Documentation/block/cfq-iosched.txt

   1 CFQ (Complete Fairness Queueing)
   2 ===============================
   3
   4 The main aim of CFQ scheduler is to provide a fair allocation of the disk
   5 I/O bandwidth for all the processes which requests an I/O operation.
   6
   7 CFQ maintains the per process queue for the processes which request I/O
   8 operation(synchronous requests). In case of asynchronous requests, all the
   9 requests from all the processes are batched together according to their
  10 process's I/O priority.
  11
  12 CFQ ioscheduler tunables
  13 ========================
  14
  15 slice_idle
  16 ----------
  17 This specifies how long CFQ should idle for next request on certain cfq queues
  18 (for sequential workloads) and service trees (for random workloads) before
  19 queue is expired and CFQ selects next queue to dispatch from.
  20
  21 By default slice_idle is a non-zero value. That means by default we idle on
  22 queues/service trees. This can be very helpful on highly seeky media like
  23 single spindle SATA/SAS disks where we can cut down on overall number of
  24 seeks and see improved throughput.
  25
  26 Setting slice_idle to 0 will remove all the idling on queues/service tree
  27 level and one should see an overall improved throughput on faster storage
  28 devices like multiple SATA/SAS disks in hardware RAID configuration. The down
  29 side is that isolation provided from WRITES also goes down and notion of
  30 IO priority becomes weaker.
  31
  32 So depending on storage and workload, it might be useful to set slice_idle=0.
  33 In general I think for SATA/SAS disks and software RAID of SATA/SAS disks
  34 keeping slice_idle enabled should be useful. For any configurations where
  35 there are multiple spindles behind single LUN (Host based hardware RAID
  36 controller or for storage arrays), setting slice_idle=0 might end up in better
  37 throughput and acceptable latencies.
  38
  39 back_seek_max
  40 -------------
  41 This specifies, given in Kbytes, the maximum "distance" for backward seeking.
  42 The distance is the amount of space from the current head location to the
  43 sectors that are backward in terms of distance.
  44
  45 This parameter allows the scheduler to anticipate requests in the "backward"
  46 direction and consider them as being the "next" if they are within this
  47 distance from the current head location.
  48
  49 back_seek_penalty
  50 -----------------
  51 This parameter is used to compute the cost of backward seeking. If the
  52 backward distance of request is just 1/back_seek_penalty from a "front"
  53 request, then the seeking cost of two requests is considered equivalent.
  54
  55 So scheduler will not bias toward one or the other request (otherwise scheduler
  56 will bias toward front request). Default value of back_seek_penalty is 2.
  57
  58 fifo_expire_async
  59 -----------------
  60 This parameter is used to set the timeout of asynchronous requests. Default
  61 value of this is 248ms.
  62
  63 fifo_expire_sync
  64 ----------------
  65 This parameter is used to set the timeout of synchronous requests. Default
  66 value of this is 124ms. In case to favor synchronous requests over asynchronous
  67 one, this value should be decreased relative to fifo_expire_async.
  68
  69 group_idle
  70 -----------
  71 This parameter forces idling at the CFQ group level instead of CFQ
  72 queue level. This was introduced after after a bottleneck was observed
  73 in higher end storage due to idle on sequential queue and allow dispatch
  74 from a single queue. The idea with this parameter is that it can be run with
  75 slice_idle=0 and group_idle=8, so that idling does not happen on individual
  76 queues in the group but happens overall on the group and thus still keeps the
  77 IO controller working.
  78 Not idling on individual queues in the group will dispatch requests from
  79 multiple queues in the group at the same time and achieve higher throughput
  80 on higher end storage.
  81
  82 Default value for this parameter is 8ms.
  83
  84 latency
  85 -------
  86 This parameter is used to enable/disable the latency mode of the CFQ
  87 scheduler. If latency mode (called low_latency) is enabled, CFQ tries
  88 to recompute the slice time for each process based on the target_latency set
  89 for the system. This favors fairness over throughput. Disabling low
  90 latency (setting it to 0) ignores target latency, allowing each process in the
  91 system to get a full time slice.
  92
  93 By default low latency mode is enabled.
  94
  95 target_latency
  96 --------------
  97 This parameter is used to calculate the time slice for a process if cfq's
  98 latency mode is enabled. It will ensure that sync requests have an estimated
  99 latency. But if sequential workload is higher(e.g. sequential read),
 100 then to meet the latency constraints, throughput may decrease because of less
 101 time for each process to issue I/O request before the cfq queue is switched.
