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 On average CPUs, the current version of BFQ can handle devices
15 performing at most ~30K IOPS; at most ~50 KIOPS on faster CPUs. As a
16 reference, 30-50 KIOPS correspond to very high bandwidths with
17 sequential I/O (e.g., 8-12 GB/s if I/O requests are 256 KB large), and
18 to 120-200 MB/s with 4KB random I/O. BFQ has not yet been tested on
21 The table of contents follow. Impatients can just jump to Section 3.
25 1. When may BFQ be useful?
29 3. What are BFQ's tunable?
30 4. BFQ group scheduling
31 4-1 Service guarantees provided
34 1. When may BFQ be useful?
35 ==========================
37 BFQ provides the following benefits on personal and server systems.
42 Low latency for interactive applications
44 Regardless of the actual background workload, BFQ guarantees that, for
45 interactive tasks, the storage device is virtually as responsive as if
46 it was idle. For example, even if one or more of the following
47 background workloads are being executed:
48 - one or more large files are being read, written or copied,
49 - a tree of source files is being compiled,
50 - one or more virtual machines are performing I/O,
51 - a software update is in progress,
52 - indexing daemons are scanning filesystems and updating their
54 starting an application or loading a file from within an application
55 takes about the same time as if the storage device was idle. As a
56 comparison, with CFQ, NOOP or DEADLINE, and in the same conditions,
57 applications experience high latencies, or even become unresponsive
58 until the background workload terminates (also on SSDs).
60 Low latency for soft real-time applications
62 Also soft real-time applications, such as audio and video
63 players/streamers, enjoy a low latency and a low drop rate, regardless
64 of the background I/O workload. As a consequence, these applications
65 do not suffer from almost any glitch due to the background workload.
67 Higher speed for code-development tasks
69 If some additional workload happens to be executed in parallel, then
70 BFQ executes the I/O-related components of typical code-development
71 tasks (compilation, checkout, merge, ...) much more quickly than CFQ,
76 On hard disks, BFQ achieves up to 30% higher throughput than CFQ, and
77 up to 150% higher throughput than DEADLINE and NOOP, with all the
78 sequential workloads considered in our tests. With random workloads,
79 and with all the workloads on flash-based devices, BFQ achieves,
80 instead, about the same throughput as the other schedulers.
82 Strong fairness, bandwidth and delay guarantees
84 BFQ distributes the device throughput, and not just the device time,
85 among I/O-bound applications in proportion their weights, with any
86 workload and regardless of the device parameters. From these bandwidth
87 guarantees, it is possible to compute tight per-I/O-request delay
88 guarantees by a simple formula. If not configured for strict service
89 guarantees, BFQ switches to time-based resource sharing (only) for
90 applications that would otherwise cause a throughput loss.
95 Most benefits for server systems follow from the same service
96 properties as above. In particular, regardless of whether additional,
97 possibly heavy workloads are being served, BFQ guarantees:
99 . audio and video-streaming with zero or very low jitter and drop
102 . fast retrieval of WEB pages and embedded objects;
104 . real-time recording of data in live-dumping applications (e.g.,
107 . responsiveness in local and remote access to a server.
110 2. How does BFQ work?
111 =====================
113 BFQ is a proportional-share I/O scheduler, whose general structure,
114 plus a lot of code, are borrowed from CFQ.
116 - Each process doing I/O on a device is associated with a weight and a
119 - BFQ grants exclusive access to the device, for a while, to one queue
120 (process) at a time, and implements this service model by
121 associating every queue with a budget, measured in number of
124 - After a queue is granted access to the device, the budget of the
125 queue is decremented, on each request dispatch, by the size of the
128 - The in-service queue is expired, i.e., its service is suspended,
129 only if one of the following events occurs: 1) the queue finishes
130 its budget, 2) the queue empties, 3) a "budget timeout" fires.
132 - The budget timeout prevents processes doing random I/O from
133 holding the device for too long and dramatically reducing
136 - Actually, as in CFQ, a queue associated with a process issuing
137 sync requests may not be expired immediately when it empties. In
138 contrast, BFQ may idle the device for a short time interval,
139 giving the process the chance to go on being served if it issues
140 a new request in time. Device idling typically boosts the
141 throughput on rotational devices, if processes do synchronous
142 and sequential I/O. In addition, under BFQ, device idling is
143 also instrumental in guaranteeing the desired throughput
144 fraction to processes issuing sync requests (see the description
145 of the slice_idle tunable in this document, or [1, 2], for more
148 - With respect to idling for service guarantees, if several
149 processes are competing for the device at the same time, but
150 all processes (and groups, after the following commit) have
151 the same weight, then BFQ guarantees the expected throughput
152 distribution without ever idling the device. Throughput is
153 thus as high as possible in this common scenario.
