Linaro Connect Bangkok

DescriptionThis presentation is a report on the last improvements on the BFQ I/O scheduler. These improvements benefit virtually any system, from embedded devices, to personal systems, to nodes in a data center.

This first set of changes concerns throughput. In the most complex scenarios for guaranteeing I/O bandwidths, BFQ delivers up to five-time higher throughput than existing solutions. But the same mechanisms that gained BFQ this primacy become a hindrance with some 'deceptive' workloads. These workloads trick BFQ mechanisms into wrongly believing that some I/O flows need to be privileged with respect to other flows, even at the expense of losing throughput dramatically. In contrast, total throughout is the only performance parameter that matters. We took countermeasures to offset this loss of throughput, countermeasures that fully succeed with some 
unfriendly workloads.

Then, as for I/O control, the combination of several new improvements and fixes let the worst-case start-up time of applications drop by an additional 35%. We show these results not only through graphs, but also through a new demo with a Chromebook.

The last contributions shown in this presentation are about
efficiency. In fact, even the execution overhead of an I/O scheduler may limit maximum throughput with very fast drives. So, to reduce BFQ overhead, we tried to turn some properties of these drives into BFQ's advantage: we looked for costly optimizations that are no longer necessary with these drives. We found some, and added controls that automatically turn them off when not needed.