 102
 103 Though this can be overcome by disabling the latency_mode, it may increase
 104 the read latency for some applications. This parameter allows for changing
 105 target_latency through the sysfs interface which can provide the balanced
 106 throughput and read latency.
 107
 108 Default value for target_latency is 300ms.
 109
 110 slice_async
 111 -----------
 112 This parameter is same as of slice_sync but for asynchronous queue. The
 113 default value is 40ms.
 114
 115 slice_async_rq
 116 --------------
 117 This parameter is used to limit the dispatching of asynchronous request to
 118 device request queue in queue's slice time. The maximum number of request that
 119 are allowed to be dispatched also depends upon the io priority. Default value
 120 for this is 2.
 121
 122 slice_sync
 123 ----------
 124 When a queue is selected for execution, the queues IO requests are only
 125 executed for a certain amount of time(time_slice) before switching to another
 126 queue. This parameter is used to calculate the time slice of synchronous
 127 queue.
 128
 129 time_slice is computed using the below equation:-
 130 time_slice = slice_sync + (slice_sync/5 * (4 - prio)). To increase the
 131 time_slice of synchronous queue, increase the value of slice_sync. Default
 132 value is 100ms.
 133
 134 quantum
 135 -------
 136 This specifies the number of request dispatched to the device queue. In a
 137 queue's time slice, a request will not be dispatched if the number of request
 138 in the device exceeds this parameter. This parameter is used for synchronous
 139 request.
 140
 141 In case of storage with several disk, this setting can limit the parallel
 142 processing of request. Therefore, increasing the value can improve the
 143 performance although this can cause the latency of some I/O to increase due
 144 to more number of requests.
 145
 146 CFQ Group scheduling
 147 ====================
 148
 149 CFQ supports blkio cgroup and has "blkio." prefixed files in each
 150 blkio cgroup directory. It is weight-based and there are four knobs
 151 for configuration - weight[_device] and leaf_weight[_device].
 152 Internal cgroup nodes (the ones with children) can also have tasks in
 153 them, so the former two configure how much proportion the cgroup as a
 154 whole is entitled to at its parent's level while the latter two
 155 configure how much proportion the tasks in the cgroup have compared to
 156 its direct children.
 157
 158 Another way to think about it is assuming that each internal node has
 159 an implicit leaf child node which hosts all the tasks whose weight is
 160 configured by leaf_weight[_device]. Let's assume a blkio hierarchy
 161 composed of five cgroups - root, A, B, AA and AB - with the following
 162 weights where the names represent the hierarchy.
 163
 164         weight leaf_weight
 165  root :  125    125
 166  A    :  500    750
 167  B    :  250    500
 168  AA   :  500    500
 169  AB   : 1000    500
 170
 171 root never has a parent making its weight is meaningless. For backward
 172 compatibility, weight is always kept in sync with leaf_weight. B, AA
 173 and AB have no child and thus its tasks have no children cgroup to
 174 compete with. They always get 100% of what the cgroup won at the
 175 parent level. Considering only the weights which matter, the hierarchy
 176 looks like the following.
 177
 178           root
 179        /    |   \
 180       A     B    leaf
 181      500   250   125
 182    /  |  \
 183   AA  AB  leaf
 184  500 1000 750
 185
 186 If all cgroups have active IOs and competing with each other, disk
 187 time will be distributed like the following.
 188
 189 Distribution below root. The total active weight at this level is
 190 A:500 + B:250 + C:125 = 875.
 191
 192  root-leaf :   125 /  875      =~ 14%
 193  A         :   500 /  875      =~ 57%
 194  B(-leaf)  :   250 /  875      =~ 28%
 195
 196 A has children and further distributes its 57% among the children and
 197 the implicit leaf node. The total active weight at this level is
 198 AA:500 + AB:1000 + A-leaf:750 = 2250.
 199
 200  A-leaf    : ( 750 / 2250) * A =~ 19%
 201  AA(-leaf) : ( 500 / 2250) * A =~ 12%
 202  AB(-leaf) : (1000 / 2250) * A =~ 25%
 203
 204 CFQ IOPS Mode for group scheduling
 205 ===================================
 206 Basic CFQ design is to provide priority based time slices. Higher priority
 207 process gets bigger time slice and lower priority process gets smaller time
 208 slice. Measuring time becomes harder if storage is fast and supports NCQ and
 209 it would be better to dispatch multiple requests from multiple cfq queues in
 210 request queue at a time. In such scenario, it is not possible to measure time
 211 consumed by single queue accurately.