155 - If low-latency mode is enabled (default configuration), BFQ
156 executes some special heuristics to detect interactive and soft
157 real-time applications (e.g., video or audio players/streamers),
158 and to reduce their latency. The most important action taken to
159 achieve this goal is to give to the queues associated with these
160 applications more than their fair share of the device
161 throughput. For brevity, we call just "weight-raising" the whole
162 sets of actions taken by BFQ to privilege these queues. In
163 particular, BFQ provides a milder form of weight-raising for
164 interactive applications, and a stronger form for soft real-time
167 - BFQ automatically deactivates idling for queues born in a burst of
168 queue creations. In fact, these queues are usually associated with
169 the processes of applications and services that benefit mostly
170 from a high throughput. Examples are systemd during boot, or git
173 - As CFQ, BFQ merges queues performing interleaved I/O, i.e.,
174 performing random I/O that becomes mostly sequential if
175 merged. Differently from CFQ, BFQ achieves this goal with a more
176 reactive mechanism, called Early Queue Merge (EQM). EQM is so
177 responsive in detecting interleaved I/O (cooperating processes),
178 that it enables BFQ to achieve a high throughput, by queue
179 merging, even for queues for which CFQ needs a different
180 mechanism, preemption, to get a high throughput. As such EQM is a
181 unified mechanism to achieve a high throughput with interleaved
184 - Queues are scheduled according to a variant of WF2Q+, named
185 B-WF2Q+, and implemented using an augmented rb-tree to preserve an
186 O(log N) overall complexity. See [2] for more details. B-WF2Q+ is
187 also ready for hierarchical scheduling. However, for a cleaner
188 logical breakdown, the code that enables and completes
189 hierarchical support is provided in the next commit, which focuses
190 exactly on this feature.
192 - B-WF2Q+ guarantees a tight deviation with respect to an ideal,
193 perfectly fair, and smooth service. In particular, B-WF2Q+
194 guarantees that each queue receives a fraction of the device
195 throughput proportional to its weight, even if the throughput
196 fluctuates, and regardless of: the device parameters, the current
197 workload and the budgets assigned to the queue.
199 - The last, budget-independence, property (although probably
200 counterintuitive in the first place) is definitely beneficial, for
201 the following reasons:
203 - First, with any proportional-share scheduler, the maximum
204 deviation with respect to an ideal service is proportional to
205 the maximum budget (slice) assigned to queues. As a consequence,
206 BFQ can keep this deviation tight not only because of the
207 accurate service of B-WF2Q+, but also because BFQ *does not*
208 need to assign a larger budget to a queue to let the queue
209 receive a higher fraction of the device throughput.
211 - Second, BFQ is free to choose, for every process (queue), the
212 budget that best fits the needs of the process, or best
213 leverages the I/O pattern of the process. In particular, BFQ
214 updates queue budgets with a simple feedback-loop algorithm that
215 allows a high throughput to be achieved, while still providing
216 tight latency guarantees to time-sensitive applications. When
217 the in-service queue expires, this algorithm computes the next
218 budget of the queue so as to:
220 - Let large budgets be eventually assigned to the queues
221 associated with I/O-bound applications performing sequential
222 I/O: in fact, the longer these applications are served once
223 got access to the device, the higher the throughput is.
225 - Let small budgets be eventually assigned to the queues
226 associated with time-sensitive applications (which typically
227 perform sporadic and short I/O), because, the smaller the
228 budget assigned to a queue waiting for service is, the sooner
229 B-WF2Q+ will serve that queue (Subsec 3.3 in [2]).
231 - If several processes are competing for the device at the same time,
232 but all processes and groups have the same weight, then BFQ
233 guarantees the expected throughput distribution without ever idling
234 the device. It uses preemption instead. Throughput is then much
235 higher in this common scenario.
237 - ioprio classes are served in strict priority order, i.e.,
238 lower-priority queues are not served as long as there are
239 higher-priority queues. Among queues in the same class, the
240 bandwidth is distributed in proportion to the weight of each
241 queue. A very thin extra bandwidth is however guaranteed to
242 the Idle class, to prevent it from starving.
245 3. What are BFQ's tunable?
246 ==========================
248 The tunables back_seek-max, back_seek_penalty, fifo_expire_async and
249 fifo_expire_sync below are the same as in CFQ. Their description is
250 just copied from that for CFQ. Some considerations in the description
251 of slice_idle are copied from CFQ too.
253 per-process ioprio and weight
254 -----------------------------
256 Unless the cgroups interface is used (see "4. BFQ group scheduling"),
257 weights can be assigned to processes only indirectly, through I/O
258 priorities, and according to the relation:
259 weight = (IOPRIO_BE_NR - ioprio) * 10.
