1 BFQ (Budget Fair Queueing)
2 ==========================
4 BFQ is a proportional-share I/O scheduler, with some extra
5 low-latency capabilities. In addition to cgroups support (blkio or io
6 controllers), BFQ's main features are:
7 - BFQ guarantees a high system and application responsiveness, and a
8 low latency for time-sensitive applications, such as audio or video
10 - BFQ distributes bandwidth, and not just time, among processes or
11 groups (switching back to time distribution when needed to keep
14 In its default configuration, BFQ privileges latency over
15 throughput. So, when needed for achieving a lower latency, BFQ builds
16 schedules that may lead to a lower throughput. If your main or only
17 goal, for a given device, is to achieve the maximum-possible
18 throughput at all times, then do switch off all low-latency heuristics
19 for that device, by setting low_latency to 0. See Section 3 for
20 details on how to configure BFQ for the desired tradeoff between
21 latency and throughput, or on how to maximize throughput.
23 BFQ has a non-null overhead, which limits the maximum IOPS that a CPU
24 can process for a device scheduled with BFQ. To give an idea of the
25 limits on slow or average CPUs, here are, first, the limits of BFQ for
26 three different CPUs, on, respectively, an average laptop, an old
27 desktop, and a cheap embedded system, in case full hierarchical
28 support is enabled (i.e., CONFIG_BFQ_GROUP_IOSCHED is set), but
29 CONFIG_DEBUG_BLK_CGROUP is not set (Section 4-2):
30 - Intel i7-4850HQ: 400 KIOPS
31 - AMD A8-3850: 250 KIOPS
32 - ARM CortexTM-A53 Octa-core: 80 KIOPS
34 If CONFIG_DEBUG_BLK_CGROUP is set (and of course full hierarchical
35 support is enabled), then the sustainable throughput with BFQ
36 decreases, because all blkio.bfq* statistics are created and updated
37 (Section 4-2). For BFQ, this leads to the following maximum
38 sustainable throughputs, on the same systems as above:
39 - Intel i7-4850HQ: 310 KIOPS
40 - AMD A8-3850: 200 KIOPS
41 - ARM CortexTM-A53 Octa-core: 56 KIOPS
43 BFQ works for multi-queue devices too.
45 The table of contents follow. Impatients can just jump to Section 3.
49 1. When may BFQ be useful?
53 3. What are BFQ's tunables and how to properly configure BFQ?
54 4. BFQ group scheduling
55 4-1 Service guarantees provided
58 1. When may BFQ be useful?
59 ==========================
61 BFQ provides the following benefits on personal and server systems.
66 Low latency for interactive applications
68 Regardless of the actual background workload, BFQ guarantees that, for
69 interactive tasks, the storage device is virtually as responsive as if
70 it was idle. For example, even if one or more of the following
71 background workloads are being executed:
72 - one or more large files are being read, written or copied,
73 - a tree of source files is being compiled,
74 - one or more virtual machines are performing I/O,
75 - a software update is in progress,
76 - indexing daemons are scanning filesystems and updating their
78 starting an application or loading a file from within an application
79 takes about the same time as if the storage device was idle. As a
80 comparison, with CFQ, NOOP or DEADLINE, and in the same conditions,
81 applications experience high latencies, or even become unresponsive
82 until the background workload terminates (also on SSDs).
84 Low latency for soft real-time applications
86 Also soft real-time applications, such as audio and video
87 players/streamers, enjoy a low latency and a low drop rate, regardless
88 of the background I/O workload. As a consequence, these applications
89 do not suffer from almost any glitch due to the background workload.
91 Higher speed for code-development tasks
93 If some additional workload happens to be executed in parallel, then
94 BFQ executes the I/O-related components of typical code-development
95 tasks (compilation, checkout, merge, ...) much more quickly than CFQ,
100 On hard disks, BFQ achieves up to 30% higher throughput than CFQ, and
101 up to 150% higher throughput than DEADLINE and NOOP, with all the
102 sequential workloads considered in our tests. With random workloads,
103 and with all the workloads on flash-based devices, BFQ achieves,
104 instead, about the same throughput as the other schedulers.
106 Strong fairness, bandwidth and delay guarantees
108 BFQ distributes the device throughput, and not just the device time,
109 among I/O-bound applications in proportion their weights, with any
110 workload and regardless of the device parameters. From these bandwidth
111 guarantees, it is possible to compute tight per-I/O-request delay
112 guarantees by a simple formula. If not configured for strict service
113 guarantees, BFQ switches to time-based resource sharing (only) for
114 applications that would otherwise cause a throughput loss.
