summaryrefslogtreecommitdiffstats
path: root/block
diff options
context:
space:
mode:
authorPaolo Valente <paolo.valente@linaro.org>2019-01-29 12:06:32 +0100
committerJens Axboe <axboe@kernel.dk>2019-01-31 12:50:23 -0700
commit530c4cbb3c62f9e42dbf39279fb346f2d2ab4dbb (patch)
treeaad632d87cc3978218cbad308e0b0bd250025e34 /block
parentac8b0cb415f3aa9162009d39624501d37031533b (diff)
downloadlinux-530c4cbb3c62f9e42dbf39279fb346f2d2ab4dbb.tar.bz2
block, bfq: unconditionally plug I/O in asymmetric scenarios
bfq detects the creation of multiple bfq_queues shortly after each other, namely a burst of queue creations in the terminology used in the code. If the burst is large, then no queue in the burst is granted - either I/O-dispatch plugging when the queue remains temporarily idle while in service; - or weight raising, because it causes even longer plugging. In fact, such a plugging tends to lower throughput, while these bursts are typically due to applications or services that spawn multiple processes, to reach a common goal as soon as possible. Examples are a "git grep" or the booting of a system. Unfortunately, disabling plugging may cause a loss of service guarantees in asymmetric scenarios, i.e., if queue weights are differentiated or if more than one group is active. This commit addresses this issue by no longer disabling I/O-dispatch plugging for queues in large bursts. Signed-off-by: Paolo Valente <paolo.valente@linaro.org> Signed-off-by: Jens Axboe <axboe@kernel.dk>
Diffstat (limited to 'block')
-rw-r--r--block/bfq-iosched.c346
1 files changed, 165 insertions, 181 deletions
diff --git a/block/bfq-iosched.c b/block/bfq-iosched.c
index a6fe60114ade..c1bb5e5fcdc4 100644
--- a/block/bfq-iosched.c
+++ b/block/bfq-iosched.c
@@ -3479,191 +3479,175 @@ static bool idling_boosts_thr_without_issues(struct bfq_data *bfqd,
bfqd->wr_busy_queues == 0;
}
+/*
+ * There is a case where idling must be performed not for
+ * throughput concerns, but to preserve service guarantees.
+ *
+ * To introduce this case, we can note that allowing the drive
+ * to enqueue more than one request at a time, and hence
+ * delegating de facto final scheduling decisions to the
+ * drive's internal scheduler, entails loss of control on the
+ * actual request service order. In particular, the critical
+ * situation is when requests from different processes happen
+ * to be present, at the same time, in the internal queue(s)
+ * of the drive. In such a situation, the drive, by deciding
+ * the service order of the internally-queued requests, does
+ * determine also the actual throughput distribution among
+ * these processes. But the drive typically has no notion or
+ * concern about per-process throughput distribution, and
+ * makes its decisions only on a per-request basis. Therefore,
+ * the service distribution enforced by the drive's internal
+ * scheduler is likely to coincide with the desired
+ * device-throughput distribution only in a completely
+ * symmetric scenario where:
+ * (i) each of these processes must get the same throughput as
+ * the others;
+ * (ii) the I/O of each process has the same properties, in
+ * terms of locality (sequential or random), direction
+ * (reads or writes), request sizes, greediness
+ * (from I/O-bound to sporadic), and so on.
+ * In fact, in such a scenario, the drive tends to treat
+ * the requests of each of these processes in about the same
+ * way as the requests of the others, and thus to provide
+ * each of these processes with about the same throughput
+ * (which is exactly the desired throughput distribution). In
+ * contrast, in any asymmetric scenario, device idling is
+ * certainly needed to guarantee that bfqq receives its
+ * assigned fraction of the device throughput (see [1] for
+ * details).
+ * The problem is that idling may significantly reduce
+ * throughput with certain combinations of types of I/O and
+ * devices. An important example is sync random I/O, on flash
+ * storage with command queueing. So, unless bfqq falls in the
+ * above cases where idling also boosts throughput, it would
+ * be important to check conditions (i) and (ii) accurately,
+ * so as to avoid idling when not strictly needed for service
+ * guarantees.
