On Mon, Dec 11 2017 at 11:00am -0500, Scott Bauer <scott.bauer@xxxxxxxxx> wrote: > Signed-off-by: Scott Bauer <scott.bauer@xxxxxxxxx> > --- > Documentation/device-mapper/dm-unstripe.txt | 82 +++++++++++++++++++++++++++++ > 1 file changed, 82 insertions(+) > create mode 100644 Documentation/device-mapper/dm-unstripe.txt > > diff --git a/Documentation/device-mapper/dm-unstripe.txt b/Documentation/device-mapper/dm-unstripe.txt > new file mode 100644 > index 000000000000..4e1a0a39a689 > --- /dev/null > +++ b/Documentation/device-mapper/dm-unstripe.txt > @@ -0,0 +1,82 @@ > +Device-Mapper Unstripe > +===================== > + > +The device-mapper Unstripe (dm-unstripe) target provides a transparent > +mechanism to unstripe a RAID 0 striping to access segregated disks. > + > +This module should be used by users who understand what the underlying > +disks look like behind the software/hardware RAID. > + > +Parameters: > +<drive (ex: /dev/nvme0n1)> <drive #> <# of drives> <stripe sectors> > + > + > +<drive> > + The block device you wish to unstripe. > + > +<drive #> > + The physical drive you wish to expose via this "virtual" device > + mapper target. This must be 0 indexed. > + > +<# of drives> > + The number of drives in the RAID 0. > + > +<stripe sectors> > + The amount of 512B sectors in the raid striping, or zero, if you > + wish you use max_hw_sector_size. > + > + > +Why use this module? > +===================== > + > +As a use case: > + > + > + As an example: > + > + Intel NVMe drives contain two cores on the physical device. > + Each core of the drive has segregated access to its LBA range. > + The current LBA model has a RAID 0 128k stripe across the two cores: > + > + Core 0: Core 1: > + __________ __________ > + | LBA 511| | LBA 768| > + | LBA 0 | | LBA 256| > + ⎻⎻⎻⎻⎻⎻⎻⎻⎻⎻ ⎻⎻⎻⎻⎻⎻⎻⎻⎻⎻ > + > + The purpose of this unstriping is to provide better QoS in noisy > + neighbor environments. When two partitions are created on the > + aggregate drive without this unstriping, reads on one partition > + can affect writes on another partition. With the striping concurrent > + reads and writes and I/O on opposite cores have lower completion times, > + and better tail latencies. > + > + With the module we were able to segregate a fio script that has read and > + write jobs that are independent of each other. Compared to when we run > + the test on a combined drive with partitions, we were able to get a 92% > + reduction in five-9ths read latency using this device mapper target. > + > + > + One could use the module to Logical de-pop a HDD if you have sufficient > + geometry information regarding the drive. OK, but I'm left wondering: why doesn't the user avoid striping across the cores? Do the Intel NVMe drives not provide the ability to present 1 device per NVMe core? This DM target seems like a pretty nasty workaround for what should be fixed in the NVMe drive's firmware. Mainly because there is no opportunity to use both striped and unstriped access to the same NVMe drive. So why impose striped on the user in the first place? Mike -- dm-devel mailing list dm-devel@xxxxxxxxxx https://www.redhat.com/mailman/listinfo/dm-devel
