summaryrefslogtreecommitdiff
path: root/Documentation/block
diff options
context:
space:
mode:
authorJens Axboe <axboe@kernel.dk>2018-10-12 10:14:46 -0600
committerJens Axboe <axboe@kernel.dk>2018-11-07 13:42:32 -0700
commitf382fb0bcef4c37dc049e9f6963e3baf204d815c (patch)
treedbfbe5689176a03ea1590497f965b40b2f8fd532 /Documentation/block
parent404b8f5a03d840f74669fd55e26f8e3564cc2dd8 (diff)
block: remove legacy IO schedulers
Retain the deadline documentation, as that carries over to mq-deadline as well. Tested-by: Ming Lei <ming.lei@redhat.com> Reviewed-by: Omar Sandoval <osandov@fb.com> Signed-off-by: Jens Axboe <axboe@kernel.dk>
Diffstat (limited to 'Documentation/block')
-rw-r--r--Documentation/block/cfq-iosched.txt291
1 files changed, 0 insertions, 291 deletions
diff --git a/Documentation/block/cfq-iosched.txt b/Documentation/block/cfq-iosched.txt
deleted file mode 100644
index 895bd3813115..000000000000
--- a/Documentation/block/cfq-iosched.txt
+++ /dev/null
@@ -1,291 +0,0 @@
-CFQ (Complete Fairness Queueing)
-===============================
-
-The main aim of CFQ scheduler is to provide a fair allocation of the disk
-I/O bandwidth for all the processes which requests an I/O operation.
-
-CFQ maintains the per process queue for the processes which request I/O
-operation(synchronous requests). In case of asynchronous requests, all the
-requests from all the processes are batched together according to their
-process's I/O priority.
-
-CFQ ioscheduler tunables
-========================
-
-slice_idle
-----------
-This specifies how long CFQ should idle for next request on certain cfq queues
-(for sequential workloads) and service trees (for random workloads) before
-queue is expired and CFQ selects next queue to dispatch from.
-
-By default slice_idle is a non-zero value. That means by default we idle on
-queues/service trees. This can be very helpful on highly seeky media like
-single spindle SATA/SAS disks where we can cut down on overall number of
-seeks and see improved throughput.
-
-Setting slice_idle to 0 will remove all the idling on queues/service tree
-level and one should see an overall improved throughput on faster storage
-devices like multiple SATA/SAS disks in hardware RAID configuration. The down
-side is that isolation provided from WRITES also goes down and notion of
-IO priority becomes weaker.
-
-So depending on storage and workload, it might be useful to set slice_idle=0.
-In general I think for SATA/SAS disks and software RAID of SATA/SAS disks
-keeping slice_idle enabled should be useful. For any configurations where
-there are multiple spindles behind single LUN (Host based hardware RAID
-controller or for storage arrays), setting slice_idle=0 might end up in better
-throughput and acceptable latencies.
-
-back_seek_max
--------------
-This specifies, given in Kbytes, the maximum "distance" for backward seeking.
-The distance is the amount of space from the current head location to the
-sectors that are backward in terms of distance.
-
-This parameter allows the scheduler to anticipate requests in the "backward"
-direction and consider them as being the "next" if they are within this
-distance from the current head location.
-
-back_seek_penalty
------------------
-This parameter is used to compute the cost of backward seeking. If the
-backward distance of request is just 1/back_seek_penalty from a "front"
-request, then the seeking cost of two requests is considered equivalent.
-
-So scheduler will not bias toward one or the other request (otherwise scheduler
-will bias toward front request). Default value of back_seek_penalty is 2.
-
-fifo_expire_async
------------------
-This parameter is used to set the timeout of asynchronous requests. Default
-value of this is 248ms.
-
-fifo_expire_sync
-----------------
-This parameter is used to set the timeout of synchronous requests. Default
-value of this is 124ms. In case to favor synchronous requests over asynchronous
-one, this value should be decreased relative to fifo_expire_async.
-
-group_idle
------------
-This parameter forces idling at the CFQ group level instead of CFQ
-queue level. This was introduced after a bottleneck was observed
-in higher end storage due to idle on sequential queue and allow dispatch
-from a single queue. The idea with this parameter is that it can be run with
-slice_idle=0 and group_idle=8, so that idling does not happen on individual
-queues in the group but happens overall on the group and thus still keeps the
-IO controller working.
