Linux kernel mirror (for testing) git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git
kernel os linux

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>

authored by

Yamin Friedman and committed by
Jason Gunthorpe
da662979 f4915455

+49
+45
drivers/infiniband/core/cq.c
··· 18 18 #define IB_POLL_FLAGS \ 19 19 (IB_CQ_NEXT_COMP | IB_CQ_REPORT_MISSED_EVENTS) 20 20 21 + static void ib_cq_rdma_dim_work(struct work_struct *w) 22 + { 23 + struct dim *dim = container_of(w, struct dim, work); 24 + struct ib_cq *cq = dim->priv; 25 + 26 + u16 usec = rdma_dim_prof[dim->profile_ix].usec; 27 + u16 comps = rdma_dim_prof[dim->profile_ix].comps; 28 + 29 + dim->state = DIM_START_MEASURE; 30 + 31 + cq->device->ops.modify_cq(cq, comps, usec); 32 + } 33 + 34 + static void rdma_dim_init(struct ib_cq *cq) 35 + { 36 + struct dim *dim; 37 + 38 + if (!cq->device->ops.modify_cq || !cq->device->use_cq_dim || 39 + cq->poll_ctx == IB_POLL_DIRECT) 40 + return; 41 + 42 + dim = kzalloc(sizeof(struct dim), GFP_KERNEL); 43 + if (!dim) 44 + return; 45 + 46 + dim->state = DIM_START_MEASURE; 47 + dim->tune_state = DIM_GOING_RIGHT; 48 + dim->profile_ix = RDMA_DIM_START_PROFILE; 49 + dim->priv = cq; 50 + cq->dim = dim; 51 + 52 + INIT_WORK(&dim->work, ib_cq_rdma_dim_work); 53 + } 54 + 21 55 static int __ib_process_cq(struct ib_cq *cq, int budget, struct ib_wc *wcs, 22 56 int batch) 23 57 { ··· 112 78 static int ib_poll_handler(struct irq_poll *iop, int budget) 113 79 { 114 80 struct ib_cq *cq = container_of(iop, struct ib_cq, iop); 81 + struct dim *dim = cq->dim; 115 82 int completed; 116 83 117 84 completed = __ib_process_cq(cq, budget, cq->wc, IB_POLL_BATCH); ··· 121 86 if (ib_req_notify_cq(cq, IB_POLL_FLAGS) > 0) 122 87 irq_poll_sched(&cq->iop); 123 88 } 89 + 90 + if (dim) 91 + rdma_dim(dim, completed); 124 92 125 93 return completed; 126 94 } ··· 143 105 if (completed >= IB_POLL_BUDGET_WORKQUEUE || 144 106 ib_req_notify_cq(cq, IB_POLL_FLAGS) > 0) 145 107 queue_work(cq->comp_wq, &cq->work); 108 + else if (cq->dim) 109 + rdma_dim(cq->dim, completed); 146 110 } 147 111 148 112 static void ib_cq_completion_workqueue(struct ib_cq *cq, void *private) ··· 200 160 goto out_free_wc; 201 161 202 162 rdma_restrack_kadd(&cq->res); 163 + 164 + rdma_dim_init(cq); 203 165 204 166 switch (cq->poll_ctx) { 205 167 case IB_POLL_DIRECT: ··· 265 223 266 224 rdma_restrack_del(&cq->res); 267 225 cq->device->ops.destroy_cq(cq, udata); 226 + if (cq->dim) 227 + cancel_work_sync(&cq->dim->work); 228 + kfree(cq->dim); 268 229 kfree(cq->wc); 269 230 kfree(cq); 270 231 }
+4
include/rdma/ib_verbs.h
··· 61 61 #include <linux/cgroup_rdma.h> 62 62 #include <linux/irqflags.h> 63 63 #include <linux/preempt.h> 64 + #include <linux/dim.h> 64 65 #include <uapi/rdma/ib_user_verbs.h> 65 66 #include <rdma/rdma_counter.h> 66 67 #include <rdma/restrack.h> ··· 1510 1509 struct work_struct work; 1511 1510 }; 1512 1511 struct workqueue_struct *comp_wq; 1512 + struct dim *dim; 1513 1513 /* 1514 1514 * Implementation details of the RDMA core, don't use in drivers: 1515 1515 */ ··· 2578 2576 u16 is_switch:1; 2579 2577 /* Indicates kernel verbs support, should not be used in drivers */ 2580 2578 u16 kverbs_provider:1; 2579 + /* CQ adaptive moderation (RDMA DIM) */ 2580 + u16 use_cq_dim:1; 2581 2581 u8 node_type; 2582 2582 u8 phys_port_cnt; 2583 2583 struct ib_device_attr attrs;