If the rxe driver is able to generate packets faster than the
IP code can process them it will start dropping packets and
ip_local_out() will return NET_XMIT_DROP. The requester side of
the driver detects this and retries the packet. The responder
does not and the requester recovers by taking a retry timer delay
and resubmitting the read operation from the last received packet.
This can and does occur for large RDMA read responses for multi-MB
reads. This causes a steep drop off in performance.
This patch modifies read_reply() in rxe_resp.c to retry the
send if err == -EAGAIN. When IP does drop a packet it requires
more time to recover than a simple retry takes so a subroutine
read_retry_delay() is added that dynamically estimates the time
required for this recovery and inserts a delay before the retry.
With this patch applied the performance of large reads is very
stable. For example with a 1Gb/sec (112.5 GB/sec) Ethernet link
between two systems, without this patch ib_read_bw shows the
following performance.
RDMA_Read BW Test
Dual-port : OFF Device : rxe0
Number of qps : 1 Transport type : IB
Connection type : RC Using SRQ : OFF
PCIe relax order: ON
ibv_wr* API : OFF
TX depth : 128
CQ Moderation : 100
Mtu : 1024[B]
Link type : Ethernet
GID index : 2
Outstand reads : 128
rdma_cm QPs : OFF
Data ex. method : Ethernet
<snip>
#bytes #iterations BW peak[MB/sec] BW average[MB/sec] MsgRate[Mpps]
2 1000 0.56 0.56 0.294363
4 1000 0.66 0.66 0.173862
8 1000 1.32 1.20 0.157406
16 1000 2.66 2.40 0.157357
32 1000 5.54 5.46 0.179006
64 1000 18.22 16.94 0.277533
128 1000 21.61 20.91 0.171322
256 1000 44.02 38.90 0.159316
512 1000 70.39 64.86 0.132843
1024 1000 106.50 100.49 0.102904
2048 1000 106.46 105.29 0.053908
4096 1000 107.85 107.85 0.027609
8192 1000 109.09 109.09 0.013963
16384 1000 110.17 110.17 0.007051
32768 1000 110.27 110.27 0.003529
65536 1000 110.33 110.33 0.001765
131072 1000 110.35 110.35 0.000883
262144 1000 110.36 110.36 0.000441
524288 1000 110.37 110.36 0.000221
1048576 1000 110.37 110.37 0.000110
2097152 1000 24.19 24.10 0.000012
4194304 1000 18.70 18.65 0.000005
8388608 1000 18.09 17.82 0.000002
No NET_XMIT_DROP returns are seen up to 1MiB but at 2MiB and above they
are constant.
With the patch applied ib_read_bw shows the following
performance:
RDMA_Read BW Test
Dual-port : OFF Device : rxe0
Number of qps : 1 Transport type : IB
Connection type : RC Using SRQ : OFF
PCIe relax order: ON
ibv_wr* API : OFF
TX depth : 128
CQ Moderation : 100
Mtu : 1024[B]
Link type : Ethernet
GID index : 2
Outstand reads : 128
rdma_cm QPs : OFF
Data ex. method : Ethernet
<snip>
#bytes #iterations BW peak[MB/sec] BW average[MB/sec] MsgRate[Mpps]
2 1000 0.34 0.33 0.175541
4 1000 0.69 0.68 0.179279
8 1000 2.02 1.75 0.229972
16 1000 2.72 2.63 0.172632
32 1000 5.42 4.94 0.161824
64 1000 10.63 9.67 0.158487
128 1000 31.06 28.11 0.230288
256 1000 40.48 36.75 0.150543
512 1000 70.00 66.00 0.135164
1024 1000 94.43 89.26 0.091402
2048 1000 106.38 104.34 0.053424
4096 1000 109.48 109.16 0.027946
8192 1000 108.96 108.96 0.013946
16384 1000 110.18 110.18 0.007052
32768 1000 110.28 110.28 0.003529
65536 1000 110.33 110.33 0.001765
131072 1000 110.35 110.35 0.000883
262144 1000 110.36 110.35 0.000441
524288 1000 110.35 110.31 0.000221
1048576 1000 110.37 110.37 0.000110
2097152 1000 110.37 110.37 0.000055
4194304 1000 110.37 110.36 0.000028
8388608 1000 110.37 110.37 0.000014
The delay algorithm computes approximately 50 usecs as the correct delay
to insert before retrying a read_reply() send.
Bob Pearson (1):
RDMA-rxe: Allow retry sends for rdma read responses
drivers/infiniband/sw/rxe/rxe_resp.c | 62 +++++++++++++++++++++++++--
drivers/infiniband/sw/rxe/rxe_verbs.h | 9 ++++
2 files changed, 68 insertions(+), 3 deletions(-)
base-commit: 91d088a0304941b88c915cc800617ff4068cdd39
--
2.37.2
