block, bfq: do not merge queues on flash storage with queueing

To boost throughput with a set of processes doing interleaved I/O
(i.e., a set of processes whose individual I/O is random, but whose
merged cumulative I/O is sequential), BFQ merges the queues associated
with these processes, i.e., redirects the I/O of these processes into a
common, shared queue. In the shared queue, I/O requests are ordered by
their position on the medium, thus sequential I/O gets dispatched to
the device when the shared queue is served.

Queue merging costs execution time, because, to detect which queues to
merge, BFQ must maintain a list of the head I/O requests of active
queues, ordered by request positions. Measurements showed that this
costs about 10% of BFQ's total per-request processing time.

Request processing time becomes more and more critical as the speed of
the underlying storage device grows. Yet, fortunately, queue merging
is basically useless on the very devices that are so fast to make
request processing time critical. To reach a high throughput, these
devices must have many requests queued at the same time. But, in this
configuration, the internal scheduling algorithms of these devices do
also the job of queue merging: they reorder requests so as to obtain
as much as possible a sequential I/O pattern. As a consequence, with
processes doing interleaved I/O, the throughput reached by one such
device is likely to be the same, with and without queue merging.

In view of this fact, this commit disables queue merging, and all
related housekeeping, for non-rotational devices with internal
queueing. The total, single-lock-protected, per-request processing
time of BFQ drops to, e.g., 1.9 us on an Intel Core i7-2760QM@2.40GHz
(time measured with simple code instrumentation, and using the
throughput-sync.sh script of the S suite [1], in performance-profiling
mode). To put this result into context, the total,
single-lock-protected, per-request execution time of the lightest I/O
scheduler available in blk-mq, mq-deadline, is 0.7 us (mq-deadline is
~800 LOC, against ~10500 LOC for BFQ).

Disabling merging provides a further, remarkable benefit in terms of
throughput. Merging tends to make many workloads artificially more
uneven, mainly because of shared queues remaining non empty for
incomparably more time than normal queues. So, if, e.g., one of the
queues in a set of merged queues has a higher weight than a normal
queue, then the shared queue may inherit such a high weight and, by
staying almost always active, may force BFQ to perform I/O plugging
most of the time. This evidently makes it harder for BFQ to let the
device reach a high throughput.

As a practical example of this problem, and of the benefits of this
commit, we measured again the throughput in the nasty scenario
considered in previous commit messages: dbench test (in the Phoronix
suite), with 6 clients, on a filesystem with journaling, and with the
journaling daemon enjoying a higher weight than normal processes. With
this commit, the throughput grows from ~150 MB/s to ~200 MB/s on a
PLEXTOR PX-256M5 SSD. This is the same peak throughput reached by any
of the other I/O schedulers. As such, this is also likely to be the
maximum possible throughput reachable with this workload on this
device, because I/O is mostly random, and the other schedulers
basically just pass I/O requests to the drive as fast as possible.

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



Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Tested-by: Francesco Pollicino <fra.fra.800@gmail.com>
Signed-off-by: Alessio Masola <alessio.masola@gmail.com>
Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Jens Axboe <axboe@kernel.dk>