Re: [PATCH v6 RESEND] io_uring: releasing CPU resources when polling

[Pavel Begunkov <asml.silence@gmail.com>]

On 7/9/24 10:29, hexue wrote:
> io_uring use polling mode could improve the IO performence, but it will
> spend 100% of CPU resources to do polling.
> 
> This set a signal "IORING_SETUP_HY_POLL" to application, aim to provide
> a interface for user to enable a new hybrid polling at io_uring level.
> 
> A new hybrid poll is implemented on the io_uring layer. Once IO issued,
> it will not polling immediately, but block first and re-run before IO
> complete, then poll to reap IO. This poll function could be a suboptimal
> solution when running on a single thread, it offers the performance lower
> than regular polling but higher than IRQ, and CPU utilization is also lower
> than polling.
> 
> Test Result
> fio-3.35, Gen 4 device
> -------------------------------------------------------------------------------------
> Performance
> -------------------------------------------------------------------------------------
>                    write          read           randwrite       randread
> regular poll    BW=3939MiB/s    BW=6596MiB/s    IOPS=190K       IOPS=526K
> IRQ             BW=3927MiB/s    BW=6567MiB/s    IOPS=181K       IOPS=216K
> hybrid poll     BW=3933MiB/s    BW=6600MiB/s    IOPS=190K       IOPS=390K(suboptimal)
> -------------------------------------------------------------------------------------
> CPU Utilization
> -------------------------------------------------------------------------------------
>                  write   read    randwrite       randread
> regular poll    100%    100%    100%            100%
> IRQ             38%     53%     100%            100%
> hybrid poll     76%     32%     70%              85%
> -------------------------------------------------------------------------------------
> 
> --
...
> diff --git a/io_uring/rw.c b/io_uring/rw.c
> index 1a2128459cb4..5505f4292ce5 100644
> --- a/io_uring/rw.c
> +++ b/io_uring/rw.c
> +static int io_uring_hybrid_poll(struct io_kiocb *req,
> +				struct io_comp_batch *iob, unsigned int poll_flags)
> +{
> +	struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
> +	struct io_ring_ctx *ctx = req->ctx;
> +	int ret;
> +	u64 runtime, sleep_time;
> +
> +	sleep_time = io_delay(ctx, req);
> +
> +	/* it doesn't implement with io_uring passthrough now */
> +	ret = req->file->f_op->iopoll(&rw->kiocb, iob, poll_flags);

->iopoll vs ->uring_cmd_iopoll, same comment as in my
previous review


> +
> +	req->iopoll_end = ktime_get_ns();
> +	runtime = req->iopoll_end - req->iopoll_start - sleep_time;
> +	if (runtime < 0)
> +		return 0;
> +
> +	/* use minimize sleep time if there are different speed
> +	 * drivers, it could get more completions from fast one
> +	 */
> +	if (ctx->available_time > runtime)
> +		ctx->available_time = runtime;
> +	return ret;
> +}
> +
>   int io_do_iopoll(struct io_ring_ctx *ctx, bool force_nonspin)
>   {
>   	struct io_wq_work_node *pos, *start, *prev;
> @@ -1133,7 +1203,9 @@ int io_do_iopoll(struct io_ring_ctx *ctx, bool force_nonspin)
>   		if (READ_ONCE(req->iopoll_completed))
>   			break;
>   
> -		if (req->opcode == IORING_OP_URING_CMD) {
> +		if (ctx->flags & IORING_SETUP_HY_POLL) {
> +			ret = io_uring_hybrid_poll(req, &iob, poll_flags);
> +		} else if (req->opcode == IORING_OP_URING_CMD) {
>   			struct io_uring_cmd *ioucmd;
>   
>   			ioucmd = io_kiocb_to_cmd(req, struct io_uring_cmd);

-- 
Pavel Begunkov