1CFQ (Complete Fairness Queueing)
   4The main aim of CFQ scheduler is to provide a fair allocation of the disk
   5I/O bandwidth for all the processes which requests an I/O operation.
   7CFQ maintains the per process queue for the processes which request I/O
   8operation(synchronous requests). In case of asynchronous requests, all the
   9requests from all the processes are batched together according to their
  10process's I/O priority.
  12CFQ ioscheduler tunables
  17This specifies how long CFQ should idle for next request on certain cfq queues
  18(for sequential workloads) and service trees (for random workloads) before
  19queue is expired and CFQ selects next queue to dispatch from.
  21By default slice_idle is a non-zero value. That means by default we idle on
  22queues/service trees. This can be very helpful on highly seeky media like
  23single spindle SATA/SAS disks where we can cut down on overall number of
  24seeks and see improved throughput.
  26Setting slice_idle to 0 will remove all the idling on queues/service tree
  27level and one should see an overall improved throughput on faster storage
  28devices like multiple SATA/SAS disks in hardware RAID configuration. The down
  29side is that isolation provided from WRITES also goes down and notion of
  30IO priority becomes weaker.
  32So depending on storage and workload, it might be useful to set slice_idle=0.
  33In general I think for SATA/SAS disks and software RAID of SATA/SAS disks
  34keeping slice_idle enabled should be useful. For any configurations where
  35there are multiple spindles behind single LUN (Host based hardware RAID
  36controller or for storage arrays), setting slice_idle=0 might end up in better
  37throughput and acceptable latencies.
  41This specifies, given in Kbytes, the maximum "distance" for backward seeking.
  42The distance is the amount of space from the current head location to the
  43sectors that are backward in terms of distance.
  45This parameter allows the scheduler to anticipate requests in the "backward"
  46direction and consider them as being the "next" if they are within this
  47distance from the current head location.
  51This parameter is used to compute the cost of backward seeking. If the
  52backward distance of request is just 1/back_seek_penalty from a "front"
  53request, then the seeking cost of two requests is considered equivalent.
  55So scheduler will not bias toward one or the other request (otherwise scheduler
  56will bias toward front request). Default value of back_seek_penalty is 2.
  60This parameter is used to set the timeout of asynchronous requests. Default
  61value of this is 248ms.
  65This parameter is used to set the timeout of synchronous requests. Default
  66value of this is 124ms. In case to favor synchronous requests over asynchronous
  67one, this value should be decreased relative to fifo_expire_async.
  71This parameter forces idling at the CFQ group level instead of CFQ
  72queue level. This was introduced after after a bottleneck was observed
  73in higher end storage due to idle on sequential queue and allow dispatch
  74from a single queue. The idea with this parameter is that it can be run with
  75slice_idle=0 and group_idle=8, so that idling does not happen on individual
  76queues in the group but happens overall on the group and thus still keeps the
  77IO controller working.
  78Not idling on individual queues in the group will dispatch requests from
  79multiple queues in the group at the same time and achieve higher throughput
  80on higher end storage.
  82Default value for this parameter is 8ms.
  86This parameter is used to enable/disable the latency mode of the CFQ
  87scheduler. If latency mode (called low_latency) is enabled, CFQ tries
  88to recompute the slice time for each process based on the target_latency set
  89for the system. This favors fairness over throughput. Disabling low
  90latency (setting it to 0) ignores target latency, allowing each process in the
  91system to get a full time slice.
  93By default low latency mode is enabled.
  97This parameter is used to calculate the time slice for a process if cfq's
  98latency mode is enabled. It will ensure that sync requests have an estimated
  99latency. But if sequential workload is higher(e.g. sequential read),
 100then to meet the latency constraints, throughput may decrease because of less
 101time for each process to issue I/O request before the cfq queue is switched.
 103Though this can be overcome by disabling the latency_mode, it may increase
 104the read latency for some applications. This parameter allows for changing
 105target_latency through the sysfs interface which can provide the balanced
 106throughput and read latency.
 108Default value for target_latency is 300ms.
 112This parameter is same as of slice_sync but for asynchronous queue. The
 113default value is 40ms.
 117This parameter is used to limit the dispatching of asynchronous request to
 118device request queue in queue's slice time. The maximum number of request that
 119are allowed to be dispatched also depends upon the io priority. Default value
 120for this is 2.
 124When a queue is selected for execution, the queues IO requests are only
 125executed for a certain amount of time(time_slice) before switching to another
 126queue. This parameter is used to calculate the time slice of synchronous
 129time_slice is computed using the below equation:-
 130time_slice = slice_sync + (slice_sync/5 * (4 - prio)). To increase the
 131time_slice of synchronous queue, increase the value of slice_sync. Default
 132value is 100ms.
 136This specifies the number of request dispatched to the device queue. In a
 137queue's time slice, a request will not be dispatched if the number of request
 138in the device exceeds this parameter. This parameter is used for synchronous
 141In case of storage with several disk, this setting can limit the parallel
 142processing of request. Therefore, increasing the value can improve the
 143performance although this can cause the latency of some I/O to increase due
 144to more number of requests.
