linux_dsm_epyc7002/Documentation/block/bfq-iosched.rst

==========================
BFQ (Budget Fair Queueing)
==========================

BFQ is a proportional-share I/O scheduler, with some extra
low-latency capabilities. In addition to cgroups support (blkio or io
controllers), BFQ's main features are:

- BFQ guarantees a high system and application responsiveness, and a
  low latency for time-sensitive applications, such as audio or video
  players;
- BFQ distributes bandwidth, and not just time, among processes or
  groups (switching back to time distribution when needed to keep
  throughput high).

In its default configuration, BFQ privileges latency over
throughput. So, when needed for achieving a lower latency, BFQ builds
schedules that may lead to a lower throughput. If your main or only
goal, for a given device, is to achieve the maximum-possible
throughput at all times, then do switch off all low-latency heuristics
for that device, by setting low_latency to 0. See Section 3 for
details on how to configure BFQ for the desired tradeoff between
latency and throughput, or on how to maximize throughput.

As every I/O scheduler, BFQ adds some overhead to per-I/O-request
processing. To give an idea of this overhead, the total,
single-lock-protected, per-request processing time of BFQ---i.e., the
sum of the execution times of the request insertion, dispatch and
completion hooks---is, e.g., 1.9 us on an Intel Core i7-2760QM@2.40GHz
(dated CPU for notebooks; time measured with simple code
instrumentation, and using the throughput-sync.sh script of the S
suite [1], in performance-profiling mode). To put this result into
context, the total, single-lock-protected, per-request execution time
of the lightest I/O scheduler available in blk-mq, mq-deadline, is 0.7
us (mq-deadline is ~800 LOC, against ~10500 LOC for BFQ).

Scheduling overhead further limits the maximum IOPS that a CPU can
process (already limited by the execution of the rest of the I/O
stack). To give an idea of the limits with BFQ, on slow or average
CPUs, here are, first, the limits of BFQ for three different CPUs, on,
respectively, an average laptop, an old desktop, and a cheap embedded
system, in case full hierarchical support is enabled (i.e.,
CONFIG_BFQ_GROUP_IOSCHED is set), but CONFIG_BFQ_CGROUP_DEBUG is not
set (Section 4-2):
- Intel i7-4850HQ: 400 KIOPS
- AMD A8-3850: 250 KIOPS
- ARM CortexTM-A53 Octa-core: 80 KIOPS

If CONFIG_BFQ_CGROUP_DEBUG is set (and of course full hierarchical
support is enabled), then the sustainable throughput with BFQ
decreases, because all blkio.bfq* statistics are created and updated
(Section 4-2). For BFQ, this leads to the following maximum
sustainable throughputs, on the same systems as above:
- Intel i7-4850HQ: 310 KIOPS
- AMD A8-3850: 200 KIOPS
- ARM CortexTM-A53 Octa-core: 56 KIOPS

BFQ works for multi-queue devices too.

.. The table of contents follow. Impatients can just jump to Section 3.

.. CONTENTS

   1. When may BFQ be useful?
    1-1 Personal systems
    1-2 Server systems
   2. How does BFQ work?
   3. What are BFQ's tunables and how to properly configure BFQ?
   4. BFQ group scheduling
    4-1 Service guarantees provided
    4-2 Interface

1. When may BFQ be useful?
==========================

BFQ provides the following benefits on personal and server systems.

1-1 Personal systems
--------------------

Low latency for interactive applications
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^

Regardless of the actual background workload, BFQ guarantees that, for
interactive tasks, the storage device is virtually as responsive as if
it was idle. For example, even if one or more of the following
background workloads are being executed:

- one or more large files are being read, written or copied,
- a tree of source files is being compiled,
- one or more virtual machines are performing I/O,
- a software update is in progress,
- indexing daemons are scanning filesystems and updating their
  databases,

starting an application or loading a file from within an application
takes about the same time as if the storage device was idle. As a
comparison, with CFQ, NOOP or DEADLINE, and in the same conditions,
applications experience high latencies, or even become unresponsive
until the background workload terminates (also on SSDs).

Low latency for soft real-time applications
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
Also soft real-time applications, such as audio and video
players/streamers, enjoy a low latency and a low drop rate, regardless
of the background I/O workload. As a consequence, these applications
do not suffer from almost any glitch due to the background workload.

