On 08/04/2010 01:58 PM, Paolo Valente wrote: > Hi, > I have been working for a few years (with Fabio Checconi) on a disk > scheduler providing definitely lower latencies than cfq, as well as a > higher throughput with most of the test workloads we used (or the same > throughput as cfq with the other workloads). We named this scheduler > bfq (budget fair queueing). I hope this is the right list for announcing > this work. > > One of the things we measured in our tests is the cold-cache execution > time of a command as, e.g., "bash -c exit", "xterm /bin/true" or > "konsole -e /bin/true", while the disk was also accessed by different > combinations of sequential, or random, readers and/or > writers. Depending on which of these background workloads was used, > these execution times were five to nine times lower with bfq under > 2.6.32. Under 2.6.35 they were instead from six to fourteen times > lower. The highest price paid for these lower latencies was a 20% loss > of aggregated disk throughput for konsole in case of background > workloads made only of sequential requests (due to the fact that bfq > of course privileges, more than cfq, the seeky IO needed to load > konsole and its dependencies). In contrast, with shorter commands, as > bash or xterm, bfq also provided up to 30% higher aggregated > throughput. > > We saw from 15% to 30% higher aggregated throughput also in our > only-aggregated-throughput tests. You can find in [1] all the details > on our tests and on other nice features of bfq, such as the fact that > it perfectly distributes the disk throughput as desired, independently > of disk physical parameters like, e.g., ZBR. in [1] you can also find > a detailed description of bfq and a short report on the maturity level > of the code (TODO list), plus all the scripts used for the tests. > > The results I mentioned so far have been achieved with the last > version of bfq, released about two months ago as patchsets for 2.6.33 > or 2.6.34. From a few days a patchset for 2.6.35 is available too, as > well as a backport to 2.6.32. The latter has been prepared by Mauro > Andreolini, who also helped me a lot with debugging. All these patches > can be found here [2]. Mauro also built a binary kernel package for > current lucid, and hosted it into a PPA, which can be found here [3]. > > A few days after being released, this version of bfq has been > introduced as the default disk scheduler in the Zen Kernel. It has > been adopted as the default disk scheduler in Gentoo Linux too. I > also recorded downloads from users with other distributions, as, e.g., > Ubuntu and ArchLinux. As of now we received only positive feedbacks > from the users. > > Paolo > > [1] http://algo.ing.unimo.it/people/paolo/disk_sched/ > [2] http://algo.ing.unimo.it/people/paolo/disk_sched/sources.php > [3] Ubuntu PPA: ppa:mauro-andreolini/ubuntu-kernel-bfq > Hi Paolo, Have you tried to post this to the upstream developers of CFQ and IO schedulers? Regards, Ric -- devel mailing list devel@xxxxxxxxxxxxxxxxxxxxxxx https://admin.fedoraproject.org/mailman/listinfo/devel
