1 ==========================
2 BFQ (Budget Fair Queueing)
3 ==========================
5 BFQ is a proportional-share I/O scheduler, with some extra
6 low-latency capabilities. In addition to cgroups support (blkio or io
7 controllers), BFQ's main features are:
9 - BFQ guarantees a high system and application responsiveness, and a
10 low latency for time-sensitive applications, such as audio or video
12 - BFQ distributes bandwidth, and not just time, among processes or
13 groups (switching back to time distribution when needed to keep
16 In its default configuration, BFQ privileges latency over
17 throughput. So, when needed for achieving a lower latency, BFQ builds
18 schedules that may lead to a lower throughput. If your main or only
19 goal, for a given device, is to achieve the maximum-possible
20 throughput at all times, then do switch off all low-latency heuristics
21 for that device, by setting low_latency to 0. See Section 3 for
22 details on how to configure BFQ for the desired tradeoff between
23 latency and throughput, or on how to maximize throughput.
25 As every I/O scheduler, BFQ adds some overhead to per-I/O-request
26 processing. To give an idea of this overhead, the total,
27 single-lock-protected, per-request processing time of BFQ---i.e., the
28 sum of the execution times of the request insertion, dispatch and
29 completion hooks---is, e.g., 1.9 us on an Intel Core i7-2760QM@2.40GHz
30 (dated CPU for notebooks; time measured with simple code
31 instrumentation, and using the throughput-sync.sh script of the S
32 suite [1], in performance-profiling mode). To put this result into
33 context, the total, single-lock-protected, per-request execution time
34 of the lightest I/O scheduler available in blk-mq, mq-deadline, is 0.7
35 us (mq-deadline is ~800 LOC, against ~10500 LOC for BFQ).
37 Scheduling overhead further limits the maximum IOPS that a CPU can
38 process (already limited by the execution of the rest of the I/O
39 stack). To give an idea of the limits with BFQ, on slow or average
40 CPUs, here are, first, the limits of BFQ for three different CPUs, on,
41 respectively, an average laptop, an old desktop, and a cheap embedded
42 system, in case full hierarchical support is enabled (i.e.,
43 CONFIG_BFQ_GROUP_IOSCHED is set), but CONFIG_BFQ_CGROUP_DEBUG is not
45 - Intel i7-4850HQ: 400 KIOPS
46 - AMD A8-3850: 250 KIOPS
47 - ARM CortexTM-A53 Octa-core: 80 KIOPS
49 If CONFIG_BFQ_CGROUP_DEBUG is set (and of course full hierarchical
50 support is enabled), then the sustainable throughput with BFQ
51 decreases, because all blkio.bfq* statistics are created and updated
52 (Section 4-2). For BFQ, this leads to the following maximum
53 sustainable throughputs, on the same systems as above:
54 - Intel i7-4850HQ: 310 KIOPS
55 - AMD A8-3850: 200 KIOPS
56 - ARM CortexTM-A53 Octa-core: 56 KIOPS
58 BFQ works for multi-queue devices too.
60 .. The table of contents follow. Impatients can just jump to Section 3.
64 1. When may BFQ be useful?
68 3. What are BFQ's tunables and how to properly configure BFQ?
69 4. BFQ group scheduling
70 4-1 Service guarantees provided
73 1. When may BFQ be useful?
74 ==========================
76 BFQ provides the following benefits on personal and server systems.
81 Low latency for interactive applications
82 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
84 Regardless of the actual background workload, BFQ guarantees that, for
85 interactive tasks, the storage device is virtually as responsive as if
86 it was idle. For example, even if one or more of the following
87 background workloads are being executed:
89 - one or more large files are being read, written or copied,
90 - a tree of source files is being compiled,
91 - one or more virtual machines are performing I/O,
92 - a software update is in progress,
93 - indexing daemons are scanning filesystems and updating their
96 starting an application or loading a file from within an application
97 takes about the same time as if the storage device was idle. As a
98 comparison, with CFQ, NOOP or DEADLINE, and in the same conditions,
99 applications experience high latencies, or even become unresponsive
100 until the background workload terminates (also on SSDs).