Higher speed for code-development tasks
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^

If some additional workload happens to be executed in parallel, then
BFQ executes the I/O-related components of typical code-development
tasks (compilation, checkout, merge, ...) much more quickly than CFQ,
NOOP or DEADLINE.

High throughput
^^^^^^^^^^^^^^^

On hard disks, BFQ achieves up to 30% higher throughput than CFQ, and
up to 150% higher throughput than DEADLINE and NOOP, with all the
sequential workloads considered in our tests. With random workloads,
and with all the workloads on flash-based devices, BFQ achieves,
instead, about the same throughput as the other schedulers.

Strong fairness, bandwidth and delay guarantees
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^

BFQ distributes the device throughput, and not just the device time,
among I/O-bound applications in proportion their weights, with any
workload and regardless of the device parameters. From these bandwidth
guarantees, it is possible to compute tight per-I/O-request delay
guarantees by a simple formula. If not configured for strict service
guarantees, BFQ switches to time-based resource sharing (only) for
applications that would otherwise cause a throughput loss.

1-2 Server systems
------------------

Most benefits for server systems follow from the same service
properties as above. In particular, regardless of whether additional,
possibly heavy workloads are being served, BFQ guarantees:

* audio and video-streaming with zero or very low jitter and drop
  rate;

* fast retrieval of WEB pages and embedded objects;

* real-time recording of data in live-dumping applications (e.g.,
  packet logging);

* responsiveness in local and remote access to a server.


2. How does BFQ work?
=====================

BFQ is a proportional-share I/O scheduler, whose general structure,
plus a lot of code, are borrowed from CFQ.

- Each process doing I/O on a device is associated with a weight and a
  `(bfq_)queue`.

- BFQ grants exclusive access to the device, for a while, to one queue
  (process) at a time, and implements this service model by
  associating every queue with a budget, measured in number of
  sectors.

  - After a queue is granted access to the device, the budget of the
    queue is decremented, on each request dispatch, by the size of the
    request.

  - The in-service queue is expired, i.e., its service is suspended,
    only if one of the following events occurs: 1) the queue finishes
    its budget, 2) the queue empties, 3) a "budget timeout" fires.

    - The budget timeout prevents processes doing random I/O from
      holding the device for too long and dramatically reducing
      throughput.

    - Actually, as in CFQ, a queue associated with a process issuing
      sync requests may not be expired immediately when it empties. In
      contrast, BFQ may idle the device for a short time interval,
      giving the process the chance to go on being served if it issues
      a new request in time. Device idling typically boosts the
      throughput on rotational devices and on non-queueing flash-based
      devices, if processes do synchronous and sequential I/O. In
      addition, under BFQ, device idling is also instrumental in
      guaranteeing the desired throughput fraction to processes
      issuing sync requests (see the description of the slice_idle
      tunable in this document, or [1, 2], for more details).

      - With respect to idling for service guarantees, if several
	processes are competing for the device at the same time, but
	all processes and groups have the same weight, then BFQ
	guarantees the expected throughput distribution without ever
	idling the device. Throughput is thus as high as possible in
	this common scenario.

     - On flash-based storage with internal queueing of commands
       (typically NCQ), device idling happens to be always detrimental
       for throughput. So, with these devices, BFQ performs idling
       only when strictly needed for service guarantees, i.e., for
       guaranteeing low latency or fairness. In these cases, overall
       throughput may be sub-optimal. No solution currently exists to
       provide both strong service guarantees and optimal throughput
       on devices with internal queueing.

  - If low-latency mode is enabled (default configuration), BFQ
    executes some special heuristics to detect interactive and soft
    real-time applications (e.g., video or audio players/streamers),
    and to reduce their latency. The most important action taken to
    achieve this goal is to give to the queues associated with these
    applications more than their fair share of the device
    throughput. For brevity, we call just "weight-raising" the whole
    sets of actions taken by BFQ to privilege these queues. In
    particular, BFQ provides a milder form of weight-raising for
    interactive applications, and a stronger form for soft real-time
    applications.

  - BFQ automatically deactivates idling for queues born in a burst of
    queue creations. In fact, these queues are usually associated with
    the processes of applications and services that benefit mostly
    from a high throughput. Examples are systemd during boot, or git
    grep.