+ *
+ * Unfortunately, it is extremely difficult to thoroughly
+ * check condition (ii). And, in case there are active groups,
+ * it becomes very difficult to check condition (i) too. In
+ * fact, if there are active groups, then, for condition (i)
+ * to become false, it is enough that an active group contains
+ * more active processes or sub-groups than some other active
+ * group. More precisely, for condition (i) to hold because of
+ * such a group, it is not even necessary that the group is
+ * (still) active: it is sufficient that, even if the group
+ * has become inactive, some of its descendant processes still
+ * have some request already dispatched but still waiting for
+ * completion. In fact, requests have still to be guaranteed
+ * their share of the throughput even after being
+ * dispatched. In this respect, it is easy to show that, if a
+ * group frequently becomes inactive while still having
+ * in-flight requests, and if, when this happens, the group is
+ * not considered in the calculation of whether the scenario
+ * is asymmetric, then the group may fail to be guaranteed its
+ * fair share of the throughput (basically because idling may
+ * not be performed for the descendant processes of the group,
+ * but it had to be). We address this issue with the
+ * following bi-modal behavior, implemented in the function
+ * bfq_symmetric_scenario().
+ *
+ * If there are groups with requests waiting for completion
+ * (as commented above, some of these groups may even be
+ * already inactive), then the scenario is tagged as
+ * asymmetric, conservatively, without checking any of the
+ * conditions (i) and (ii). So the device is idled for bfqq.
+ * This behavior matches also the fact that groups are created
+ * exactly if controlling I/O is a primary concern (to
+ * preserve bandwidth and latency guarantees).
+ *
+ * On the opposite end, if there are no groups with requests
+ * waiting for completion, then only condition (i) is actually
+ * controlled, i.e., provided that condition (i) holds, idling
+ * is not performed, regardless of whether condition (ii)
+ * holds. In other words, only if condition (i) does not hold,
+ * then idling is allowed, and the device tends to be
+ * prevented from queueing many requests, possibly of several
+ * processes. Since there are no groups with requests waiting
+ * for completion, then, to control condition (i) it is enough
+ * to check just whether all the queues with requests waiting
+ * for completion also have the same weight.
+ *
+ * Not checking condition (ii) evidently exposes bfqq to the
+ * risk of getting less throughput than its fair share.
+ * However, for queues with the same weight, a further
+ * mechanism, preemption, mitigates or even eliminates this
+ * problem. And it does so without consequences on overall
+ * throughput. This mechanism and its benefits are explained
+ * in the next three paragraphs.
+ *
+ * Even if a queue, say Q, is expired when it remains idle, Q
+ * can still preempt the new in-service queue if the next
+ * request of Q arrives soon (see the comments on
+ * bfq_bfqq_update_budg_for_activation). If all queues and
+ * groups have the same weight, this form of preemption,
+ * combined with the hole-recovery heuristic described in the
+ * comments on function bfq_bfqq_update_budg_for_activation,
+ * are enough to preserve a correct bandwidth distribution in
+ * the mid term, even without idling. In fact, even if not
+ * idling allows the internal queues of the device to contain
+ * many requests, and thus to reorder requests, we can rather
+ * safely assume that the internal scheduler still preserves a
+ * minimum of mid-term fairness.
+ *
+ * More precisely, this preemption-based, idleless approach
+ * provides fairness in terms of IOPS, and not sectors per
+ * second. This can be seen with a simple example. Suppose
+ * that there are two queues with the same weight, but that
+ * the first queue receives requests of 8 sectors, while the
+ * second queue receives requests of 1024 sectors. In
+ * addition, suppose that each of the two queues contains at
+ * most one request at a time, which implies that each queue
+ * always remains idle after it is served. Finally, after
+ * remaining idle, each queue receives very quickly a new
+ * request. It follows that the two queues are served
+ * alternatively, preempting each other if needed. This
+ * implies that, although both queues have the same weight,
+ * the queue with large requests receives a service that is
+ * 1024/8 times as high as the service received by the other
+ * queue.
+ *
+ * The motivation for using preemption instead of idling (for
+ * queues with the same weight) is that, by not idling,
+ * service guarantees are preserved (completely or at least in
+ * part) without minimally sacrificing throughput. And, if
+ * there is no active group, then the primary expectation for
+ * this device is probably a high throughput.
+ *
+ * We are now left only with explaining the additional
+ * compound condition that is checked below for deciding
+ * whether the scenario is asymmetric. To explain this
+ * compound condition, we need to add that the function
+ * bfq_symmetric_scenario checks the weights of only
+ * non-weight-raised queues, for efficiency reasons (see
+ * comments on bfq_weights_tree_add()). Then the fact that
+ * bfqq is weight-raised is checked explicitly here. More
+ * precisely, the compound condition below takes into account
+ * also the fact that, even if bfqq is being weight-raised,
+ * the scenario is still symmetric if all queues with requests
+ * waiting for completion happen to be
+ * weight-raised. Actually, we should be even more precise
+ * here, and differentiate between interactive weight raising
+ * and soft real-time weight raising.