102 Low latency for soft real-time applications
103 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
104 Also soft real-time applications, such as audio and video
105 players/streamers, enjoy a low latency and a low drop rate, regardless
106 of the background I/O workload. As a consequence, these applications
107 do not suffer from almost any glitch due to the background workload.
109 Higher speed for code-development tasks
110 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
112 If some additional workload happens to be executed in parallel, then
113 BFQ executes the I/O-related components of typical code-development
114 tasks (compilation, checkout, merge, ...) much more quickly than CFQ,
120 On hard disks, BFQ achieves up to 30% higher throughput than CFQ, and
121 up to 150% higher throughput than DEADLINE and NOOP, with all the
122 sequential workloads considered in our tests. With random workloads,
123 and with all the workloads on flash-based devices, BFQ achieves,
124 instead, about the same throughput as the other schedulers.
126 Strong fairness, bandwidth and delay guarantees
127 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
129 BFQ distributes the device throughput, and not just the device time,
130 among I/O-bound applications in proportion their weights, with any
131 workload and regardless of the device parameters. From these bandwidth
132 guarantees, it is possible to compute tight per-I/O-request delay
133 guarantees by a simple formula. If not configured for strict service
134 guarantees, BFQ switches to time-based resource sharing (only) for
135 applications that would otherwise cause a throughput loss.
140 Most benefits for server systems follow from the same service
141 properties as above. In particular, regardless of whether additional,
142 possibly heavy workloads are being served, BFQ guarantees:
144 * audio and video-streaming with zero or very low jitter and drop
147 * fast retrieval of WEB pages and embedded objects;
149 * real-time recording of data in live-dumping applications (e.g.,
152 * responsiveness in local and remote access to a server.
155 2. How does BFQ work?
156 =====================
158 BFQ is a proportional-share I/O scheduler, whose general structure,
159 plus a lot of code, are borrowed from CFQ.
161 - Each process doing I/O on a device is associated with a weight and a
164 - BFQ grants exclusive access to the device, for a while, to one queue
165 (process) at a time, and implements this service model by
166 associating every queue with a budget, measured in number of
169 - After a queue is granted access to the device, the budget of the
170 queue is decremented, on each request dispatch, by the size of the
173 - The in-service queue is expired, i.e., its service is suspended,
174 only if one of the following events occurs: 1) the queue finishes
175 its budget, 2) the queue empties, 3) a "budget timeout" fires.
177 - The budget timeout prevents processes doing random I/O from
178 holding the device for too long and dramatically reducing
181 - Actually, as in CFQ, a queue associated with a process issuing
182 sync requests may not be expired immediately when it empties. In
183 contrast, BFQ may idle the device for a short time interval,
184 giving the process the chance to go on being served if it issues
185 a new request in time. Device idling typically boosts the
186 throughput on rotational devices and on non-queueing flash-based
187 devices, if processes do synchronous and sequential I/O. In
188 addition, under BFQ, device idling is also instrumental in
189 guaranteeing the desired throughput fraction to processes
190 issuing sync requests (see the description of the slice_idle
191 tunable in this document, or [1, 2], for more details).
193 - With respect to idling for service guarantees, if several
194 processes are competing for the device at the same time, but
195 all processes and groups have the same weight, then BFQ
196 guarantees the expected throughput distribution without ever
197 idling the device. Throughput is thus as high as possible in
198 this common scenario.
200 - On flash-based storage with internal queueing of commands
201 (typically NCQ), device idling happens to be always detrimental
202 for throughput. So, with these devices, BFQ performs idling
203 only when strictly needed for service guarantees, i.e., for
204 guaranteeing low latency or fairness. In these cases, overall
205 throughput may be sub-optimal. No solution currently exists to
206 provide both strong service guarantees and optimal throughput
207 on devices with internal queueing.