119 Most benefits for server systems follow from the same service
120 properties as above. In particular, regardless of whether additional,
121 possibly heavy workloads are being served, BFQ guarantees:
123 . audio and video-streaming with zero or very low jitter and drop
126 . fast retrieval of WEB pages and embedded objects;
128 . real-time recording of data in live-dumping applications (e.g.,
131 . responsiveness in local and remote access to a server.
134 2. How does BFQ work?
135 =====================
137 BFQ is a proportional-share I/O scheduler, whose general structure,
138 plus a lot of code, are borrowed from CFQ.
140 - Each process doing I/O on a device is associated with a weight and a
143 - BFQ grants exclusive access to the device, for a while, to one queue
144 (process) at a time, and implements this service model by
145 associating every queue with a budget, measured in number of
148 - After a queue is granted access to the device, the budget of the
149 queue is decremented, on each request dispatch, by the size of the
152 - The in-service queue is expired, i.e., its service is suspended,
153 only if one of the following events occurs: 1) the queue finishes
154 its budget, 2) the queue empties, 3) a "budget timeout" fires.
156 - The budget timeout prevents processes doing random I/O from
157 holding the device for too long and dramatically reducing
160 - Actually, as in CFQ, a queue associated with a process issuing
161 sync requests may not be expired immediately when it empties. In
162 contrast, BFQ may idle the device for a short time interval,
163 giving the process the chance to go on being served if it issues
164 a new request in time. Device idling typically boosts the
165 throughput on rotational devices and on non-queueing flash-based
166 devices, if processes do synchronous and sequential I/O. In
167 addition, under BFQ, device idling is also instrumental in
168 guaranteeing the desired throughput fraction to processes
169 issuing sync requests (see the description of the slice_idle
170 tunable in this document, or [1, 2], for more details).
172 - With respect to idling for service guarantees, if several
173 processes are competing for the device at the same time, but
174 all processes and groups have the same weight, then BFQ
175 guarantees the expected throughput distribution without ever
176 idling the device. Throughput is thus as high as possible in
177 this common scenario.
179 - On flash-based storage with internal queueing of commands
180 (typically NCQ), device idling happens to be always detrimental
181 for throughput. So, with these devices, BFQ performs idling
182 only when strictly needed for service guarantees, i.e., for
183 guaranteeing low latency or fairness. In these cases, overall
184 throughput may be sub-optimal. No solution currently exists to
185 provide both strong service guarantees and optimal throughput
186 on devices with internal queueing.
188 - If low-latency mode is enabled (default configuration), BFQ
189 executes some special heuristics to detect interactive and soft
190 real-time applications (e.g., video or audio players/streamers),
191 and to reduce their latency. The most important action taken to
192 achieve this goal is to give to the queues associated with these
193 applications more than their fair share of the device
194 throughput. For brevity, we call just "weight-raising" the whole
195 sets of actions taken by BFQ to privilege these queues. In
196 particular, BFQ provides a milder form of weight-raising for
197 interactive applications, and a stronger form for soft real-time
200 - BFQ automatically deactivates idling for queues born in a burst of
201 queue creations. In fact, these queues are usually associated with
202 the processes of applications and services that benefit mostly
203 from a high throughput. Examples are systemd during boot, or git
206 - As CFQ, BFQ merges queues performing interleaved I/O, i.e.,
207 performing random I/O that becomes mostly sequential if
208 merged. Differently from CFQ, BFQ achieves this goal with a more
209 reactive mechanism, called Early Queue Merge (EQM). EQM is so
210 responsive in detecting interleaved I/O (cooperating processes),
211 that it enables BFQ to achieve a high throughput, by queue
212 merging, even for queues for which CFQ needs a different
213 mechanism, preemption, to get a high throughput. As such EQM is a
214 unified mechanism to achieve a high throughput with interleaved
217 - Queues are scheduled according to a variant of WF2Q+, named
218 B-WF2Q+, and implemented using an augmented rb-tree to preserve an
219 O(log N) overall complexity. See [2] for more details. B-WF2Q+ is
220 also ready for hierarchical scheduling, details in Section 4.
222 - B-WF2Q+ guarantees a tight deviation with respect to an ideal,
223 perfectly fair, and smooth service. In particular, B-WF2Q+
224 guarantees that each queue receives a fraction of the device
225 throughput proportional to its weight, even if the throughput
226 fluctuates, and regardless of: the device parameters, the current
227 workload and the budgets assigned to the queue.