-Not idling on individual queues in the group will dispatch requests from
-multiple queues in the group at the same time and achieve higher throughput
-on higher end storage.
-
-Default value for this parameter is 8ms.
-
-low_latency
------------
-This parameter is used to enable/disable the low latency mode of the CFQ
-scheduler. If enabled, CFQ tries to recompute the slice time for each process
-based on the target_latency set for the system. This favors fairness over
-throughput. Disabling low latency (setting it to 0) ignores target latency,
-allowing each process in the system to get a full time slice.
-
-By default low latency mode is enabled.
-
-target_latency
---------------
-This parameter is used to calculate the time slice for a process if cfq's
-latency mode is enabled. It will ensure that sync requests have an estimated
-latency. But if sequential workload is higher(e.g. sequential read),
-then to meet the latency constraints, throughput may decrease because of less
-time for each process to issue I/O request before the cfq queue is switched.
-
-Though this can be overcome by disabling the latency_mode, it may increase
-the read latency for some applications. This parameter allows for changing
-target_latency through the sysfs interface which can provide the balanced
-throughput and read latency.
-
-Default value for target_latency is 300ms.
-
-slice_async
------------
-This parameter is same as of slice_sync but for asynchronous queue. The
-default value is 40ms.
-
-slice_async_rq
---------------
-This parameter is used to limit the dispatching of asynchronous request to
-device request queue in queue's slice time. The maximum number of request that
-are allowed to be dispatched also depends upon the io priority. Default value
-for this is 2.
-
-slice_sync
-----------
-When a queue is selected for execution, the queues IO requests are only
-executed for a certain amount of time(time_slice) before switching to another
-queue. This parameter is used to calculate the time slice of synchronous
-queue.
-
-time_slice is computed using the below equation:-
-time_slice = slice_sync + (slice_sync/5 * (4 - prio)). To increase the
-time_slice of synchronous queue, increase the value of slice_sync. Default
-value is 100ms.
-
-quantum
--------
-This specifies the number of request dispatched to the device queue. In a
-queue's time slice, a request will not be dispatched if the number of request
-in the device exceeds this parameter. This parameter is used for synchronous
-request.
-
-In case of storage with several disk, this setting can limit the parallel
-processing of request. Therefore, increasing the value can improve the
-performance although this can cause the latency of some I/O to increase due
-to more number of requests.
-
-CFQ Group scheduling
-====================
-
-CFQ supports blkio cgroup and has "blkio." prefixed files in each
-blkio cgroup directory. It is weight-based and there are four knobs
-for configuration - weight[_device] and leaf_weight[_device].
-Internal cgroup nodes (the ones with children) can also have tasks in
-them, so the former two configure how much proportion the cgroup as a
-whole is entitled to at its parent's level while the latter two
-configure how much proportion the tasks in the cgroup have compared to
-its direct children.
-
-Another way to think about it is assuming that each internal node has
-an implicit leaf child node which hosts all the tasks whose weight is
-configured by leaf_weight[_device]. Let's assume a blkio hierarchy
-composed of five cgroups - root, A, B, AA and AB - with the following
-weights where the names represent the hierarchy.
-
- weight leaf_weight
- root : 125 125
- A : 500 750
- B : 250 500
- AA : 500 500
- AB : 1000 500
-
-root never has a parent making its weight is meaningless. For backward
-compatibility, weight is always kept in sync with leaf_weight. B, AA
-and AB have no child and thus its tasks have no children cgroup to
-compete with. They always get 100% of what the cgroup won at the
-parent level. Considering only the weights which matter, the hierarchy
-looks like the following.
-
- root
- / | \
- A B leaf
- 500 250 125
- / | \
- AA AB leaf
- 500 1000 750
-
-If all cgroups have active IOs and competing with each other, disk
-time will be distributed like the following.
-
-Distribution below root. The total active weight at this level is
-A:500 + B:250 + C:125 = 875.
-
- root-leaf : 125 / 875 =~ 14%
- A : 500 / 875 =~ 57%
- B(-leaf) : 250 / 875 =~ 28%
-
-A has children and further distributes its 57% among the children and
-the implicit leaf node. The total active weight at this level is
-AA:500 + AB:1000 + A-leaf:750 = 2250.