209 - If low-latency mode is enabled (default configuration), BFQ
210 executes some special heuristics to detect interactive and soft
211 real-time applications (e.g., video or audio players/streamers),
212 and to reduce their latency. The most important action taken to
213 achieve this goal is to give to the queues associated with these
214 applications more than their fair share of the device
215 throughput. For brevity, we call just "weight-raising" the whole
216 sets of actions taken by BFQ to privilege these queues. In
217 particular, BFQ provides a milder form of weight-raising for
218 interactive applications, and a stronger form for soft real-time
221 - BFQ automatically deactivates idling for queues born in a burst of
222 queue creations. In fact, these queues are usually associated with
223 the processes of applications and services that benefit mostly
224 from a high throughput. Examples are systemd during boot, or git
227 - As CFQ, BFQ merges queues performing interleaved I/O, i.e.,
228 performing random I/O that becomes mostly sequential if
229 merged. Differently from CFQ, BFQ achieves this goal with a more
230 reactive mechanism, called Early Queue Merge (EQM). EQM is so
231 responsive in detecting interleaved I/O (cooperating processes),
232 that it enables BFQ to achieve a high throughput, by queue
233 merging, even for queues for which CFQ needs a different
234 mechanism, preemption, to get a high throughput. As such EQM is a
235 unified mechanism to achieve a high throughput with interleaved
238 - Queues are scheduled according to a variant of WF2Q+, named
239 B-WF2Q+, and implemented using an augmented rb-tree to preserve an
240 O(log N) overall complexity. See [2] for more details. B-WF2Q+ is
241 also ready for hierarchical scheduling, details in Section 4.
243 - B-WF2Q+ guarantees a tight deviation with respect to an ideal,
244 perfectly fair, and smooth service. In particular, B-WF2Q+
245 guarantees that each queue receives a fraction of the device
246 throughput proportional to its weight, even if the throughput
247 fluctuates, and regardless of: the device parameters, the current
248 workload and the budgets assigned to the queue.
250 - The last, budget-independence, property (although probably
251 counterintuitive in the first place) is definitely beneficial, for
252 the following reasons:
254 - First, with any proportional-share scheduler, the maximum
255 deviation with respect to an ideal service is proportional to
256 the maximum budget (slice) assigned to queues. As a consequence,
257 BFQ can keep this deviation tight not only because of the
258 accurate service of B-WF2Q+, but also because BFQ *does not*
259 need to assign a larger budget to a queue to let the queue
260 receive a higher fraction of the device throughput.
262 - Second, BFQ is free to choose, for every process (queue), the
263 budget that best fits the needs of the process, or best
264 leverages the I/O pattern of the process. In particular, BFQ
265 updates queue budgets with a simple feedback-loop algorithm that
266 allows a high throughput to be achieved, while still providing
267 tight latency guarantees to time-sensitive applications. When
268 the in-service queue expires, this algorithm computes the next
269 budget of the queue so as to:
271 - Let large budgets be eventually assigned to the queues
272 associated with I/O-bound applications performing sequential
273 I/O: in fact, the longer these applications are served once
274 got access to the device, the higher the throughput is.
276 - Let small budgets be eventually assigned to the queues
277 associated with time-sensitive applications (which typically
278 perform sporadic and short I/O), because, the smaller the
279 budget assigned to a queue waiting for service is, the sooner
280 B-WF2Q+ will serve that queue (Subsec 3.3 in [2]).
282 - If several processes are competing for the device at the same time,
283 but all processes and groups have the same weight, then BFQ
284 guarantees the expected throughput distribution without ever idling
285 the device. It uses preemption instead. Throughput is then much
286 higher in this common scenario.
288 - ioprio classes are served in strict priority order, i.e.,
289 lower-priority queues are not served as long as there are
290 higher-priority queues. Among queues in the same class, the
291 bandwidth is distributed in proportion to the weight of each
292 queue. A very thin extra bandwidth is however guaranteed to
293 the Idle class, to prevent it from starving.
296 3. What are BFQ's tunables and how to properly configure BFQ?
