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.
23 Selecting IO schedulers
24 -----------------------
25 Refer to Documentation/block/switching-sched.txt for information on
26 selecting an io scheduler on a per-device basis.
28 Anticipatory IO scheduler Policies
29 ----------------------------------
30 The as-iosched implementation implements several layers of policies
31 to determine when an IO request is dispatched to the disk controller.
32 Here are the policies outlined, in order of application.
34 1. one-way Elevator algorithm.
36 The elevator algorithm is similar to that used in deadline scheduler, with
37 the addition that it allows limited backward movement of the elevator
38 (i.e. seeks backwards). A seek backwards can occur when choosing between
39 two IO requests where one is behind the elevator's current position, and
40 the other is in front of the elevator's position. If the seek distance to
41 the request in back of the elevator is less than half the seek distance to
42 the request in front of the elevator, then the request in back can be chosen.
43 Backward seeks are also limited to a maximum of MAXBACK (1024*1024) sectors.
44 This favors forward movement of the elevator, while allowing opportunistic
45 "short" backward seeks.
47 2. FIFO expiration times for reads and for writes.
49 This is again very similar to the deadline IO scheduler. The expiration
50 times for requests on these lists is tunable using the parameters read_expire
51 and write_expire discussed below. When a read or a write expires in this way,
52 the IO scheduler will interrupt its current elevator sweep or read anticipation
53 to service the expired request.
55 3. Read and write request batching
57 A batch is a collection of read requests or a collection of write
58 requests. The as scheduler alternates dispatching read and write batches
59 to the driver. In the case a read batch, the scheduler submits read
60 requests to the driver as long as there are read requests to submit, and
61 the read batch time limit has not been exceeded (read_batch_expire).
62 The read batch time limit begins counting down only when there are
63 competing write requests pending.
65 In the case of a write batch, the scheduler submits write requests to
66 the driver as long as there are write requests available, and the
67 write batch time limit has not been exceeded (write_batch_expire).
68 However, the length of write batches will be gradually shortened
69 when read batches frequently exceed their time limit.
71 When changing between batch types, the scheduler waits for all requests
72 from the previous batch to complete before scheduling requests for the
75 The read and write fifo expiration times described in policy 2 above
76 are checked only when in scheduling IO of a batch for the corresponding
77 (read/write) type. So for example, the read FIFO timeout values are
78 tested only during read batches. Likewise, the write FIFO timeout
79 values are tested only during write batches. For this reason,
80 it is generally not recommended for the read batch time
81 to be longer than the write expiration time, nor for the write batch
82 time to exceed the read expiration time (see tunable parameters below).
84 When the IO scheduler changes from a read to a write batch,
85 it begins the elevator from the request that is on the head of the
86 write expiration FIFO. Likewise, when changing from a write batch to
87 a read batch, scheduler begins the elevator from the first entry
88 on the read expiration FIFO.
92 Read anticipation occurs only when scheduling a read batch.
93 This implementation of read anticipation allows only one read request
94 to be dispatched to the disk controller at a time. In
95 contrast, many write requests may be dispatched to the disk controller
96 at a time during a write batch. It is this characteristic that can make
97 the anticipatory scheduler perform anomalously with controllers supporting
98 TCQ, or with hardware striped RAID devices. Setting the antic_expire
99 queue parameter (see below) to zero disables this behavior, and the
100 anticipatory scheduler behaves essentially like the deadline scheduler.
102 When read anticipation is enabled (antic_expire is not zero), reads
103 are dispatched to the disk controller one at a time.
104 At the end of each read request, the IO scheduler examines its next
105 candidate read request from its sorted read list. If that next request
106 is from the same process as the request that just completed,
107 or if the next request in the queue is "very close" to the
108 just completed request, it is dispatched immediately. Otherwise,
109 statistics (average think time, average seek distance) on the process
110 that submitted the just completed request are examined. If it seems
111 likely that that process will submit another request soon, and that
112 request is likely to be near the just completed request, then the IO
113 scheduler will stop dispatching more read requests for up to (antic_expire)
114 milliseconds, hoping that process will submit a new request near the one
115 that just completed. If such a request is made, then it is dispatched
116 immediately. If the antic_expire wait time expires, then the IO scheduler
117 will dispatch the next read request from the sorted read queue.
119 To decide whether an anticipatory wait is worthwhile, the scheduler
120 maintains statistics for each process that can be used to compute
121 mean "think time" (the time between read requests), and mean seek
122 distance for that process. One observation is that these statistics
123 are associated with each process, but those statistics are not associated
124 with a specific IO device. So for example, if a process is doing IO
125 on several file systems on separate devices, the statistics will be
126 a combination of IO behavior from all those devices.
129 Tuning the anticipatory IO scheduler
130 ------------------------------------
131 When using 'as', the anticipatory IO scheduler there are 5 parameters under
132 /sys/block/*/queue/iosched/. All are units of milliseconds.
136 Controls how long until a read request becomes "expired". It also controls the
137 interval between which expired requests are served, so set to 50, a request
138 might take anywhere < 100ms to be serviced _if_ it is the next on the
139 expired list. Obviously request expiration strategies won't make the disk
140 go faster. The result basically equates to the timeslice a single reader
141 gets in the presence of other IO. 100*((seek time / read_expire) + 1) is
142 very roughly the % streaming read efficiency your disk should get with
146 Controls how much time a batch of reads is given before pending writes are
147 served. A higher value is more efficient. This might be set below read_expire
148 if writes are to be given higher priority than reads, but reads are to be
149 as efficient as possible when there are no writes. Generally though, it
150 should be some multiple of read_expire.
153 * write_batch_expire are equivalent to the above, for writes.
156 Controls the maximum amount of time we can anticipate a good read (one
157 with a short seek distance from the most recently completed request) before
158 giving up. Many other factors may cause anticipation to be stopped early,
159 or some processes will not be "anticipated" at all. Should be a bit higher
160 for big seek time devices though not a linear correspondence - most
161 processes have only a few ms thinktime.
163 In addition to the tunables above there is a read-only file named est_time
164 which, when read, will show:
166 - The probability of a task exiting without a cooperating task
167 submitting an anticipated IO.
169 - The current mean think time.
171 - The seek distance used to determine if an incoming IO is better.