  - As CFQ, BFQ merges queues performing interleaved I/O, i.e.,
    performing random I/O that becomes mostly sequential if
    merged. Differently from CFQ, BFQ achieves this goal with a more
    reactive mechanism, called Early Queue Merge (EQM). EQM is so
    responsive in detecting interleaved I/O (cooperating processes),
    that it enables BFQ to achieve a high throughput, by queue
    merging, even for queues for which CFQ needs a different
    mechanism, preemption, to get a high throughput. As such EQM is a
    unified mechanism to achieve a high throughput with interleaved
    I/O.

  - Queues are scheduled according to a variant of WF2Q+, named
    B-WF2Q+, and implemented using an augmented rb-tree to preserve an
    O(log N) overall complexity.  See [2] for more details. B-WF2Q+ is
    also ready for hierarchical scheduling, details in Section 4.

  - B-WF2Q+ guarantees a tight deviation with respect to an ideal,
    perfectly fair, and smooth service. In particular, B-WF2Q+
    guarantees that each queue receives a fraction of the device
    throughput proportional to its weight, even if the throughput
    fluctuates, and regardless of: the device parameters, the current
    workload and the budgets assigned to the queue.

  - The last, budget-independence, property (although probably
    counterintuitive in the first place) is definitely beneficial, for
    the following reasons:

    - First, with any proportional-share scheduler, the maximum
      deviation with respect to an ideal service is proportional to
      the maximum budget (slice) assigned to queues. As a consequence,
      BFQ can keep this deviation tight not only because of the
      accurate service of B-WF2Q+, but also because BFQ *does not*
      need to assign a larger budget to a queue to let the queue
      receive a higher fraction of the device throughput.

    - Second, BFQ is free to choose, for every process (queue), the
      budget that best fits the needs of the process, or best
      leverages the I/O pattern of the process. In particular, BFQ
      updates queue budgets with a simple feedback-loop algorithm that
      allows a high throughput to be achieved, while still providing
      tight latency guarantees to time-sensitive applications. When
      the in-service queue expires, this algorithm computes the next
      budget of the queue so as to:

      - Let large budgets be eventually assigned to the queues
	associated with I/O-bound applications performing sequential
	I/O: in fact, the longer these applications are served once
	got access to the device, the higher the throughput is.

      - Let small budgets be eventually assigned to the queues
	associated with time-sensitive applications (which typically
	perform sporadic and short I/O), because, the smaller the
	budget assigned to a queue waiting for service is, the sooner
	B-WF2Q+ will serve that queue (Subsec 3.3 in [2]).

- If several processes are competing for the device at the same time,
  but all processes and groups have the same weight, then BFQ
  guarantees the expected throughput distribution without ever idling
  the device. It uses preemption instead. Throughput is then much
  higher in this common scenario.

- ioprio classes are served in strict priority order, i.e.,
  lower-priority queues are not served as long as there are
  higher-priority queues.  Among queues in the same class, the
  bandwidth is distributed in proportion to the weight of each
  queue. A very thin extra bandwidth is however guaranteed to
  the Idle class, to prevent it from starving.


3. What are BFQ's tunables and how to properly configure BFQ?
=============================================================

Most BFQ tunables affect service guarantees (basically latency and
fairness) and throughput. For full details on how to choose the
desired tradeoff between service guarantees and throughput, see the
parameters slice_idle, strict_guarantees and low_latency. For details
on how to maximise throughput, see slice_idle, timeout_sync and
max_budget. The other performance-related parameters have been
inherited from, and have been preserved mostly for compatibility with
CFQ. So far, no performance improvement has been reported after
changing the latter parameters in BFQ.

In particular, the tunables back_seek-max, back_seek_penalty,
fifo_expire_async and fifo_expire_sync below are the same as in
CFQ. Their description is just copied from that for CFQ. Some
considerations in the description of slice_idle are copied from CFQ
too.

per-process ioprio and weight
-----------------------------

Unless the cgroups interface is used (see "4. BFQ group scheduling"),
weights can be assigned to processes only indirectly, through I/O
priorities, and according to the relation:
weight = (IOPRIO_BE_NR - ioprio) * 10.