261 Beware that, if low-latency is set, then BFQ automatically raises the
262 weight of the queues associated with interactive and soft real-time
263 applications. Unset this tunable if you need/want to control weights.
268 This parameter specifies how long BFQ should idle for next I/O
269 request, when certain sync BFQ queues become empty. By default
270 slice_idle is a non-zero value. Idling has a double purpose: boosting
271 throughput and making sure that the desired throughput distribution is
272 respected (see the description of how BFQ works, and, if needed, the
273 papers referred there).
275 As for throughput, idling can be very helpful on highly seeky media
276 like single spindle SATA/SAS disks where we can cut down on overall
277 number of seeks and see improved throughput.
279 Setting slice_idle to 0 will remove all the idling on queues and one
280 should see an overall improved throughput on faster storage devices
281 like multiple SATA/SAS disks in hardware RAID configuration.
283 So depending on storage and workload, it might be useful to set
284 slice_idle=0. In general for SATA/SAS disks and software RAID of
285 SATA/SAS disks keeping slice_idle enabled should be useful. For any
286 configurations where there are multiple spindles behind single LUN
287 (Host based hardware RAID controller or for storage arrays), setting
288 slice_idle=0 might end up in better throughput and acceptable
291 Idling is however necessary to have service guarantees enforced in
292 case of differentiated weights or differentiated I/O-request lengths.
293 To see why, suppose that a given BFQ queue A must get several I/O
294 requests served for each request served for another queue B. Idling
295 ensures that, if A makes a new I/O request slightly after becoming
296 empty, then no request of B is dispatched in the middle, and thus A
297 does not lose the possibility to get more than one request dispatched
298 before the next request of B is dispatched. Note that idling
299 guarantees the desired differentiated treatment of queues only in
300 terms of I/O-request dispatches. To guarantee that the actual service
301 order then corresponds to the dispatch order, the strict_guarantees
302 tunable must be set too.
304 There is an important flipside for idling: apart from the above cases
305 where it is beneficial also for throughput, idling can severely impact
306 throughput. One important case is random workload. Because of this
307 issue, BFQ tends to avoid idling as much as possible, when it is not
308 beneficial also for throughput. As a consequence of this behavior, and
309 of further issues described for the strict_guarantees tunable,
310 short-term service guarantees may be occasionally violated. And, in
311 some cases, these guarantees may be more important than guaranteeing
312 maximum throughput. For example, in video playing/streaming, a very
313 low drop rate may be more important than maximum throughput. In these
314 cases, consider setting the strict_guarantees parameter.
319 If this parameter is set (default: unset), then BFQ
321 - always performs idling when the in-service queue becomes empty;
323 - forces the device to serve one I/O request at a time, by dispatching a
324 new request only if there is no outstanding request.
326 In the presence of differentiated weights or I/O-request sizes, both
327 the above conditions are needed to guarantee that every BFQ queue
328 receives its allotted share of the bandwidth. The first condition is
329 needed for the reasons explained in the description of the slice_idle
330 tunable. The second condition is needed because all modern storage
331 devices reorder internally-queued requests, which may trivially break
332 the service guarantees enforced by the I/O scheduler.
334 Setting strict_guarantees may evidently affect throughput.
339 This specifies, given in Kbytes, the maximum "distance" for backward seeking.
340 The distance is the amount of space from the current head location to the
341 sectors that are backward in terms of distance.
343 This parameter allows the scheduler to anticipate requests in the "backward"
344 direction and consider them as being the "next" if they are within this
345 distance from the current head location.
350 This parameter is used to compute the cost of backward seeking. If the
351 backward distance of request is just 1/back_seek_penalty from a "front"
352 request, then the seeking cost of two requests is considered equivalent.
354 So scheduler will not bias toward one or the other request (otherwise scheduler
355 will bias toward front request). Default value of back_seek_penalty is 2.
360 This parameter is used to set the timeout of asynchronous requests. Default
361 value of this is 248ms.
366 This parameter is used to set the timeout of synchronous requests. Default
367 value of this is 124ms. In case to favor synchronous requests over asynchronous
368 one, this value should be decreased relative to fifo_expire_async.
373 This parameter is used to enable/disable BFQ's low latency mode. By
374 default, low latency mode is enabled. If enabled, interactive and soft
375 real-time applications are privileged and experience a lower latency,
376 as explained in more detail in the description of how BFQ works.
378 DO NOT enable this mode if you need full control on bandwidth
379 distribution. In fact, if it is enabled, then BFQ automatically
380 increases the bandwidth share of privileged applications, as the main
381 means to guarantee a lower latency to them.
386 Maximum amount of device time that can be given to a task (queue) once
387 it has been selected for service. On devices with costly seeks,
388 increasing this time usually increases maximum throughput. On the
389 opposite end, increasing this time coarsens the granularity of the
390 short-term bandwidth and latency guarantees, especially if the
391 following parameter is set to zero.
