1 Anticipatory IO scheduler
2 -------------------------
3 Nick Piggin <piggin@cyberone.com.au> 13 Sep 2003
5 Attention! Database servers, especially those using "TCQ" disks should
6 investigate performance with the 'deadline' IO scheduler. Any system with high
7 disk performance requirements should do so, in fact.
9 If you see unusual performance characteristics of your disk systems, or you
10 see big performance regressions versus the deadline scheduler, please email
11 me. Database users don't bother unless you're willing to test a lot of patches
12 from me ;) its a known issue.
14 Also, users with hardware RAID controllers, doing striping, may find
15 highly variable performance results with using the as-iosched. The
16 as-iosched anticipatory implementation is based on the notion that a disk
17 device has only one physical seeking head. A striped RAID controller
18 actually has a head for each physical device in the logical RAID device.
20 However, setting the antic_expire (see tunable parameters below) produces
21 very similar behavior to the deadline IO scheduler.
24 Selecting IO schedulers
25 -----------------------
26 To choose IO schedulers at boot time, use the argument 'elevator=deadline'.
27 'noop', 'as' and 'cfq' (the default) are also available. IO schedulers are
28 assigned globally at boot time only presently. It's also possible to change
29 the IO scheduler for a determined device on the fly, as described in
30 Documentation/block/switching-sched.txt.
33 Anticipatory IO scheduler Policies
34 ----------------------------------
35 The as-iosched implementation implements several layers of policies
36 to determine when an IO request is dispatched to the disk controller.
37 Here are the policies outlined, in order of application.
39 1. one-way Elevator algorithm.
41 The elevator algorithm is similar to that used in deadline scheduler, with
42 the addition that it allows limited backward movement of the elevator
43 (i.e. seeks backwards). A seek backwards can occur when choosing between
44 two IO requests where one is behind the elevator's current position, and
45 the other is in front of the elevator's position. If the seek distance to
46 the request in back of the elevator is less than half the seek distance to
47 the request in front of the elevator, then the request in back can be chosen.
48 Backward seeks are also limited to a maximum of MAXBACK (1024*1024) sectors.
49 This favors forward movement of the elevator, while allowing opportunistic
50 "short" backward seeks.
52 2. FIFO expiration times for reads and for writes.
54 This is again very similar to the deadline IO scheduler. The expiration
55 times for requests on these lists is tunable using the parameters read_expire
56 and write_expire discussed below. When a read or a write expires in this way,
57 the IO scheduler will interrupt its current elevator sweep or read anticipation
58 to service the expired request.
60 3. Read and write request batching
62 A batch is a collection of read requests or a collection of write
63 requests. The as scheduler alternates dispatching read and write batches
64 to the driver. In the case a read batch, the scheduler submits read
65 requests to the driver as long as there are read requests to submit, and
66 the read batch time limit has not been exceeded (read_batch_expire).
67 The read batch time limit begins counting down only when there are
68 competing write requests pending.
70 In the case of a write batch, the scheduler submits write requests to
71 the driver as long as there are write requests available, and the
72 write batch time limit has not been exceeded (write_batch_expire).
73 However, the length of write batches will be gradually shortened
74 when read batches frequently exceed their time limit.
76 When changing between batch types, the scheduler waits for all requests
77 from the previous batch to complete before scheduling requests for the
80 The read and write fifo expiration times described in policy 2 above
81 are checked only when in scheduling IO of a batch for the corresponding
82 (read/write) type. So for example, the read FIFO timeout values are
83 tested only during read batches. Likewise, the write FIFO timeout
84 values are tested only during write batches. For this reason,
85 it is generally not recommended for the read batch time
86 to be longer than the write expiration time, nor for the write batch
87 time to exceed the read expiration time (see tunable parameters below).
89 When the IO scheduler changes from a read to a write batch,
90 it begins the elevator from the request that is on the head of the
91 write expiration FIFO. Likewise, when changing from a write batch to
92 a read batch, scheduler begins the elevator from the first entry
93 on the read expiration FIFO.
97 Read anticipation occurs only when scheduling a read batch.
98 This implementation of read anticipation allows only one read request
99 to be dispatched to the disk controller at a time. In
100 contrast, many write requests may be dispatched to the disk controller
101 at a time during a write batch. It is this characteristic that can make
102 the anticipatory scheduler perform anomalously with controllers supporting
103 TCQ, or with hardware striped RAID devices. Setting the antic_expire
104 queue parameter (see below) to zero disables this behavior, and the
105 anticipatory scheduler behaves essentially like the deadline scheduler.
107 When read anticipation is enabled (antic_expire is not zero), reads
108 are dispatched to the disk controller one at a time.
109 At the end of each read request, the IO scheduler examines its next
110 candidate read request from its sorted read list. If that next request
111 is from the same process as the request that just completed,
112 or if the next request in the queue is "very close" to the
113 just completed request, it is dispatched immediately. Otherwise,
114 statistics (average think time, average seek distance) on the process
115 that submitted the just completed request are examined. If it seems
116 likely that that process will submit another request soon, and that
117 request is likely to be near the just completed request, then the IO
118 scheduler will stop dispatching more read requests for up time (antic_expire)
119 milliseconds, hoping that process will submit a new request near the one
120 that just completed. If such a request is made, then it is dispatched
121 immediately. If the antic_expire wait time expires, then the IO scheduler
122 will dispatch the next read request from the sorted read queue.
124 To decide whether an anticipatory wait is worthwhile, the scheduler
125 maintains statistics for each process that can be used to compute
126 mean "think time" (the time between read requests), and mean seek
127 distance for that process. One observation is that these statistics
128 are associated with each process, but those statistics are not associated
129 with a specific IO device. So for example, if a process is doing IO
130 on several file systems on separate devices, the statistics will be
131 a combination of IO behavior from all those devices.
134 Tuning the anticipatory IO scheduler
135 ------------------------------------
136 When using 'as', the anticipatory IO scheduler there are 5 parameters under
137 /sys/block/*/queue/iosched/. All are units of milliseconds.
141 Controls how long until a read request becomes "expired". It also controls the
142 interval between which expired requests are served, so set to 50, a request
143 might take anywhere < 100ms to be serviced _if_ it is the next on the
144 expired list. Obviously request expiration strategies won't make the disk
145 go faster. The result basically equates to the timeslice a single reader
146 gets in the presence of other IO. 100*((seek time / read_expire) + 1) is
147 very roughly the % streaming read efficiency your disk should get with
151 Controls how much time a batch of reads is given before pending writes are
152 served. A higher value is more efficient. This might be set below read_expire
153 if writes are to be given higher priority than reads, but reads are to be
154 as efficient as possible when there are no writes. Generally though, it
155 should be some multiple of read_expire.
158 * write_batch_expire are equivalent to the above, for writes.
161 Controls the maximum amount of time we can anticipate a good read (one
162 with a short seek distance from the most recently completed request) before
163 giving up. Many other factors may cause anticipation to be stopped early,
164 or some processes will not be "anticipated" at all. Should be a bit higher
165 for big seek time devices though not a linear correspondence - most
166 processes have only a few ms thinktime.