[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <d0912e0a-4aaf-43e0-96d8-9ba2e5b620d5@linux.dev>
Date: Sat, 21 Sep 2024 21:43:12 +0800
From: Zhu Yanjun <yanjun.zhu@...ux.dev>
To: Håkon Bugge <haakon.bugge@...cle.com>,
Jason Gunthorpe <jgg@...pe.ca>, Leon Romanovsky <leon@...nel.org>,
Allison Henderson <allison.henderson@...cle.com>,
"David S . Miller" <davem@...emloft.net>, Eric Dumazet
<edumazet@...gle.com>, Jakub Kicinski <kuba@...nel.org>,
Paolo Abeni <pabeni@...hat.com>
Cc: linux-rdma@...r.kernel.org, linux-kernel@...r.kernel.org,
netdev@...r.kernel.org, rds-devel@....oracle.com
Subject: Re: [MAINLINE 2/2] rds: ib: Add Dynamic Interrupt Moderation to CQs
在 2024/9/18 16:35, Håkon Bugge 写道:
> With the support from ib_core to use Dynamic Interrupt Moderation
> (DIM) from legacy ULPs, which uses ib_create_cq(), we enable that
> feature for the receive and send CQs in RDS.
>
> A set of rds-stress runs have been done. bcopy read + write for
> payload 8448 and 16640 bytes and ack/req of 256 bytes. Number of QPs
> varies from 8 to 128, number of threads (i.e. rds-stress processes)
> from one to 16 and a depth of four. A limit has been applied such that
> the number of processes times the number of QPs never exceeds 128. All
> in all, 61 rds-stress runs.
>
> For brevity, only the rows showing a +/- 3% deviation or larger from
> base is listed. The geometric mean of the ratios (IOPS_test /
> IOPS_base) is calculated for all 61 runs, and that gives the best
> possible "average" impact of the commits.
>
> In the following, "base" is v6.11-rc7. "test" is the same
> kernel with the following two commits:
>
> * rds: ib: Add Dynamic Interrupt Moderation to CQs (this commit)
> * RDMA/core: Enable legacy ULPs to use RDMA DIM
>
> This is executed between two X8-2 with CX-5 using fw 16.35.3502. These
> BM systems were instantiated with one VF, which were used for the
> test:
>
> base test
> ACK REQ QPS THR DEP IOPS IOPS Percent
> 256 8448 8 1 4 634463 658162 3.7
> 256 8448 8 2 4 862648 997358 15.6
> 256 8448 8 4 4 950458 1113991 17.2
> 256 8448 8 8 4 932120 1127024 20.9
> 256 8448 8 16 4 944977 1133885 20.0
> 8448 256 8 2 4 858663 975563 13.6
> 8448 256 8 4 4 934884 1098854 17.5
> 8448 256 8 8 4 928247 1116015 20.2
> 8448 256 8 16 4 938864 1123455 19.7
> 256 8448 64 1 4 965985 918445 -4.9
> 8448 256 64 1 4 963280 918239 -4.7
> 256 16640 8 2 4 544670 582330 6.9
> 256 16640 8 4 4 554873 597553 7.7
> 256 16640 8 8 4 551799 597479 8.3
> 256 16640 8 16 4 553041 597898 8.1
> 16640 256 8 2 4 544644 578331 6.2
> 16640 256 8 4 4 553944 594627 7.3
> 16640 256 8 8 4 551388 594737 7.9
> 16640 256 8 16 4 552986 596581 7.9
> Geometric mean of ratios: 1.03
>
> Signed-off-by: Håkon Bugge <haakon.bugge@...cle.com>
> ---
> net/rds/ib_cm.c | 10 ++++++++++
> 1 file changed, 10 insertions(+)
>
> diff --git a/net/rds/ib_cm.c b/net/rds/ib_cm.c
> index 26b069e1999df..79603d86b6c02 100644
> --- a/net/rds/ib_cm.c
> +++ b/net/rds/ib_cm.c
> @@ -259,6 +259,7 @@ static void rds_ib_cq_comp_handler_recv(struct ib_cq *cq, void *context)
> static void poll_scq(struct rds_ib_connection *ic, struct ib_cq *cq,
> struct ib_wc *wcs)
> {
> + int ncompleted = 0;
> int nr, i;
> struct ib_wc *wc;
>
> @@ -276,7 +277,10 @@ static void poll_scq(struct rds_ib_connection *ic, struct ib_cq *cq,
> rds_ib_mr_cqe_handler(ic, wc);
>
> }
> + ncompleted += nr;
> }
> + if (cq->dim)
> + rdma_dim(cq->dim, ncompleted);
> }
>
> static void rds_ib_tasklet_fn_send(unsigned long data)
> @@ -304,6 +308,7 @@ static void poll_rcq(struct rds_ib_connection *ic, struct ib_cq *cq,
> struct ib_wc *wcs,
> struct rds_ib_ack_state *ack_state)
> {
> + int ncompleted = 0;
> int nr, i;
> struct ib_wc *wc;
>
> @@ -316,7 +321,10 @@ static void poll_rcq(struct rds_ib_connection *ic, struct ib_cq *cq,
>
> rds_ib_recv_cqe_handler(ic, wc, ack_state);
> }
> + ncompleted += nr;
> }
> + if (cq->dim)
> + rdma_dim(cq->dim, ncompleted);
> }
>
> static void rds_ib_tasklet_fn_recv(unsigned long data)
> @@ -542,6 +550,7 @@ static int rds_ib_setup_qp(struct rds_connection *conn)
> ic->i_scq_vector = ibdev_get_unused_vector(rds_ibdev);
> cq_attr.cqe = ic->i_send_ring.w_nr + fr_queue_space + 1;
> cq_attr.comp_vector = ic->i_scq_vector;
> + cq_attr.flags |= IB_CQ_MODERATE;
cq_attr.flags is added IB_CQ_MODERATE here.
> ic->i_send_cq = ib_create_cq(dev, rds_ib_cq_comp_handler_send,
> rds_ib_cq_event_handler, conn,
> &cq_attr);
> @@ -556,6 +565,7 @@ static int rds_ib_setup_qp(struct rds_connection *conn)
> ic->i_rcq_vector = ibdev_get_unused_vector(rds_ibdev);
> cq_attr.cqe = ic->i_recv_ring.w_nr;
> cq_attr.comp_vector = ic->i_rcq_vector;
> + cq_attr.flags |= IB_CQ_MODERATE;
Why is cq_attr.flags add IB_CQ_MODERATE again?
Is this cq_attr.flags changed before this line?
Zhu Yanjun
> ic->i_recv_cq = ib_create_cq(dev, rds_ib_cq_comp_handler_recv,
> rds_ib_cq_event_handler, conn,
> &cq_attr);
Powered by blists - more mailing lists