229 - The last, budget-independence, property (although probably
230 counterintuitive in the first place) is definitely beneficial, for
231 the following reasons:
233 - First, with any proportional-share scheduler, the maximum
234 deviation with respect to an ideal service is proportional to
235 the maximum budget (slice) assigned to queues. As a consequence,
236 BFQ can keep this deviation tight not only because of the
237 accurate service of B-WF2Q+, but also because BFQ *does not*
238 need to assign a larger budget to a queue to let the queue
239 receive a higher fraction of the device throughput.
241 - Second, BFQ is free to choose, for every process (queue), the
242 budget that best fits the needs of the process, or best
243 leverages the I/O pattern of the process. In particular, BFQ
244 updates queue budgets with a simple feedback-loop algorithm that
245 allows a high throughput to be achieved, while still providing
246 tight latency guarantees to time-sensitive applications. When
247 the in-service queue expires, this algorithm computes the next
248 budget of the queue so as to:
250 - Let large budgets be eventually assigned to the queues
251 associated with I/O-bound applications performing sequential
252 I/O: in fact, the longer these applications are served once
253 got access to the device, the higher the throughput is.
255 - Let small budgets be eventually assigned to the queues
256 associated with time-sensitive applications (which typically
257 perform sporadic and short I/O), because, the smaller the
258 budget assigned to a queue waiting for service is, the sooner
259 B-WF2Q+ will serve that queue (Subsec 3.3 in [2]).
261 - If several processes are competing for the device at the same time,
262 but all processes and groups have the same weight, then BFQ
263 guarantees the expected throughput distribution without ever idling
264 the device. It uses preemption instead. Throughput is then much
265 higher in this common scenario.
267 - ioprio classes are served in strict priority order, i.e.,
268 lower-priority queues are not served as long as there are
269 higher-priority queues. Among queues in the same class, the
270 bandwidth is distributed in proportion to the weight of each
271 queue. A very thin extra bandwidth is however guaranteed to
272 the Idle class, to prevent it from starving.
275 3. What are BFQ's tunables and how to properly configure BFQ?
276 =============================================================
278 Most BFQ tunables affect service guarantees (basically latency and
279 fairness) and throughput. For full details on how to choose the
280 desired tradeoff between service guarantees and throughput, see the
281 parameters slice_idle, strict_guarantees and low_latency. For details
282 on how to maximise throughput, see slice_idle, timeout_sync and
283 max_budget. The other performance-related parameters have been
284 inherited from, and have been preserved mostly for compatibility with
285 CFQ. So far, no performance improvement has been reported after
286 changing the latter parameters in BFQ.
288 In particular, the tunables back_seek-max, back_seek_penalty,
289 fifo_expire_async and fifo_expire_sync below are the same as in
290 CFQ. Their description is just copied from that for CFQ. Some
291 considerations in the description of slice_idle are copied from CFQ
294 per-process ioprio and weight
295 -----------------------------
297 Unless the cgroups interface is used (see "4. BFQ group scheduling"),
298 weights can be assigned to processes only indirectly, through I/O
299 priorities, and according to the relation:
300 weight = (IOPRIO_BE_NR - ioprio) * 10.
302 Beware that, if low-latency is set, then BFQ automatically raises the
303 weight of the queues associated with interactive and soft real-time
304 applications. Unset this tunable if you need/want to control weights.
309 This parameter specifies how long BFQ should idle for next I/O
310 request, when certain sync BFQ queues become empty. By default
311 slice_idle is a non-zero value. Idling has a double purpose: boosting
312 throughput and making sure that the desired throughput distribution is
313 respected (see the description of how BFQ works, and, if needed, the
314 papers referred there).
316 As for throughput, idling can be very helpful on highly seeky media
317 like single spindle SATA/SAS disks where we can cut down on overall
318 number of seeks and see improved throughput.
320 Setting slice_idle to 0 will remove all the idling on queues and one
321 should see an overall improved throughput on faster storage devices
322 like multiple SATA/SAS disks in hardware RAID configuration, as well
323 as flash-based storage with internal command queueing (and
326 So depending on storage and workload, it might be useful to set
327 slice_idle=0. In general for SATA/SAS disks and software RAID of
328 SATA/SAS disks keeping slice_idle enabled should be useful. For any
329 configurations where there are multiple spindles behind single LUN
330 (Host based hardware RAID controller or for storage arrays), or with
331 flash-based fast storage, setting slice_idle=0 might end up in better
332 throughput and acceptable latencies.