297 =============================================================
299 Most BFQ tunables affect service guarantees (basically latency and
300 fairness) and throughput. For full details on how to choose the
301 desired tradeoff between service guarantees and throughput, see the
302 parameters slice_idle, strict_guarantees and low_latency. For details
303 on how to maximise throughput, see slice_idle, timeout_sync and
304 max_budget. The other performance-related parameters have been
305 inherited from, and have been preserved mostly for compatibility with
306 CFQ. So far, no performance improvement has been reported after
307 changing the latter parameters in BFQ.
309 In particular, the tunables back_seek-max, back_seek_penalty,
310 fifo_expire_async and fifo_expire_sync below are the same as in
311 CFQ. Their description is just copied from that for CFQ. Some
312 considerations in the description of slice_idle are copied from CFQ
315 per-process ioprio and weight
316 -----------------------------
318 Unless the cgroups interface is used (see "4. BFQ group scheduling"),
319 weights can be assigned to processes only indirectly, through I/O
320 priorities, and according to the relation:
321 weight = (IOPRIO_BE_NR - ioprio) * 10.
323 Beware that, if low-latency is set, then BFQ automatically raises the
324 weight of the queues associated with interactive and soft real-time
325 applications. Unset this tunable if you need/want to control weights.
330 This parameter specifies how long BFQ should idle for next I/O
331 request, when certain sync BFQ queues become empty. By default
332 slice_idle is a non-zero value. Idling has a double purpose: boosting
333 throughput and making sure that the desired throughput distribution is
334 respected (see the description of how BFQ works, and, if needed, the
335 papers referred there).
337 As for throughput, idling can be very helpful on highly seeky media
338 like single spindle SATA/SAS disks where we can cut down on overall
339 number of seeks and see improved throughput.
341 Setting slice_idle to 0 will remove all the idling on queues and one
342 should see an overall improved throughput on faster storage devices
343 like multiple SATA/SAS disks in hardware RAID configuration, as well
344 as flash-based storage with internal command queueing (and
347 So depending on storage and workload, it might be useful to set
348 slice_idle=0. In general for SATA/SAS disks and software RAID of
349 SATA/SAS disks keeping slice_idle enabled should be useful. For any
350 configurations where there are multiple spindles behind single LUN
351 (Host based hardware RAID controller or for storage arrays), or with
352 flash-based fast storage, setting slice_idle=0 might end up in better
353 throughput and acceptable latencies.
355 Idling is however necessary to have service guarantees enforced in
356 case of differentiated weights or differentiated I/O-request lengths.
357 To see why, suppose that a given BFQ queue A must get several I/O
358 requests served for each request served for another queue B. Idling
359 ensures that, if A makes a new I/O request slightly after becoming
360 empty, then no request of B is dispatched in the middle, and thus A
361 does not lose the possibility to get more than one request dispatched
362 before the next request of B is dispatched. Note that idling
363 guarantees the desired differentiated treatment of queues only in
364 terms of I/O-request dispatches. To guarantee that the actual service
365 order then corresponds to the dispatch order, the strict_guarantees
366 tunable must be set too.
368 There is an important flipside for idling: apart from the above cases
369 where it is beneficial also for throughput, idling can severely impact
370 throughput. One important case is random workload. Because of this
371 issue, BFQ tends to avoid idling as much as possible, when it is not
372 beneficial also for throughput (as detailed in Section 2). As a
373 consequence of this behavior, and of further issues described for the
374 strict_guarantees tunable, short-term service guarantees may be
375 occasionally violated. And, in some cases, these guarantees may be
376 more important than guaranteeing maximum throughput. For example, in
377 video playing/streaming, a very low drop rate may be more important
378 than maximum throughput. In these cases, consider setting the
379 strict_guarantees parameter.
384 Controls the same tuning parameter as slice_idle, but in microseconds.
385 Either tunable can be used to set idling behavior. Afterwards, the
386 other tunable will reflect the newly set value in sysfs.
391 If this parameter is set (default: unset), then BFQ
393 - always performs idling when the in-service queue becomes empty;
395 - forces the device to serve one I/O request at a time, by dispatching a
396 new request only if there is no outstanding request.