Beware that, if low-latency is set, then BFQ automatically raises the
weight of the queues associated with interactive and soft real-time
applications. Unset this tunable if you need/want to control weights.

slice_idle
----------

This parameter specifies how long BFQ should idle for next I/O
request, when certain sync BFQ queues become empty. By default
slice_idle is a non-zero value. Idling has a double purpose: boosting
throughput and making sure that the desired throughput distribution is
respected (see the description of how BFQ works, and, if needed, the
papers referred there).

As for throughput, idling 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 and one
should see an overall improved throughput on faster storage devices
like multiple SATA/SAS disks in hardware RAID configuration, as well
as flash-based storage with internal command queueing (and
parallelism).

So depending on storage and workload, it might be useful to set
slice_idle=0.  In general 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), or with
flash-based fast storage, setting slice_idle=0 might end up in better
throughput and acceptable latencies.

Idling is however necessary to have service guarantees enforced in
case of differentiated weights or differentiated I/O-request lengths.
To see why, suppose that a given BFQ queue A must get several I/O
requests served for each request served for another queue B. Idling
ensures that, if A makes a new I/O request slightly after becoming
empty, then no request of B is dispatched in the middle, and thus A
does not lose the possibility to get more than one request dispatched
before the next request of B is dispatched. Note that idling
guarantees the desired differentiated treatment of queues only in
terms of I/O-request dispatches. To guarantee that the actual service
order then corresponds to the dispatch order, the strict_guarantees
tunable must be set too.

There is an important flipside for idling: apart from the above cases
where it is beneficial also for throughput, idling can severely impact
throughput. One important case is random workload. Because of this
issue, BFQ tends to avoid idling as much as possible, when it is not
beneficial also for throughput (as detailed in Section 2). As a
consequence of this behavior, and of further issues described for the
strict_guarantees tunable, short-term service guarantees may be
occasionally violated. And, in some cases, these guarantees may be
more important than guaranteeing maximum throughput. For example, in
video playing/streaming, a very low drop rate may be more important
than maximum throughput. In these cases, consider setting the
strict_guarantees parameter.

slice_idle_us
-------------

Controls the same tuning parameter as slice_idle, but in microseconds.
Either tunable can be used to set idling behavior.  Afterwards, the
other tunable will reflect the newly set value in sysfs.

strict_guarantees
-----------------

If this parameter is set (default: unset), then BFQ

- always performs idling when the in-service queue becomes empty;

- forces the device to serve one I/O request at a time, by dispatching a
  new request only if there is no outstanding request.

In the presence of differentiated weights or I/O-request sizes, both
the above conditions are needed to guarantee that every BFQ queue
receives its allotted share of the bandwidth. The first condition is
needed for the reasons explained in the description of the slice_idle
tunable.  The second condition is needed because all modern storage
devices reorder internally-queued requests, which may trivially break
the service guarantees enforced by the I/O scheduler.

Setting strict_guarantees may evidently affect throughput.

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.

low_latency
-----------

This parameter is used to enable/disable BFQ's low latency mode. By
default, low latency mode is enabled. If enabled, interactive and soft
real-time applications are privileged and experience a lower latency,
as explained in more detail in the description of how BFQ works.

DISABLE this mode if you need full control on bandwidth
distribution. In fact, if it is enabled, then BFQ automatically
increases the bandwidth share of privileged applications, as the main
means to guarantee a lower latency to them.

In addition, as already highlighted at the beginning of this document,
DISABLE this mode if your only goal is to achieve a high throughput.
In fact, privileging the I/O of some application over the rest may
entail a lower throughput. To achieve the highest-possible throughput
on a non-rotational device, setting slice_idle to 0 may be needed too
(at the cost of giving up any strong guarantee on fairness and low
latency).

timeout_sync
------------

Maximum amount of device time that can be given to a task (queue) once
it has been selected for service. On devices with costly seeks,
increasing this time usually increases maximum throughput. On the
opposite end, increasing this time coarsens the granularity of the
short-term bandwidth and latency guarantees, especially if the
following parameter is set to zero.

max_budget
----------

Maximum amount of service, measured in sectors, that can be provided
to a BFQ queue once it is set in service (of course within the limits
of the above timeout). According to what said in the description of
the algorithm, larger values increase the throughput in proportion to
the percentage of sequential I/O requests issued. The price of larger
values is that they coarsen the granularity of short-term bandwidth
and latency guarantees.