334 Idling is however necessary to have service guarantees enforced in
335 case of differentiated weights or differentiated I/O-request lengths.
336 To see why, suppose that a given BFQ queue A must get several I/O
337 requests served for each request served for another queue B. Idling
338 ensures that, if A makes a new I/O request slightly after becoming
339 empty, then no request of B is dispatched in the middle, and thus A
340 does not lose the possibility to get more than one request dispatched
341 before the next request of B is dispatched. Note that idling
342 guarantees the desired differentiated treatment of queues only in
343 terms of I/O-request dispatches. To guarantee that the actual service
344 order then corresponds to the dispatch order, the strict_guarantees
345 tunable must be set too.
347 There is an important flipside for idling: apart from the above cases
348 where it is beneficial also for throughput, idling can severely impact
349 throughput. One important case is random workload. Because of this
350 issue, BFQ tends to avoid idling as much as possible, when it is not
351 beneficial also for throughput (as detailed in Section 2). As a
352 consequence of this behavior, and of further issues described for the
353 strict_guarantees tunable, short-term service guarantees may be
354 occasionally violated. And, in some cases, these guarantees may be
355 more important than guaranteeing maximum throughput. For example, in
356 video playing/streaming, a very low drop rate may be more important
357 than maximum throughput. In these cases, consider setting the
358 strict_guarantees parameter.
363 If this parameter is set (default: unset), then BFQ
365 - always performs idling when the in-service queue becomes empty;
367 - forces the device to serve one I/O request at a time, by dispatching a
368 new request only if there is no outstanding request.
370 In the presence of differentiated weights or I/O-request sizes, both
371 the above conditions are needed to guarantee that every BFQ queue
372 receives its allotted share of the bandwidth. The first condition is
373 needed for the reasons explained in the description of the slice_idle
374 tunable. The second condition is needed because all modern storage
375 devices reorder internally-queued requests, which may trivially break
376 the service guarantees enforced by the I/O scheduler.
378 Setting strict_guarantees may evidently affect throughput.
383 This specifies, given in Kbytes, the maximum "distance" for backward seeking.
384 The distance is the amount of space from the current head location to the
385 sectors that are backward in terms of distance.
387 This parameter allows the scheduler to anticipate requests in the "backward"
388 direction and consider them as being the "next" if they are within this
389 distance from the current head location.
394 This parameter is used to compute the cost of backward seeking. If the
395 backward distance of request is just 1/back_seek_penalty from a "front"
396 request, then the seeking cost of two requests is considered equivalent.
398 So scheduler will not bias toward one or the other request (otherwise scheduler
399 will bias toward front request). Default value of back_seek_penalty is 2.
404 This parameter is used to set the timeout of asynchronous requests. Default
405 value of this is 248ms.
410 This parameter is used to set the timeout of synchronous requests. Default
411 value of this is 124ms. In case to favor synchronous requests over asynchronous
412 one, this value should be decreased relative to fifo_expire_async.
417 This parameter is used to enable/disable BFQ's low latency mode. By
418 default, low latency mode is enabled. If enabled, interactive and soft
419 real-time applications are privileged and experience a lower latency,
420 as explained in more detail in the description of how BFQ works.
422 DISABLE this mode if you need full control on bandwidth
423 distribution. In fact, if it is enabled, then BFQ automatically
424 increases the bandwidth share of privileged applications, as the main
425 means to guarantee a lower latency to them.
427 In addition, as already highlighted at the beginning of this document,
428 DISABLE this mode if your only goal is to achieve a high throughput.
429 In fact, privileging the I/O of some application over the rest may
430 entail a lower throughput. To achieve the highest-possible throughput
431 on a non-rotational device, setting slice_idle to 0 may be needed too
432 (at the cost of giving up any strong guarantee on fairness and low
438 Maximum amount of device time that can be given to a task (queue) once
439 it has been selected for service. On devices with costly seeks,
440 increasing this time usually increases maximum throughput. On the
441 opposite end, increasing this time coarsens the granularity of the
442 short-term bandwidth and latency guarantees, especially if the
443 following parameter is set to zero.
448 Maximum amount of service, measured in sectors, that can be provided
449 to a BFQ queue once it is set in service (of course within the limits
450 of the above timeout). According to what said in the description of
451 the algorithm, larger values increase the throughput in proportion to
452 the percentage of sequential I/O requests issued. The price of larger
453 values is that they coarsen the granularity of short-term bandwidth
454 and latency guarantees.