 146CFQ Group scheduling
 149CFQ supports blkio cgroup and has "blkio." prefixed files in each
 150blkio cgroup directory. It is weight-based and there are four knobs
 151for configuration - weight[_device] and leaf_weight[_device].
 152Internal cgroup nodes (the ones with children) can also have tasks in
 153them, so the former two configure how much proportion the cgroup as a
 154whole is entitled to at its parent's level while the latter two
 155configure how much proportion the tasks in the cgroup have compared to
 156its direct children.
 158Another way to think about it is assuming that each internal node has
 159an implicit leaf child node which hosts all the tasks whose weight is
 160configured by leaf_weight[_device]. Let's assume a blkio hierarchy
 161composed of five cgroups - root, A, B, AA and AB - with the following
 162weights where the names represent the hierarchy.
 164        weight leaf_weight
 165 root :  125    125
 166 A    :  500    750
 167 B    :  250    500
 168 AA   :  500    500
 169 AB   : 1000    500
 171root never has a parent making its weight is meaningless. For backward
 172compatibility, weight is always kept in sync with leaf_weight. B, AA
 173and AB have no child and thus its tasks have no children cgroup to
 174compete with. They always get 100% of what the cgroup won at the
 175parent level. Considering only the weights which matter, the hierarchy
 176looks like the following.
 178          root
 179       /    |   \
 180      A     B    leaf
 181     500   250   125
 182   /  |  \
 183  AA  AB  leaf
 184 500 1000 750
 186If all cgroups have active IOs and competing with each other, disk
 187time will be distributed like the following.
 189Distribution below root. The total active weight at this level is
 190A:500 + B:250 + C:125 = 875.
 192 root-leaf :   125 /  875      =~ 14%
 193 A         :   500 /  875      =~ 57%
 194 B(-leaf)  :   250 /  875      =~ 28%
 196A has children and further distributes its 57% among the children and
 197the implicit leaf node. The total active weight at this level is
 198AA:500 + AB:1000 + A-leaf:750 = 2250.
 200 A-leaf    : ( 750 / 2250) * A =~ 19%
 201 AA(-leaf) : ( 500 / 2250) * A =~ 12%
 202 AB(-leaf) : (1000 / 2250) * A =~ 25%
 204CFQ IOPS Mode for group scheduling
 206Basic CFQ design is to provide priority based time slices. Higher priority
 207process gets bigger time slice and lower priority process gets smaller time
 208slice. Measuring time becomes harder if storage is fast and supports NCQ and
 209it would be better to dispatch multiple requests from multiple cfq queues in
 210request queue at a time. In such scenario, it is not possible to measure time
 211consumed by single queue accurately.
 213What is possible though is to measure number of requests dispatched from a
 214single queue and also allow dispatch from multiple cfq queue at the same time.
 215This effectively becomes the fairness in terms of IOPS (IO operations per
 218If one sets slice_idle=0 and if storage supports NCQ, CFQ internally switches
 219to IOPS mode and starts providing fairness in terms of number of requests
 220dispatched. Note that this mode switching takes effect only for group
 221scheduling. For non-cgroup users nothing should change.
 223CFQ IO scheduler Idling Theory
 225Idling on a queue is primarily about waiting for the next request to come
 226on same queue after completion of a request. In this process CFQ will not
 227dispatch requests from other cfq queues even if requests are pending there.
 229The rationale behind idling is that it can cut down on number of seeks
 230on rotational media. For example, if a process is doing dependent
 231sequential reads (next read will come on only after completion of previous
 232one), then not dispatching request from other queue should help as we
 233did not move the disk head and kept on dispatching sequential IO from
 234one queue.
 236CFQ has following service trees and various queues are put on these trees.
 238        sync-idle       sync-noidle     async
 240All cfq queues doing synchronous sequential IO go on to sync-idle tree.
 241On this tree we idle on each queue individually.
 243All synchronous non-sequential queues go on sync-noidle tree. Also any
 244request which are marked with REQ_NOIDLE go on this service tree. On this
 245tree we do not idle on individual queues instead idle on the whole group
 246of queues or the tree. So if there are 4 queues waiting for IO to dispatch
 247we will idle only once last queue has dispatched the IO and there is
 248no more IO on this service tree.
 250All async writes go on async service tree. There is no idling on async
 253CFQ has some optimizations for SSDs and if it detects a non-rotational
 254media which can support higher queue depth (multiple requests at in
 255flight at a time), then it cuts down on idling of individual queues and
 256all the queues move to sync-noidle tree and only tree idle remains. This
 257tree idling provides isolation with buffered write queues on async tree.
 261Q1. Why to idle at all on queues marked with REQ_NOIDLE.
 263A1. 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.
 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.
 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.
 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.
 288Q2. When to specify REQ_NOIDLE
 289A2. 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.
 293 kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.