-
- A-leaf : ( 750 / 2250) * A =~ 19%
- AA(-leaf) : ( 500 / 2250) * A =~ 12%
- AB(-leaf) : (1000 / 2250) * A =~ 25%
-
-CFQ IOPS Mode for group scheduling
-===================================
-Basic CFQ design is to provide priority based time slices. Higher priority
-process gets bigger time slice and lower priority process gets smaller time
-slice. Measuring time becomes harder if storage is fast and supports NCQ and
-it would be better to dispatch multiple requests from multiple cfq queues in
-request queue at a time. In such scenario, it is not possible to measure time
-consumed by single queue accurately.
-
-What is possible though is to measure number of requests dispatched from a
-single queue and also allow dispatch from multiple cfq queue at the same time.
-This effectively becomes the fairness in terms of IOPS (IO operations per
-second).
-
-If one sets slice_idle=0 and if storage supports NCQ, CFQ internally switches
-to IOPS mode and starts providing fairness in terms of number of requests
-dispatched. Note that this mode switching takes effect only for group
-scheduling. For non-cgroup users nothing should change.
-
-CFQ IO scheduler Idling Theory
-===============================
-Idling on a queue is primarily about waiting for the next request to come
-on same queue after completion of a request. In this process CFQ will not
-dispatch requests from other cfq queues even if requests are pending there.
-
-The rationale behind idling is that it can cut down on number of seeks
-on rotational media. For example, if a process is doing dependent
-sequential reads (next read will come on only after completion of previous
-one), then not dispatching request from other queue should help as we
-did not move the disk head and kept on dispatching sequential IO from
-one queue.
-
-CFQ has following service trees and various queues are put on these trees.
-
- sync-idle sync-noidle async
-
-All cfq queues doing synchronous sequential IO go on to sync-idle tree.
-On this tree we idle on each queue individually.
-
-All synchronous non-sequential queues go on sync-noidle tree. Also any
-synchronous write request which is not marked with REQ_IDLE goes on this
-service tree. On this tree we do not idle on individual queues instead idle
-on the whole group of queues or the tree. So if there are 4 queues waiting
-for IO to dispatch we will idle only once last queue has dispatched the IO
-and there is no more IO on this service tree.
-
-All async writes go on async service tree. There is no idling on async
-queues.
-
-CFQ has some optimizations for SSDs and if it detects a non-rotational
-media which can support higher queue depth (multiple requests at in
-flight at a time), then it cuts down on idling of individual queues and
-all the queues move to sync-noidle tree and only tree idle remains. This
-tree idling provides isolation with buffered write queues on async tree.
-
-FAQ
-===
-Q1. Why to idle at all on queues not marked with REQ_IDLE.
-
-A1. We only do tree idle (all queues on sync-noidle tree) on queues not marked
- with REQ_IDLE. This helps in providing isolation with all the sync-idle
- queues. Otherwise in presence of many sequential readers, other
- synchronous IO might not get fair share of disk.
-
- For example, if there are 10 sequential readers doing IO and they get
- 100ms each. If a !REQ_IDLE request comes in, it will be scheduled
- roughly after 1 second. If after completion of !REQ_IDLE request we
- do not idle, and after a couple of milli seconds a another !REQ_IDLE
- request comes in, again it will be scheduled after 1second. Repeat it
- and notice how a workload can lose its disk share and suffer due to
- multiple sequential readers.
-
- fsync can generate dependent IO where bunch of data is written in the
- context of fsync, and later some journaling data is written. Journaling
- data comes in only after fsync has finished its IO (atleast for ext4
- that seemed to be the case). Now if one decides not to idle on fsync
- thread due to !REQ_IDLE, then next journaling write will not get
- scheduled for another second. A process doing small fsync, will suffer
- badly in presence of multiple sequential readers.
-
- Hence doing tree idling on threads using !REQ_IDLE flag on requests
- provides isolation from multiple sequential readers and at the same
- time we do not idle on individual threads.
-
-Q2. When to specify REQ_IDLE
-A2. I would think whenever one is doing synchronous write and expecting
- more writes to be dispatched from same context soon, should be able
- to specify REQ_IDLE on writes and that probably should work well for
- most of the cases.