456 The default value is 0, which enables auto-tuning: BFQ sets max_budget
457 to the maximum number of sectors that can be served during
458 timeout_sync, according to the estimated peak rate.
460 For specific devices, some users have occasionally reported to have
461 reached a higher throughput by setting max_budget explicitly, i.e., by
462 setting max_budget to a higher value than 0. In particular, they have
463 set max_budget to higher values than those to which BFQ would have set
464 it with auto-tuning. An alternative way to achieve this goal is to
465 just increase the value of timeout_sync, leaving max_budget equal to 0.
470 Read-only parameter, used to show the weights of the currently active
474 4. Group scheduling with BFQ
475 ============================
477 BFQ supports both cgroups-v1 and cgroups-v2 io controllers, namely
478 blkio and io. In particular, BFQ supports weight-based proportional
479 share. To activate cgroups support, set BFQ_GROUP_IOSCHED.
481 4-1 Service guarantees provided
482 -------------------------------
484 With BFQ, proportional share means true proportional share of the
485 device bandwidth, according to group weights. For example, a group
486 with weight 200 gets twice the bandwidth, and not just twice the time,
487 of a group with weight 100.
489 BFQ supports hierarchies (group trees) of any depth. Bandwidth is
490 distributed among groups and processes in the expected way: for each
491 group, the children of the group share the whole bandwidth of the
492 group in proportion to their weights. In particular, this implies
493 that, for each leaf group, every process of the group receives the
494 same share of the whole group bandwidth, unless the ioprio of the
497 The resource-sharing guarantee for a group may partially or totally
498 switch from bandwidth to time, if providing bandwidth guarantees to
499 the group lowers the throughput too much. This switch occurs on a
500 per-process basis: if a process of a leaf group causes throughput loss
501 if served in such a way to receive its share of the bandwidth, then
502 BFQ switches back to just time-based proportional share for that
508 To get proportional sharing of bandwidth with BFQ for a given device,
509 BFQ must of course be the active scheduler for that device.
511 Within each group directory, the names of the files associated with
512 BFQ-specific cgroup parameters and stats begin with the "bfq."
513 prefix. So, with cgroups-v1 or cgroups-v2, the full prefix for
514 BFQ-specific files is "blkio.bfq." or "io.bfq." For example, the group
515 parameter to set the weight of a group with BFQ is blkio.bfq.weight
518 As for cgroups-v1 (blkio controller), the exact set of stat files
519 created, and kept up-to-date by bfq, depends on whether
520 CONFIG_DEBUG_BLK_CGROUP is set. If it is set, then bfq creates all
521 the stat files documented in
522 Documentation/cgroup-v1/blkio-controller.txt. If, instead,
523 CONFIG_DEBUG_BLK_CGROUP is not set, then bfq creates only the files
524 blkio.bfq.io_service_bytes
525 blkio.bfq.io_service_bytes_recursive
526 blkio.bfq.io_serviced
527 blkio.bfq.io_serviced_recursive
529 The value of CONFIG_DEBUG_BLK_CGROUP greatly influences the maximum
530 throughput sustainable with bfq, because updating the blkio.bfq.*
531 stats is rather costly, especially for some of the stats enabled by
532 CONFIG_DEBUG_BLK_CGROUP.
537 For each group, there is only the following parameter to set.
539 weight (namely blkio.bfq.weight or io.bfq-weight): the weight of the
540 group inside its parent. Available values: 1..10000 (default 100). The
541 linear mapping between ioprio and weights, described at the beginning
542 of the tunable section, is still valid, but all weights higher than
543 IOPRIO_BE_NR*10 are mapped to ioprio 0.
545 Recall that, if low-latency is set, then BFQ automatically raises the
546 weight of the queues associated with interactive and soft real-time
547 applications. Unset this tunable if you need/want to control weights.
550 [1] P. Valente, A. Avanzini, "Evolution of the BFQ Storage I/O
551 Scheduler", Proceedings of the First Workshop on Mobile System
552 Technologies (MST-2015), May 2015.
553 http://algogroup.unimore.it/people/paolo/disk_sched/mst-2015.pdf
555 [2] P. Valente and M. Andreolini, "Improving Application
556 Responsiveness with the BFQ Disk I/O Scheduler", Proceedings of
557 the 5th Annual International Systems and Storage Conference
558 (SYSTOR '12), June 2012.
559 Slightly extended version:
560 http://algogroup.unimore.it/people/paolo/disk_sched/bfq-v1-suite-
projects / releases.git / blob