 212
 213 What is possible though is to measure number of requests dispatched from a
 214 single queue and also allow dispatch from multiple cfq queue at the same time.
 215 This effectively becomes the fairness in terms of IOPS (IO operations per
 216 second).
 217
 218 If one sets slice_idle=0 and if storage supports NCQ, CFQ internally switches
 219 to IOPS mode and starts providing fairness in terms of number of requests
 220 dispatched. Note that this mode switching takes effect only for group
 221 scheduling. For non-cgroup users nothing should change.
 222
 223 CFQ IO scheduler Idling Theory
 224 ===============================
 225 Idling on a queue is primarily about waiting for the next request to come
 226 on same queue after completion of a request. In this process CFQ will not
 227 dispatch requests from other cfq queues even if requests are pending there.
 228
 229 The rationale behind idling is that it can cut down on number of seeks
 230 on rotational media. For example, if a process is doing dependent
 231 sequential reads (next read will come on only after completion of previous
 232 one), then not dispatching request from other queue should help as we
 233 did not move the disk head and kept on dispatching sequential IO from
 234 one queue.
 235
 236 CFQ has following service trees and various queues are put on these trees.
 237
 238         sync-idle       sync-noidle     async
 239
 240 All cfq queues doing synchronous sequential IO go on to sync-idle tree.
 241 On this tree we idle on each queue individually.
 242
 243 All synchronous non-sequential queues go on sync-noidle tree. Also any
 244 request which are marked with REQ_NOIDLE go on this service tree. On this
 245 tree we do not idle on individual queues instead idle on the whole group
 246 of queues or the tree. So if there are 4 queues waiting for IO to dispatch
 247 we will idle only once last queue has dispatched the IO and there is
 248 no more IO on this service tree.
 249
 250 All async writes go on async service tree. There is no idling on async
 251 queues.
 252
 253 CFQ has some optimizations for SSDs and if it detects a non-rotational
 254 media which can support higher queue depth (multiple requests at in
 255 flight at a time), then it cuts down on idling of individual queues and
 256 all the queues move to sync-noidle tree and only tree idle remains. This
 257 tree idling provides isolation with buffered write queues on async tree.
 258
 259 FAQ
 260 ===
 261 Q1. Why to idle at all on queues marked with REQ_NOIDLE.
 262
 263 A1. We only do tree idle (all queues on sync-noidle tree) on queues marked
 264     with REQ_NOIDLE. This helps in providing isolation with all the sync-idle
 265     queues. Otherwise in presence of many sequential readers, other
 266     synchronous IO might not get fair share of disk.
 267
 268     For example, if there are 10 sequential readers doing IO and they get
 269     100ms each. If a REQ_NOIDLE request comes in, it will be scheduled
 270     roughly after 1 second. If after completion of REQ_NOIDLE request we
 271     do not idle, and after a couple of milli seconds a another REQ_NOIDLE
 272     request comes in, again it will be scheduled after 1second. Repeat it
 273     and notice how a workload can lose its disk share and suffer due to
 274     multiple sequential readers.
 275
 276     fsync can generate dependent IO where bunch of data is written in the
 277     context of fsync, and later some journaling data is written. Journaling
 278     data comes in only after fsync has finished its IO (atleast for ext4
 279     that seemed to be the case). Now if one decides not to idle on fsync
 280     thread due to REQ_NOIDLE, then next journaling write will not get
 281     scheduled for another second. A process doing small fsync, will suffer
 282     badly in presence of multiple sequential readers.
 283
 284     Hence doing tree idling on threads using REQ_NOIDLE flag on requests
 285     provides isolation from multiple sequential readers and at the same
 286     time we do not idle on individual threads.
 287
 288 Q2. When to specify REQ_NOIDLE
 289 A2. I would think whenever one is doing synchronous write and not expecting
 290     more writes to be dispatched from same context soon, should be able
 291     to specify REQ_NOIDLE on writes and that probably should work well for
 292     most of the cases.