Budget Fair Queueing I/O Scheduler ================================== This patchset transforms BFQ-v1-r1 for Linux 2.6.35 into BFQ-v1-r1 for Linux 2.6.35. For further information: http://algo.ing.unimo.it/people/paolo/disk_sched/. CHANGELOG IMPORTANT NOTE: This version of BFQ contains a bug that may cause a panic when a process is moved to a new cgroup. This functionality seems to be rarely used, as we have not yet received any report of this panic. We will be glad to fix the bug after the first report. This bug has however been removed from version v3r3 onwards. 2.6.34-v1-r1: Bugfix: replaced calls to spin_is_locked, which does not work properly CONFIG_SMP is not set, with assert_spin_locked v1: This is a new version of BFQ with respect to the versions you can find on Fabio's site: http://feanor.sssup.it/~fabio/linux/bfq. Here is what we changed with respect to the previous versions: 1) re-tuned the budget feedback mechanism: it is now slighlty more biased toward assigning high budgets, to boost the aggregated throughput more, and more quickly as new processes are started 2) introduced more tolerance toward seeky queues (I verified that the phenomenona described below used to occurr systematically): 2a: if a queue is expired after having received very little service, then it is not punished as a seeky queue, even if it occurred to consume that little service too slowly; the rationale is that, if the new active queue has been served for a too short time interval, then its possible sequential accesses may not yet prevail on the initial latencies for moving the disk head on the first sector requested 2b: the waiting time (disk idling) of a queue detected as seeky as a function of the position of the requests it issued is reduced to a very low value only after the queue has consumed a minimum fraction of the assigned budget; this prevents processes generating (partly) seeky workloads from being too ill-treated 2c: if a queue has consumed 'enough' budget upon a budget timeout, then, even if it did not consume all of its budget, that queue is not punished as any seeky queue; the rationale is that, depending on the disk zones, a queue may be served at a lower rate than the estimated peak rate. Changes 2a and 2b have been critical in lowering latencies, whereas change 2c, in addition to change 1, helped a lot increase the disk throughput. 3) slightly changed the peak rate estimator: a low-pass filter is now used instead of just keeping the highest rate sampled; the rationale is that the peak rate of a disk should be quite stable, so the filter should converge more or less smoothly to the right value; it seemed to correctly catch the peak rate with all disks we used 4) added the low latency mechanism described in detail in http://algo.ing.unimo.it/people/paolo/disk_sched/description.php.