[PATCHv2 0/7] dma mapping optimisations

[Jens Axboe]

On 8/2/22 1:36 PM, Keith Busch wrote:
> device undergoes various represenations for every IO. Each consumes
> memory and CPU cycles. When the backing storage is NVMe, the sequence
> looks something like the following:
> 
>   __user void *
>   struct iov_iter
>   struct pages[]
>   struct bio_vec[]
>   struct scatterlist[]
>   __le64[]
> 
> Applications will often use the same buffer for many IO, though, so
> these potentially costly per-IO transformations to reach the exact same
> hardware descriptor can be skipped.
> 
> The io_uring interface already provides a way for users to register
> buffers to get to the 'struct bio_vec[]'. That still leaves the
> scatterlist needed for the repeated dma_map_sg(), then transform to
> nvme's PRP list format.
> 
> This series takes the registered buffers a step further. A block driver
> can implement a new .dma_map() callback to complete the representation
> to the hardware's DMA mapped address, and return a cookie so a user can
> reference it later for any given IO. When used, the block stack can skip
> significant amounts of code, improving CPU utilization, and, if not
> bandwidth limited, IOPs.
> 
> The implementation is currently limited to mapping a registered buffer
> to a single file.

I ran this on my test box to see how we'd do. First the bad news:
smaller block size IO seems slower. I ran with QD=8 and used 24 drives,
and using t/io_uring (with registered buffers, polled, etc) and a 512b
block size I get:

IOPS=44.36M, BW=21.66GiB/s, IOS/call=1/1
IOPS=44.64M, BW=21.80GiB/s, IOS/call=2/2
IOPS=44.69M, BW=21.82GiB/s, IOS/call=1/1
IOPS=44.55M, BW=21.75GiB/s, IOS/call=2/2
IOPS=44.93M, BW=21.94GiB/s, IOS/call=1/1
IOPS=44.79M, BW=21.87GiB/s, IOS/call=1/2

and adding -D1 I get:

IOPS=43.74M, BW=21.36GiB/s, IOS/call=1/1
IOPS=44.04M, BW=21.50GiB/s, IOS/call=1/1
IOPS=43.63M, BW=21.30GiB/s, IOS/call=2/2
IOPS=43.67M, BW=21.32GiB/s, IOS/call=1/1
IOPS=43.57M, BW=21.28GiB/s, IOS/call=1/2
IOPS=43.53M, BW=21.25GiB/s, IOS/call=2/1

which does regress that workload. Since we avoid more expensive setup at
higher block sizes, I tested that too. Here's using 128k IOs with -D0:

OPS=972.18K, BW=121.52GiB/s, IOS/call=31/31
IOPS=988.79K, BW=123.60GiB/s, IOS/call=31/31
IOPS=990.40K, BW=123.80GiB/s, IOS/call=31/31
IOPS=987.80K, BW=123.48GiB/s, IOS/call=31/31
IOPS=988.12K, BW=123.52GiB/s, IOS/call=31/31

and here with -D1:

IOPS=978.36K, BW=122.30GiB/s, IOS/call=31/31
IOPS=996.75K, BW=124.59GiB/s, IOS/call=31/31
IOPS=996.55K, BW=124.57GiB/s, IOS/call=31/31
IOPS=996.52K, BW=124.56GiB/s, IOS/call=31/31
IOPS=996.54K, BW=124.57GiB/s, IOS/call=31/31
IOPS=996.51K, BW=124.56GiB/s, IOS/call=31/31

which is a notable improvement. Then I checked CPU utilization,
switching to IRQ driven IO instead. And the good news there for bs=128K
we end up using half the CPU to achieve better performance. So definite
win there!

Just a quick dump on some quick result, I didn't look further into this
just yet.

-- 
Jens Axboe