stats
164 */
165 uint16_t continue_on_error;
166 uint64_t total_err_count;
Jens Axboeddcc0b62011-10-03 14:45:27 +0200167 uint32_t first_error;
Jens Axboea64e88d2011-10-03 14:20:01 +0200168
169 uint32_t kb_base;
170};
171
Jens Axboeb75a3942011-10-03 16:03:43 +0200172struct jobs_eta {
173 uint32_t nr_running;
174 uint32_t nr_ramp;
175 uint32_t nr_pending;
176 uint32_t files_open;
177 uint32_t m_rate, t_rate;
178 uint32_t m_iops, t_iops;
179 uint32_t rate[2];
180 uint32_t iops[2];
181 uint64_t elapsed_sec;
182 uint64_t eta_sec;
Jens Axboe1d1f45a2011-10-03 19:44:41 +0200183
184 /*
185 * Network 'copy' of run_str[]
186 */
187 uint32_t nr_threads;
188 uint8_t run_str[0];
Jens Axboeb75a3942011-10-03 16:03:43 +0200189};
190
Jens Axboea64e88d2011-10-03 14:20:01 +0200191extern void show_thread_status(struct thread_stat *ts, struct group_run_stats *rs);
192extern void show_group_stats(struct group_run_stats *rs);
Jens Axboeaf9c9fb2011-10-09 21:54:10 +0200193extern int calc_thread_status(struct jobs_eta *je, int force);
Jens Axboecf451d12011-10-03 16:48:30 +0200194extern void display_thread_status(struct jobs_eta *je);
Jens Axboe5b9babb2011-10-10 12:14:30 +0200195extern void show_run_stats(void);
196extern void sum_thread_stats(struct thread_stat *dst, struct thread_stat *src, int nr);
Jens Axboe37f0c1a2011-10-11 14:08:33 +0200197extern void sum_group_stats(struct group_run_stats *dst, struct group_run_stats *src);
198extern void init_thread_stat(struct thread_stat *ts);
199extern void init_group_run_stat(struct group_run_stats *gs);
Jens Axboea64e88d2011-10-03 14:20:01 +0200200
201#endif