Re: [LSF/MM/BPF TOPIC] NVMe HDD

[Ming Lei <ming.lei@redhat.com>]

On Thu, Feb 13, 2020 at 08:24:36AM +0000, Damien Le Moal wrote:
> On 2020/02/13 16:54, Ming Lei wrote:
> > On Thu, Feb 13, 2020 at 02:40:41AM +0000, Damien Le Moal wrote:
> >> Ming,
> >>
> >> On 2020/02/13 7:03, Ming Lei wrote:
> >>> On Wed, Feb 12, 2020 at 01:47:53AM +0000, Damien Le Moal wrote:
> >>>> On 2020/02/12 4:01, Tim Walker wrote:
> >>>>> On Tue, Feb 11, 2020 at 7:28 AM Ming Lei <ming.lei@redhat.com> wrote:
> >>>>>>
> >>>>>> On Mon, Feb 10, 2020 at 02:20:10PM -0500, Tim Walker wrote:
> >>>>>>> Background:
> >>>>>>>
> >>>>>>> NVMe specification has hardened over the decade and now NVMe devices
> >>>>>>> are well integrated into our customers’ systems. As we look forward,
> >>>>>>> moving HDDs to the NVMe command set eliminates the SAS IOC and driver
> >>>>>>> stack, consolidating on a single access method for rotational and
> >>>>>>> static storage technologies. PCIe-NVMe offers near-SATA interface
> >>>>>>> costs, features and performance suitable for high-cap HDDs, and
> >>>>>>> optimal interoperability for storage automation, tiering, and
> >>>>>>> management. We will share some early conceptual results and proposed
> >>>>>>> salient design goals and challenges surrounding an NVMe HDD.
> >>>>>>
> >>>>>> HDD. performance is very sensitive to IO order. Could you provide some
> >>>>>> background info about NVMe HDD? Such as:
> >>>>>>
> >>>>>> - number of hw queues
> >>>>>> - hw queue depth
> >>>>>> - will NVMe sort/merge IO among all SQs or not?
> >>>>>>
> >>>>>>>
> >>>>>>>
> >>>>>>> Discussion Proposal:
> >>>>>>>
> >>>>>>> We’d like to share our views and solicit input on:
> >>>>>>>
> >>>>>>> -What Linux storage stack assumptions do we need to be aware of as we
> >>>>>>> develop these devices with drastically different performance
> >>>>>>> characteristics than traditional NAND? For example, what schedular or
> >>>>>>> device driver level changes will be needed to integrate NVMe HDDs?
> >>>>>>
> >>>>>> IO merge is often important for HDD. IO merge is usually triggered when
> >>>>>> .queue_rq() returns STS_RESOURCE, so far this condition won't be
> >>>>>> triggered for NVMe SSD.
> >>>>>>
> >>>>>> Also blk-mq kills BDI queue congestion and ioc batching, and causes
> >>>>>> writeback performance regression[1][2].
> >>>>>>
> >>>>>> What I am thinking is that if we need to switch to use independent IO
> >>>>>> path for handling SSD and HDD. IO, given the two mediums are so
> >>>>>> different from performance viewpoint.
> >>>>>>
> >>>>>> [1] https://urldefense.proofpoint.com/v2/url?u=https-3A__lore.kernel.org_linux-2Dscsi_Pine.LNX.4.44L0.1909181213141.1507-2D100000-40iolanthe.rowland.org_&d=DwIFaQ&c=IGDlg0lD0b-nebmJJ0Kp8A&r=NW1X0yRHNNEluZ8sOGXBxCbQJZPWcIkPT0Uy3ynVsFU&m=pSnHpt_uQQ73JV4VIQg1C_PVAcLvqBBtmyxQHwWjGSM&s=tsnFP8bQIAq7G66B75LTe3vo4K14HbL9JJKsxl_LPAw&e=
> >>>>>> [2] https://urldefense.proofpoint.com/v2/url?u=https-3A__lore.kernel.org_linux-2Dscsi_20191226083706.GA17974-40ming.t460p_&d=DwIFaQ&c=IGDlg0lD0b-nebmJJ0Kp8A&r=NW1X0yRHNNEluZ8sOGXBxCbQJZPWcIkPT0Uy3ynVsFU&m=pSnHpt_uQQ73JV4VIQg1C_PVAcLvqBBtmyxQHwWjGSM&s=GJwSxXtc_qZHKnrTqSbytUjuRrrQgZpvV3bxZYFDHe4&e=
> >>>>>>
> >>>>>>
> >>>>>> Thanks,
> >>>>>> Ming
> >>>>>>
> >>>>>
> >>>>> I would expect the drive would support a reasonable number of queues
> >>>>> and a relatively deep queue depth, more in line with NVMe practices
> >>>>> than SAS HDD's typical 128. But it probably doesn't make sense to
> >>>>> queue up thousands of commands on something as slow as an HDD, and
> >>>>> many customers keep queues < 32 for latency management.
> >>>>
> >>>> Exposing an HDD through multiple-queues each with a high queue depth is
> >>>> simply asking for troubles. Commands will end up spending so much time
> >>>> sitting in the queues that they will timeout. This can already be observed