The default value is 0, which enables auto-tuning: BFQ sets max_budget
to the maximum number of sectors that can be served during
timeout_sync, according to the estimated peak rate.

For specific devices, some users have occasionally reported to have
reached a higher throughput by setting max_budget explicitly, i.e., by
setting max_budget to a higher value than 0. In particular, they have
set max_budget to higher values than those to which BFQ would have set
it with auto-tuning. An alternative way to achieve this goal is to
just increase the value of timeout_sync, leaving max_budget equal to 0.

4. Group scheduling with BFQ
============================

BFQ supports both cgroups-v1 and cgroups-v2 io controllers, namely
blkio and io. In particular, BFQ supports weight-based proportional
share. To activate cgroups support, set BFQ_GROUP_IOSCHED.

4-1 Service guarantees provided
-------------------------------

With BFQ, proportional share means true proportional share of the
device bandwidth, according to group weights. For example, a group
with weight 200 gets twice the bandwidth, and not just twice the time,
of a group with weight 100.

BFQ supports hierarchies (group trees) of any depth. Bandwidth is
distributed among groups and processes in the expected way: for each
group, the children of the group share the whole bandwidth of the
group in proportion to their weights. In particular, this implies
that, for each leaf group, every process of the group receives the
same share of the whole group bandwidth, unless the ioprio of the
process is modified.

The resource-sharing guarantee for a group may partially or totally
switch from bandwidth to time, if providing bandwidth guarantees to
the group lowers the throughput too much. This switch occurs on a
per-process basis: if a process of a leaf group causes throughput loss
if served in such a way to receive its share of the bandwidth, then
BFQ switches back to just time-based proportional share for that
process.

4-2 Interface
-------------

To get proportional sharing of bandwidth with BFQ for a given device,
BFQ must of course be the active scheduler for that device.

Within each group directory, the names of the files associated with
BFQ-specific cgroup parameters and stats begin with the "bfq."
prefix. So, with cgroups-v1 or cgroups-v2, the full prefix for
BFQ-specific files is "blkio.bfq." or "io.bfq." For example, the group
parameter to set the weight of a group with BFQ is blkio.bfq.weight
or io.bfq.weight.

As for cgroups-v1 (blkio controller), the exact set of stat files
created, and kept up-to-date by bfq, depends on whether
CONFIG_BFQ_CGROUP_DEBUG is set. If it is set, then bfq creates all
the stat files documented in
Documentation/admin-guide/cgroup-v1/blkio-controller.rst. If, instead,
CONFIG_BFQ_CGROUP_DEBUG is not set, then bfq creates only the files::

  blkio.bfq.io_service_bytes
  blkio.bfq.io_service_bytes_recursive
  blkio.bfq.io_serviced
  blkio.bfq.io_serviced_recursive

The value of CONFIG_BFQ_CGROUP_DEBUG greatly influences the maximum
throughput sustainable with bfq, because updating the blkio.bfq.*
stats is rather costly, especially for some of the stats enabled by
CONFIG_BFQ_CGROUP_DEBUG.

Parameters to set
-----------------

For each group, there is only the following parameter to set.

weight (namely blkio.bfq.weight or io.bfq-weight): the weight of the
group inside its parent. Available values: 1..1000 (default 100). The
linear mapping between ioprio and weights, described at the beginning
of the tunable section, is still valid, but all weights higher than
IOPRIO_BE_NR*10 are mapped to ioprio 0.

Recall that, if low-latency is set, then BFQ automatically raises the
weight of the queues associated with interactive and soft real-time
applications. Unset this tunable if you need/want to control weights.


[1]
    P. Valente, A. Avanzini, "Evolution of the BFQ Storage I/O
    Scheduler", Proceedings of the First Workshop on Mobile System
    Technologies (MST-2015), May 2015.

    http://algogroup.unimore.it/people/paolo/disk_sched/mst-2015.pdf

[2]
    P. Valente and M. Andreolini, "Improving Application
    Responsiveness with the BFQ Disk I/O Scheduler", Proceedings of
    the 5th Annual International Systems and Storage Conference
    (SYSTOR '12), June 2012.

    Slightly extended version:

    http://algogroup.unimore.it/people/paolo/disk_sched/bfq-v1-suite-results.pdf

[3]
   https://github.com/Algodev-github/S