398 In the presence of differentiated weights or I/O-request sizes, both
399 the above conditions are needed to guarantee that every BFQ queue
400 receives its allotted share of the bandwidth. The first condition is
401 needed for the reasons explained in the description of the slice_idle
402 tunable. The second condition is needed because all modern storage
403 devices reorder internally-queued requests, which may trivially break
404 the service guarantees enforced by the I/O scheduler.
406 Setting strict_guarantees may evidently affect throughput.
411 This specifies, given in Kbytes, the maximum "distance" for backward seeking.
412 The distance is the amount of space from the current head location to the
413 sectors that are backward in terms of distance.
415 This parameter allows the scheduler to anticipate requests in the "backward"
416 direction and consider them as being the "next" if they are within this
417 distance from the current head location.
422 This parameter is used to compute the cost of backward seeking. If the
423 backward distance of request is just 1/back_seek_penalty from a "front"
424 request, then the seeking cost of two requests is considered equivalent.
426 So scheduler will not bias toward one or the other request (otherwise scheduler
427 will bias toward front request). Default value of back_seek_penalty is 2.
432 This parameter is used to set the timeout of asynchronous requests. Default
433 value of this is 250ms.
438 This parameter is used to set the timeout of synchronous requests. Default
439 value of this is 125ms. In case to favor synchronous requests over asynchronous
440 one, this value should be decreased relative to fifo_expire_async.
445 This parameter is used to enable/disable BFQ's low latency mode. By
446 default, low latency mode is enabled. If enabled, interactive and soft
447 real-time applications are privileged and experience a lower latency,
448 as explained in more detail in the description of how BFQ works.
450 DISABLE this mode if you need full control on bandwidth
451 distribution. In fact, if it is enabled, then BFQ automatically
452 increases the bandwidth share of privileged applications, as the main
453 means to guarantee a lower latency to them.
455 In addition, as already highlighted at the beginning of this document,
456 DISABLE this mode if your only goal is to achieve a high throughput.
457 In fact, privileging the I/O of some application over the rest may
458 entail a lower throughput. To achieve the highest-possible throughput
459 on a non-rotational device, setting slice_idle to 0 may be needed too
460 (at the cost of giving up any strong guarantee on fairness and low
466 Maximum amount of device time that can be given to a task (queue) once
467 it has been selected for service. On devices with costly seeks,
468 increasing this time usually increases maximum throughput. On the
469 opposite end, increasing this time coarsens the granularity of the
470 short-term bandwidth and latency guarantees, especially if the
471 following parameter is set to zero.
476 Maximum amount of service, measured in sectors, that can be provided
477 to a BFQ queue once it is set in service (of course within the limits
478 of the above timeout). According to what said in the description of
479 the algorithm, larger values increase the throughput in proportion to
480 the percentage of sequential I/O requests issued. The price of larger
481 values is that they coarsen the granularity of short-term bandwidth
482 and latency guarantees.
484 The default value is 0, which enables auto-tuning: BFQ sets max_budget
485 to the maximum number of sectors that can be served during
486 timeout_sync, according to the estimated peak rate.
488 For specific devices, some users have occasionally reported to have
489 reached a higher throughput by setting max_budget explicitly, i.e., by
490 setting max_budget to a higher value than 0. In particular, they have
491 set max_budget to higher values than those to which BFQ would have set
492 it with auto-tuning. An alternative way to achieve this goal is to
493 just increase the value of timeout_sync, leaving max_budget equal to 0.
495 4. Group scheduling with BFQ
496 ============================
498 BFQ supports both cgroups-v1 and cgroups-v2 io controllers, namely
499 blkio and io. In particular, BFQ supports weight-based proportional
500 share. To activate cgroups support, set BFQ_GROUP_IOSCHED.
502 4-1 Service guarantees provided
503 -------------------------------
505 With BFQ, proportional share means true proportional share of the
506 device bandwidth, according to group weights. For example, a group
507 with weight 200 gets twice the bandwidth, and not just twice the time,
508 of a group with weight 100.
