On 15/10/2020 22:11, Pavel Begunkov wrote:
> see [3/4] and [4/4] for motivation, the other two are just preps.
> I wasn't expecting to find performance difference, but my naive nop
> test yeilds 5030 vs 5160 KIOPS, before and after [3/4] correspondingly.
Forgot to add that I want to review it thoroughly on silly bugs,
so until then it's more of RFC.
> The test is submitting 32 linked nops and waits for them to complete.
> The testing is tuned for consistentcy, and the results are consistent
> across reboots.
The test was basically like the diff below.
diff --git a/tools/io_uring/io_uring-bench.c b/tools/io_uring/io_uring-bench.c
index 7703f0118385..84c4487c4d4e 100644
--- a/tools/io_uring/io_uring-bench.c
+++ b/tools/io_uring/io_uring-bench.c
@@ -96,13 +96,13 @@ static volatile int finish;
/*
* OPTIONS: Set these to test the various features of io_uring.
*/
-static int polled = 1; /* use IO polling */
+static int polled = 0; /* use IO polling */
static int fixedbufs = 1; /* use fixed user buffers */
static int register_files = 1; /* use fixed files */
static int buffered = 0; /* use buffered IO, not O_DIRECT */
static int sq_thread_poll = 0; /* use kernel submission/poller thread */
static int sq_thread_cpu = -1; /* pin above thread to this CPU */
-static int do_nop = 0; /* no-op SQ ring commands */
+static int do_nop = 1; /* no-op SQ ring commands */
static int io_uring_register_buffers(struct submitter *s)
{
@@ -149,6 +149,7 @@ static void init_io(struct submitter *s, unsigned index)
if (do_nop) {
sqe->opcode = IORING_OP_NOP;
+ sqe->flags = IOSQE_IO_LINK;
return;
}
--
Pavel Begunkov
