aboutsummaryrefslogtreecommitdiff
path: root/Documentation/iostats.txt
blob: 04d394a2e06ceff45ca2b6616dc27503d9d92221 (plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
=====================
I/O statistics fields
=====================

Since 2.4.20 (and some versions before, with patches), and 2.5.45,
more extensive disk statistics have been introduced to help measure disk
activity. Tools such as ``sar`` and ``iostat`` typically interpret these and do
the work for you, but in case you are interested in creating your own
tools, the fields are explained here.

In 2.4 now, the information is found as additional fields in
``/proc/partitions``.  In 2.6 and upper, the same information is found in two
places: one is in the file ``/proc/diskstats``, and the other is within
the sysfs file system, which must be mounted in order to obtain
the information. Throughout this document we'll assume that sysfs
is mounted on ``/sys``, although of course it may be mounted anywhere.
Both ``/proc/diskstats`` and sysfs use the same source for the information
and so should not differ.

Here are examples of these different formats::

   2.4:
      3     0   39082680 hda 446216 784926 9550688 4382310 424847 312726 5922052 19310380 0 3376340 23705160
      3     1    9221278 hda1 35486 0 35496 38030 0 0 0 0 0 38030 38030

   2.6+ sysfs:
      446216 784926 9550688 4382310 424847 312726 5922052 19310380 0 3376340 23705160
      35486    38030    38030    38030

   2.6+ diskstats:
      3    0   hda 446216 784926 9550688 4382310 424847 312726 5922052 19310380 0 3376340 23705160
      3    1   hda1 35486 38030 38030 38030

On 2.4 you might execute ``grep 'hda ' /proc/partitions``. On 2.6+, you have
a choice of ``cat /sys/block/hda/stat`` or ``grep 'hda ' /proc/diskstats``.

The advantage of one over the other is that the sysfs choice works well
if you are watching a known, small set of disks.  ``/proc/diskstats`` may
be a better choice if you are watching a large number of disks because
you'll avoid the overhead of 50, 100, or 500 or more opens/closes with
each snapshot of your disk statistics.

In 2.4, the statistics fields are those after the device name. In
the above example, the first field of statistics would be 446216.
By contrast, in 2.6+ if you look at ``/sys/block/hda/stat``, you'll
find just the eleven fields, beginning with 446216.  If you look at
``/proc/diskstats``, the eleven fields will be preceded by the major and
minor device numbers, and device name.  Each of these formats provides
eleven fields of statistics, each meaning exactly the same things.
All fields except field 9 are cumulative since boot.  Field 9 should
go to zero as I/Os complete; all others only increase (unless they
overflow and wrap).  Yes, these are (32-bit or 64-bit) unsigned long
(native word size) numbers, and on a very busy or long-lived system they
may wrap. Applications should be prepared to deal with that; unless
your observations are measured in large numbers of minutes or hours,
they should not wrap twice before you notice them.

Each set of stats only applies to the indicated device; if you want
system-wide stats you'll have to find all the devices and sum them all up.

Field  1 -- # of reads completed
    This is the total number of reads completed successfully.

Field  2 -- # of reads merged, field 6 -- # of writes merged
    Reads and writes which are adjacent to each other may be merged for
    efficiency.  Thus two 4K reads may become one 8K read before it is
    ultimately handed to the disk, and so it will be counted (and queued)
    as only one I/O.  This field lets you know how often this was done.

Field  3 -- # of sectors read
    This is the total number of sectors read successfully.

Field  4 -- # of milliseconds spent reading
    This is the total number of milliseconds spent by all reads (as
    measured from __make_request() to end_that_request_last()).

Field  5 -- # of writes completed
    This is the total number of writes completed successfully.

Field  6 -- # of writes merged
    See the description of field 2.

Field  7 -- # of sectors written
    This is the total number of sectors written successfully.

Field  8 -- # of milliseconds spent writing
    This is the total number of milliseconds spent by all writes (as
    measured from __make_request() to end_that_request_last()).

Field  9 -- # of I/Os currently in progress
    The only field that should go to zero. Incremented as requests are
    given to appropriate struct request_queue and decremented as they finish.

Field 10 -- # of milliseconds spent doing I/Os
    This field increases so long as field 9 is nonzero.

Field 11 -- weighted # of milliseconds spent doing I/Os
    This field is incremented at each I/O start, I/O completion, I/O
    merge, or read of these stats by the number of I/Os in progress
    (field 9) times the number of milliseconds spent doing I/O since the
    last update of this field.  This can provide an easy measure of both
    I/O completion time and the backlog that may be accumulating.


To avoid introducing performance bottlenecks, no locks are held while
modifying these counters.  This implies that minor inaccuracies may be
introduced when changes collide, so (for instance) adding up all the
read I/Os issued per partition should equal those made to the disks ...
but due to the lack of locking it may only be very close.

In 2.6+, there are counters for each CPU, which make the lack of locking
almost a non-issue.  When the statistics are read, the per-CPU counters
are summed (possibly overflowing the unsigned long variable they are
summed to) and the result given to the user.  There is no convenient
user interface for accessing the per-CPU counters themselves.

Disks vs Partitions
-------------------

There were significant changes between 2.4 and 2.6+ in the I/O subsystem.
As a result, some statistic information disappeared. The translation from
a disk address relative to a partition to the disk address relative to
the host disk happens much earlier.  All merges and timings now happen
at the disk level rather than at both the disk and partition level as
in 2.4.  Consequently, you'll see a different statistics output on 2.6+ for
partitions from that for disks.  There are only *four* fields available
for partitions on 2.6+ machines.  This is reflected in the examples above.

Field  1 -- # of reads issued
    This is the total number of reads issued to this partition.

Field  2 -- # of sectors read
    This is the total number of sectors requested to be read from this
    partition.

Field  3 -- # of writes issued
    This is the total number of writes issued to this partition.

Field  4 -- # of sectors written
    This is the total number of sectors requested to be written to
    this partition.

Note that since the address is translated to a disk-relative one, and no
record of the partition-relative address is kept, the subsequent success
or failure of the read cannot be attributed to the partition.  In other
words, the number of reads for partitions is counted slightly before time
of queuing for partitions, and at completion for whole disks.  This is
a subtle distinction that is probably uninteresting for most cases.

More significant is the error induced by counting the numbers of
reads/writes before merges for partitions and after for disks. Since a
typical workload usually contains a lot of successive and adjacent requests,
the number of reads/writes issued can be several times higher than the
number of reads/writes completed.

In 2.6.25, the full statistic set is again available for partitions and
disk and partition statistics are consistent again. Since we still don't
keep record of the partition-relative address, an operation is attributed to
the partition which contains the first sector of the request after the
eventual merges. As requests can be merged across partition, this could lead
to some (probably insignificant) inaccuracy.

Additional notes
----------------

In 2.6+, sysfs is not mounted by default.  If your distribution of
Linux hasn't added it already, here's the line you'll want to add to
your ``/etc/fstab``::

	none /sys sysfs defaults 0 0


In 2.6+, all disk statistics were removed from ``/proc/stat``.  In 2.4, they
appear in both ``/proc/partitions`` and ``/proc/stat``, although the ones in
``/proc/stat`` take a very different format from those in ``/proc/partitions``
(see proc(5), if your system has it.)

-- ricklind@us.ibm.com