510 BFQ supports hierarchies (group trees) of any depth. Bandwidth is
511 distributed among groups and processes in the expected way: for each
512 group, the children of the group share the whole bandwidth of the
513 group in proportion to their weights. In particular, this implies
514 that, for each leaf group, every process of the group receives the
515 same share of the whole group bandwidth, unless the ioprio of the
518 The resource-sharing guarantee for a group may partially or totally
519 switch from bandwidth to time, if providing bandwidth guarantees to
520 the group lowers the throughput too much. This switch occurs on a
521 per-process basis: if a process of a leaf group causes throughput loss
522 if served in such a way to receive its share of the bandwidth, then
523 BFQ switches back to just time-based proportional share for that
529 To get proportional sharing of bandwidth with BFQ for a given device,
530 BFQ must of course be the active scheduler for that device.
532 Within each group directory, the names of the files associated with
533 BFQ-specific cgroup parameters and stats begin with the "bfq."
534 prefix. So, with cgroups-v1 or cgroups-v2, the full prefix for
535 BFQ-specific files is "blkio.bfq." or "io.bfq." For example, the group
536 parameter to set the weight of a group with BFQ is blkio.bfq.weight
539 As for cgroups-v1 (blkio controller), the exact set of stat files
540 created, and kept up-to-date by bfq, depends on whether
541 CONFIG_BFQ_CGROUP_DEBUG is set. If it is set, then bfq creates all
542 the stat files documented in
543 Documentation/admin-guide/cgroup-v1/blkio-controller.rst. If, instead,
544 CONFIG_BFQ_CGROUP_DEBUG is not set, then bfq creates only the files::
546 blkio.bfq.io_service_bytes
547 blkio.bfq.io_service_bytes_recursive
548 blkio.bfq.io_serviced
549 blkio.bfq.io_serviced_recursive
551 The value of CONFIG_BFQ_CGROUP_DEBUG greatly influences the maximum
552 throughput sustainable with bfq, because updating the blkio.bfq.*
553 stats is rather costly, especially for some of the stats enabled by
554 CONFIG_BFQ_CGROUP_DEBUG.
559 For each group, the following parameters can be set:
562 This specifies the default weight for the cgroup inside its parent.
563 Available values: 1..1000 (default: 100).
565 For cgroup v1, it is set by writing the value to `blkio.bfq.weight`.
567 For cgroup v2, it is set by writing the value to `io.bfq.weight`.
568 (with an optional prefix of `default` and a space).
570 The linear mapping between ioprio and weights, described at the beginning
571 of the tunable section, is still valid, but all weights higher than
572 IOPRIO_BE_NR*10 are mapped to ioprio 0.
574 Recall that, if low-latency is set, then BFQ automatically raises the
575 weight of the queues associated with interactive and soft real-time
576 applications. Unset this tunable if you need/want to control weights.
579 This specifies a per-device weight for the cgroup. The syntax is
580 `minor:major weight`. A weight of `0` may be used to reset to the default
583 For cgroup v1, it is set by writing the value to `blkio.bfq.weight_device`.
585 For cgroup v2, the file name is `io.bfq.weight`.
589 P. Valente, A. Avanzini, "Evolution of the BFQ Storage I/O
590 Scheduler", Proceedings of the First Workshop on Mobile System
591 Technologies (MST-2015), May 2015.
593 http://algogroup.unimore.it/people/paolo/disk_sched/mst-2015.pdf
596 P. Valente and M. Andreolini, "Improving Application
597 Responsiveness with the BFQ Disk I/O Scheduler", Proceedings of
598 the 5th Annual International Systems and Storage Conference
599 (SYSTOR '12), June 2012.
601 Slightly extended version:
603 http://algogroup.unimore.it/people/paolo/disk_sched/bfq-v1-suite-results.pdf
606 https://github.com/Algodev-github/S
projects / linux-2.6-microblaze.git / blob