396 Maximum amount of service, measured in sectors, that can be provided
397 to a BFQ queue once it is set in service (of course within the limits
398 of the above timeout). According to what said in the description of
399 the algorithm, larger values increase the throughput in proportion to
400 the percentage of sequential I/O requests issued. The price of larger
401 values is that they coarsen the granularity of short-term bandwidth
402 and latency guarantees.
404 The default value is 0, which enables auto-tuning: BFQ sets max_budget
405 to the maximum number of sectors that can be served during
406 timeout_sync, according to the estimated peak rate.
411 Read-only parameter, used to show the weights of the currently active
418 BFQ exports a few parameters to control/tune the behavior of
419 low-latency heuristics.
423 Factor by which the weight of a weight-raised queue is multiplied. If
424 the queue is deemed soft real-time, then the weight is further
425 multiplied by an additional, constant factor.
429 Maximum duration of a weight-raising period for an interactive task
430 (ms). If set to zero (default value), then this value is computed
431 automatically, as a function of the peak rate of the device. In any
432 case, when the value of this parameter is read, it always reports the
433 current duration, regardless of whether it has been set manually or
434 computed automatically.
438 Maximum service rate below which a queue is deemed to be associated
439 with a soft real-time application, and is then weight-raised
440 accordingly (sectors/sec).
444 Minimum idle period after which interactive weight-raising may be
445 reactivated for a queue (in ms).
449 Maximum weight-raising duration for soft real-time queues (in ms). The
450 start time from which this duration is considered is automatically
451 moved forward if the queue is detected to be still soft real-time
452 before the current soft real-time weight-raising period finishes.
454 wr_min_inter_arr_async
456 Minimum period between I/O request arrivals after which weight-raising
457 may be reactivated for an already busy async queue (in ms).
460 4. Group scheduling with BFQ
461 ============================
463 BFQ supports both cgroups-v1 and cgroups-v2 io controllers, namely
464 blkio and io. In particular, BFQ supports weight-based proportional
465 share. To activate cgroups support, set BFQ_GROUP_IOSCHED.
467 4-1 Service guarantees provided
468 -------------------------------
470 With BFQ, proportional share means true proportional share of the
471 device bandwidth, according to group weights. For example, a group
472 with weight 200 gets twice the bandwidth, and not just twice the time,
473 of a group with weight 100.
475 BFQ supports hierarchies (group trees) of any depth. Bandwidth is
476 distributed among groups and processes in the expected way: for each
477 group, the children of the group share the whole bandwidth of the
478 group in proportion to their weights. In particular, this implies
479 that, for each leaf group, every process of the group receives the
480 same share of the whole group bandwidth, unless the ioprio of the
483 The resource-sharing guarantee for a group may partially or totally
484 switch from bandwidth to time, if providing bandwidth guarantees to
485 the group lowers the throughput too much. This switch occurs on a
486 per-process basis: if a process of a leaf group causes throughput loss
487 if served in such a way to receive its share of the bandwidth, then
488 BFQ switches back to just time-based proportional share for that
494 To get proportional sharing of bandwidth with BFQ for a given device,
495 BFQ must of course be the active scheduler for that device.
497 Within each group directory, the names of the files associated with
498 BFQ-specific cgroup parameters and stats begin with the "bfq."
499 prefix. So, with cgroups-v1 or cgroups-v2, the full prefix for
500 BFQ-specific files is "blkio.bfq." or "io.bfq." For example, the group
501 parameter to set the weight of a group with BFQ is blkio.bfq.weight
507 For each group, there is only the following parameter to set.
509 weight (namely blkio.bfq.weight or io.bfq-weight): the weight of the
510 group inside its parent. Available values: 1..10000 (default 100). The
511 linear mapping between ioprio and weights, described at the beginning
512 of the tunable section, is still valid, but all weights higher than
513 IOPRIO_BE_NR*10 are mapped to ioprio 0.
515 Recall that, if low-latency is set, then BFQ automatically raises the
516 weight of the queues associated with interactive and soft real-time
517 applications. Unset this tunable if you need/want to control weights.
520 [1] P. Valente, A. Avanzini, "Evolution of the BFQ Storage I/O
521 Scheduler", Proceedings of the First Workshop on Mobile System
522 Technologies (MST-2015), May 2015.
523 http://algogroup.unimore.it/people/paolo/disk_sched/mst-2015.pdf
525 [2] P. Valente and M. Andreolini, "Improving Application
526 Responsiveness with the BFQ Disk I/O Scheduler", Proceedings of
527 the 5th Annual International Systems and Storage Conference
528 (SYSTOR '12), June 2012.
529 Slightly extended version:
530 http://algogroup.unimore.it/people/paolo/disk_sched/bfq-v1-suite-