diff options
author | Yamin Friedman | 2019-07-08 13:59:03 +0300 |
---|---|---|
committer | Jason Gunthorpe | 2019-07-08 16:37:22 -0300 |
commit | da6629793aa6944db6c8a908ca1a52d87f1489aa (patch) | |
tree | 80f393dd5133579ea97a408d26a372227a25b066 /include/rdma | |
parent | f4915455dcf07c4f237d6160a4b6adb0575d2909 (diff) |
RDMA/core: Provide RDMA DIM support for ULPs
Added the interface in the infiniband driver that applies the rdma_dim
adaptive moderation. There is now a special function for allocating an
ib_cq that uses rdma_dim.
Performance improvement (ConnectX-5 100GbE, x86) running FIO benchmark over
NVMf between two equal end-hosts with 56 cores across a Mellanox switch
using null_blk device:
READS without DIM:
blk size | BW | IOPS | 99th percentile latency | 99.99th latency
512B | 3.8GiB/s | 7.7M | 1401 usec | 2442 usec
4k | 7.0GiB/s | 1.8M | 4817 usec | 6587 usec
64k | 10.7GiB/s| 175k | 9896 usec | 10028 usec
IO WRITES without DIM:
blk size | BW | IOPS | 99th percentile latency | 99.99th latency
512B | 3.6GiB/s | 7.5M | 1434 usec | 2474 usec
4k | 6.3GiB/s | 1.6M | 938 usec | 1221 usec
64k | 10.7GiB/s| 175k | 8979 usec | 12780 usec
IO READS with DIM:
blk size | BW | IOPS | 99th percentile latency | 99.99th latency
512B | 4GiB/s | 8.2M | 816 usec | 889 usec
4k | 10.1GiB/s| 2.65M| 3359 usec | 5080 usec
64k | 10.7GiB/s| 175k | 9896 usec | 10028 usec
IO WRITES with DIM:
blk size | BW | IOPS | 99th percentile latency | 99.99th latency
512B | 3.9GiB/s | 8.1M | 799 usec | 922 usec
4k | 9.6GiB/s | 2.5M | 717 usec | 1004 usec
64k | 10.7GiB/s| 176k | 8586 usec | 12256 usec
The rdma_dim algorithm was designed to measure the effectiveness of
moderation on the flow in a general way and thus should be appropriate
for all RDMA storage protocols.
rdma_dim is configured to be the default option based on performance
improvement seen after extensive tests.
Signed-off-by: Yamin Friedman <yaminf@mellanox.com>
Reviewed-by: Max Gurtovoy <maxg@mellanox.com>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Diffstat (limited to 'include/rdma')
-rw-r--r-- | include/rdma/ib_verbs.h | 4 |
1 files changed, 4 insertions, 0 deletions
diff --git a/include/rdma/ib_verbs.h b/include/rdma/ib_verbs.h index 4053be51b7fa..c5f8a9f17063 100644 --- a/include/rdma/ib_verbs.h +++ b/include/rdma/ib_verbs.h @@ -61,6 +61,7 @@ #include <linux/cgroup_rdma.h> #include <linux/irqflags.h> #include <linux/preempt.h> +#include <linux/dim.h> #include <uapi/rdma/ib_user_verbs.h> #include <rdma/rdma_counter.h> #include <rdma/restrack.h> @@ -1509,6 +1510,7 @@ struct ib_cq { struct work_struct work; }; struct workqueue_struct *comp_wq; + struct dim *dim; /* * Implementation details of the RDMA core, don't use in drivers: */ @@ -2576,6 +2578,8 @@ struct ib_device { u16 is_switch:1; /* Indicates kernel verbs support, should not be used in drivers */ u16 kverbs_provider:1; + /* CQ adaptive moderation (RDMA DIM) */ + u16 use_cq_dim:1; u8 node_type; u8 phys_port_cnt; struct ib_device_attr attrs; |