+ *
+ * As a side note, it is worth considering that the above
+ * device-idling countermeasures may however fail in the
+ * following unlucky scenario: if idling is (correctly)
+ * disabled in a time period during which all symmetry
+ * sub-conditions hold, and hence the device is allowed to
+ * enqueue many requests, but at some later point in time some
+ * sub-condition stops to hold, then it may become impossible
+ * to let requests be served in the desired order until all
+ * the requests already queued in the device have been served.
+ */
static bool idling_needed_for_service_guarantees(struct bfq_data *bfqd,
struct bfq_queue *bfqq)
{
- /*
- * There is a case where idling must be performed not for
- * throughput concerns, but to preserve service guarantees.
- *
- * To introduce this case, we can note that allowing the drive
- * to enqueue more than one request at a time, and thereby
- * delegating de facto final scheduling decisions to the
- * drive's internal scheduler, entails loss of control on the
- * actual request service order. In particular, the critical
- * situation is when requests from different processes happen
- * to be present, at the same time, in the internal queue(s)
- * of the drive. In such a situation, the drive, by deciding
- * the service order of the internally-queued requests, does
- * determine also the actual throughput distribution among
- * these processes. But the drive typically has no notion or
- * concern about per-process throughput distribution, and
- * makes its decisions only on a per-request basis. Therefore,
- * the service distribution enforced by the drive's internal
- * scheduler is likely to coincide with the desired
- * device-throughput distribution only in a completely
- * symmetric scenario where:
- * (i) each of these processes must get the same throughput as
- * the others;
- * (ii) the I/O of each process has the same properties, in
- * terms of locality (sequential or random), direction
- * (reads or writes), request sizes, greediness
- * (from I/O-bound to sporadic), and so on.
- * In fact, in such a scenario, the drive tends to treat
- * the requests of each of these processes in about the same
- * way as the requests of the others, and thus to provide
- * each of these processes with about the same throughput
- * (which is exactly the desired throughput distribution). In
- * contrast, in any asymmetric scenario, device idling is
- * certainly needed to guarantee that bfqq receives its
- * assigned fraction of the device throughput (see [1] for
- * details).
- * The problem is that idling may significantly reduce
- * throughput with certain combinations of types of I/O and
- * devices. An important example is sync random I/O, on flash
- * storage with command queueing. So, unless bfqq falls in the
- * above cases where idling also boosts throughput, it would
- * be important to check conditions (i) and (ii) accurately,
- * so as to avoid idling when not strictly needed for service
- * guarantees.
- *
- * Unfortunately, it is extremely difficult to thoroughly
- * check condition (ii). And, in case there are active groups,
- * it becomes very difficult to check condition (i) too. In
- * fact, if there are active groups, then, for condition (i)
- * to become false, it is enough that an active group contains
- * more active processes or sub-groups than some other active
- * group. More precisely, for condition (i) to hold because of
- * such a group, it is not even necessary that the group is
- * (still) active: it is sufficient that, even if the group
- * has become inactive, some of its descendant processes still
- * have some request already dispatched but still waiting for
- * completion. In fact, requests have still to be guaranteed
- * their share of the throughput even after being
- * dispatched. In this respect, it is easy to show that, if a
- * group frequently becomes inactive while still having
- * in-flight requests, and if, when this happens, the group is
- * not considered in the calculation of whether the scenario
- * is asymmetric, then the group may fail to be guaranteed its
- * fair share of the throughput (basically because idling may
- * not be performed for the descendant processes of the group,
- * but it had to be). We address this issue with the
- * following bi-modal behavior, implemented in the function
- * bfq_symmetric_scenario().
- *
- * If there are groups with requests waiting for completion
- * (as commented above, some of these groups may even be
- * already inactive), then the scenario is tagged as
- * asymmetric, conservatively, without checking any of the
- * conditions (i) and (ii). So the device is idled for bfqq.
- * This behavior matches also the fact that groups are created
- * exactly if controlling I/O is a primary concern (to
- * preserve bandwidth and latency guarantees).
- *
- * On the opposite end, if there are no groups with requests
- * waiting for completion, then only condition (i) is actually
- * controlled, i.e., provided that condition (i) holds, idling
- * is not performed, regardless of whether condition (ii)
- * holds. In other words, only if condition (i) does not hold,
- * then idling is allowed, and the device tends to be
- * prevented from queueing many requests, possibly of several
- * processes. Since there are no groups with requests waiting
- * for completion, then, to control condition (i) it is enough
- * to check just whether all the queues with requests waiting
- * for completion also have the same weight.
