[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <20161019183845.GL18532@htj.duckdns.org>
Date: Wed, 19 Oct 2016 14:38:45 -0400
From: Tejun Heo <tj@...nel.org>
To: Adam Manzanares <adam.manzanares@...t.com>
Cc: axboe@...nel.dk, dan.j.williams@...el.com, hare@...e.de,
martin.petersen@...cle.com, mchristi@...hat.com,
toshi.kani@....com, ming.lei@...onical.com,
sathya.prakash@...adcom.com, chaitra.basappa@...adcom.com,
suganath-prabu.subramani@...adcom.com, linux-block@...r.kernel.org,
linux-ide@...r.kernel.org, linux-kernel@...r.kernel.org,
MPT-FusionLinux.pdl@...adcom.com, linux-scsi@...r.kernel.org
Subject: Re: [PATCH v6 0/3] Enabling ATA Command Priorities
On Mon, Oct 17, 2016 at 11:27:27AM -0700, Adam Manzanares wrote:
> This patch builds ATA commands with high priority if the iocontext of a process
> is set to real time. The goal of the patch is to improve tail latencies of
> workloads that use higher queue depths. This requires setting the iocontext
> ioprio on the request when it is initialized
>
> This patch has been tested with an Ultrastar HE8 HDD and cuts the
> the p99.99 tail latency of foreground IO from 2s down to 72ms when
> using the deadline scheduler. This patch works independently of the
> scheduler so it can be used with all of the currently available
> request based schedulers.
>
> Foreground IO, for the previously described results, is an async fio job
> submitting 4K read requests at a QD of 1 to the HDD. The foreground IO is set
> with the iopriority class of real time. The background workload is another fio
> job submitting read requests at a QD of 32 to the same HDD with default
> iopriority.
>
> This feature is enabled for ATA devices by setting the ata ncq_prio_on device
> attribute to 1. An ATA device is also checked to see if the device supports per
> command priority.
Applied 1-3 to libata/for-4.10 w/ some modifications.
Thanks.
--
tejun
Powered by blists - more mailing lists