> >>>> with the smartpqi SAS HBA which exposes single drives as multiqueue block
> >>>> devices with high queue depth. Exercising these drives heavily leads to
> >>>> thousands of commands being queued and to timeouts. It is fairly easy to
> >>>> trigger this without a manual change to the QD. This is on my to-do list of
> >>>> fixes for some time now (lacking time to do it).
> >>>
> >>> Just wondering why smartpqi SAS won't set one proper .cmd_per_lun for
> >>> avoiding the issue, looks the driver simply assigns .can_queue to it,
> >>> then it isn't strange to see the timeout issue. If .can_queue is a bit
> >>> big, HDD. is easily saturated too long.
> >>>
> >>>>
> >>>> NVMe HDDs need to have an interface setup that match their speed, that is,
> >>>> something like a SAS interface: *single* queue pair with a max QD of 256 or
> >>>> less depending on what the drive can take. Their is no TASK_SET_FULL
> >>>> notification on NVMe, so throttling has to come from the max QD of the SQ,
> >>>> which the drive will advertise to the host.
> >>>>
> >>>>> Merge and elevator are important to HDD performance. I don't believe
> >>>>> NVMe should attempt to merge/sort across SQs. Can NVMe merge/sort
> >>>>> within a SQ without driving large differences between SSD & HDD data
> >>>>> paths?
> >>>>
> >>>> As far as I know, there is no merging going on once requests are passed to
> >>>> the driver and added to an SQ. So this is beside the point.
> >>>> The current default block scheduler for NVMe SSDs is "none". This is
> >>>> decided based on the number of queues of the device. For NVMe drives that
> >>>> have only a single queue *AND* the QUEUE_FLAG_NONROT flag cleared in their
> >>>> request queue will can fallback to the default spinning rust mq-deadline
> >>>> elevator. That will achieve command merging and LBA ordering needed for
> >>>> good performance on HDDs.
> >>>
> >>> mq-deadline basically won't merge IO if STS_RESOURCE isn't returned from
> >>> .queue_rq(), or blk_mq_get_dispatch_budget always return true. NVMe's
> >>> .queue_rq() basically always returns STS_OK.
> >>
> >> I am confused: when an elevator is set, ->queue_rq() is called for requests
> >> obtained from the elevator (with e->type->ops.dispatch_request()), after
> >> the requests went through it. And merging will happen at that stage when
> >> new requests are inserted in the elevator.
> > 
> > When request is queued to lld via .queue_rq(), the request has been
> > removed from scheduler queue. And IO merge is just done inside or
> > against scheduler queue.
> 
> Yes, for incoming new BIOs, not for requests passed to the LLD.
> 
> >> If the ->queue_rq() returns BLK_STS_RESOURCE or BLK_STS_DEV_RESOURCE, the
> >> request is indeed requeued which offer more chances of further merging, but
> >> that is not the same as no merging happening.
> >> Am I missing your point here ?
> > 
> > BLK_STS_RESOURCE or BLK_STS_DEV_RESOURCE or getting no budget can be
> > thought as device saturation feedback, then more requests can be
> > gathered in scheduler queue since we don't dequeue request from
> > scheduler queue when that happens, then IO merge is possible.
> > 
> > Without any device saturation feedback from driver, block layer just
> > dequeues request from scheduler queue with same speed of submission to
> > hardware, then no IO can be merged.
> 
> Got it. And since queue full will mean no more tags, submission will block
> on get_request() and there will be no chance in the elevator to merge
> anything (aside from opportunistic merging in plugs), isn't it ?
> So I guess NVMe HDDs will need some tuning in this area.

scheduler queue depth is usually 2 times of hw queue depth, so requests
ar usually enough for merging.

For NVMe, there isn't ns queue depth, such as scsi's device queue depth,
meantime the hw queue depth is big enough, so no chance to trigger merge.

Thanks,
Ming