- *
- * Not checking condition (ii) evidently exposes bfqq to the
- * risk of getting less throughput than its fair share.
- * However, for queues with the same weight, a further
- * mechanism, preemption, mitigates or even eliminates this
- * problem. And it does so without consequences on overall
- * throughput. This mechanism and its benefits are explained
- * in the next three paragraphs.
- *
- * Even if a queue, say Q, is expired when it remains idle, Q
- * can still preempt the new in-service queue if the next
- * request of Q arrives soon (see the comments on
- * bfq_bfqq_update_budg_for_activation). If all queues and
- * groups have the same weight, this form of preemption,
- * combined with the hole-recovery heuristic described in the
- * comments on function bfq_bfqq_update_budg_for_activation,
- * are enough to preserve a correct bandwidth distribution in
- * the mid term, even without idling. In fact, even if not
- * idling allows the internal queues of the device to contain
- * many requests, and thus to reorder requests, we can rather
- * safely assume that the internal scheduler still preserves a
- * minimum of mid-term fairness.
- *
- * More precisely, this preemption-based, idleless approach
- * provides fairness in terms of IOPS, and not sectors per
- * second. This can be seen with a simple example. Suppose
- * that there are two queues with the same weight, but that
- * the first queue receives requests of 8 sectors, while the
- * second queue receives requests of 1024 sectors. In
- * addition, suppose that each of the two queues contains at
- * most one request at a time, which implies that each queue
- * always remains idle after it is served. Finally, after
- * remaining idle, each queue receives very quickly a new
- * request. It follows that the two queues are served
- * alternatively, preempting each other if needed. This
- * implies that, although both queues have the same weight,
- * the queue with large requests receives a service that is
- * 1024/8 times as high as the service received by the other
- * queue.
- *
- * The motivation for using preemption instead of idling (for
- * queues with the same weight) is that, by not idling,
- * service guarantees are preserved (completely or at least in
- * part) without minimally sacrificing throughput. And, if
- * there is no active group, then the primary expectation for
- * this device is probably a high throughput.
- *
- * We are now left only with explaining the additional
- * compound condition that is checked below for deciding
- * whether the scenario is asymmetric. To explain this
- * compound condition, we need to add that the function
- * bfq_symmetric_scenario checks the weights of only
- * non-weight-raised queues, for efficiency reasons (see
- * comments on bfq_weights_tree_add()). Then the fact that
- * bfqq is weight-raised is checked explicitly here. More
- * precisely, the compound condition below takes into account
- * also the fact that, even if bfqq is being weight-raised,
- * the scenario is still symmetric if all queues with requests
- * waiting for completion happen to be
- * weight-raised. Actually, we should be even more precise
- * here, and differentiate between interactive weight raising
- * and soft real-time weight raising.
- *
- * As a side note, it is worth considering that the above
- * device-idling countermeasures may however fail in the
- * following unlucky scenario: if idling is (correctly)
- * disabled in a time period during which all symmetry
- * sub-conditions hold, and hence the device is allowed to
- * enqueue many requests, but at some later point in time some
- * sub-condition stops to hold, then it may become impossible
- * to let requests be served in the desired order until all
- * the requests already queued in the device have been served.
- */
- bool asymmetric_scenario = (bfqq->wr_coeff > 1 &&
- bfqd->wr_busy_queues <
- bfq_tot_busy_queues(bfqd)) ||
+ return (bfqq->wr_coeff > 1 &&
+ bfqd->wr_busy_queues <
+ bfq_tot_busy_queues(bfqd)) ||
!bfq_symmetric_scenario(bfqd);
-
- /*
- * Finally, there is a case where maximizing throughput is the
- * best choice even if it may cause unfairness toward
- * bfqq. Such a case is when bfqq became active in a burst of
- * queue activations. Queues that became active during a large
- * burst benefit only from throughput, as discussed in the
- * comments on bfq_handle_burst. Thus, if bfqq became active
- * in a burst and not idling the device maximizes throughput,
- * then the device must no be idled, because not idling the
- * device provides bfqq and all other queues in the burst with
- * maximum benefit. Combining this and the above case, we can
- * now establish when idling is actually needed to preserve
- * service guarantees.
- */
- return asymmetric_scenario && !bfq_bfqq_in_large_burst(bfqq);
}
/*