linux-rdma / perftest Goto Github PK
View Code? Open in Web Editor NEWInfiniband Verbs Performance Tests
License: Other
Infiniband Verbs Performance Tests
License: Other
I get immediate failures on the latest perftest when running ib_send_bw or ib_write_bw over either mlx5 or cxgb4. git bisect points to this commit as the culprit:
[root@stevo1 perftest]# ./ib_send_bw -a -R -d cxgb4_0 172.16.2.2
Couldn't allocate MR
failed to create mr
Failed to create MR
Unable to create the resources needed by comm struct
Unable to perform rdma_client function
Unable to init the socket connection
[root@stevo1 perftest]# git bisect bad
3d9815e is the first bad commit
commit 3d9815e
Author: Zohar Ben Aharon [email protected]
Date: Mon Feb 26 14:14:48 2018 +0200
Fix for commit a9df5ed7f04484b6838ee2b48246be1ab66513f1
FreeBSD not supporting contig memory allocation
Signed-off-by: Zohar Ben Aharon <[email protected]>
:040000 040000 f01b099fc410e017aadb5967c2e0cdb815ba6c52 b3c0d0589c2daff7acc14b752ad1f4175e2d673f M src
[root@stevo1 perftest]#
perftest fails to build from source on mips*/armel/armhf (on Debian):
[...]
make all-am
make[2]: Entering directory '/«BUILDDIR»/perftest-3.4+0.6.gc3435c2'
gcc -DHAVE_CONFIG_H -I. -Wdate-time -D_FORTIFY_SOURCE=2 -g -Wall -D_GNU_SOURCE -O3 -g -O2 -fdebug-prefix-map=/«BUILDDIR»/perftest-3.4+0.6.gc3435c2=. -fstack-protector-strong -Wformat -Werror=format-security -c -o src/get_clock.o src/get_clock.c
In file included from /usr/include/arm-linux-gnueabihf/sys/time.h:21:0,
from src/get_clock.c:43:
/usr/include/features.h:148:3: warning: #warning "_BSD_SOURCE and _SVID_SOURCE are deprecated, use _DEFAULT_SOURCE" [-Wcpp]
# warning "_BSD_SOURCE and _SVID_SOURCE are deprecated, use _DEFAULT_SOURCE"
^~~~~~~
In file included from src/get_clock.c:48:0:
src/get_clock.h:101:2: warning: #warning get_cycles not implemented for this architecture: attempt asm/timex.h [-Wcpp]
#warning get_cycles not implemented for this architecture: attempt asm/timex.h
^~~~~~~
src/get_clock.h:102:23: fatal error: asm/timex.h: No such file or directory
#include <asm/timex.h>
^
compilation terminated.
Makefile:786: recipe for target 'src/get_clock.o' failed
make[2]: *** [src/get_clock.o] Error 1
Debian bug: https://bugs.debian.org/863158
I have 3 nodes, Mellanox ConnectX-5 ethernet card with RoCe. Nodes are A, B, C.
I start two copies of ib_send_bw on node A (server side):
Then on node B, I start ib_send_bw (client side) to connect to port 5001 on node A:
again, on node C, I start another ib_send_bw (client side) to connect to port 5002 on node A:
10.254.28.96 is one of the ip address on node A.
The performance is very poor:
Node B output:
4194304 4826 0.00 965.20 0.000241
4194304 7695 0.00 1539.00 0.000385
4194304 2596 0.00 519.20 0.000130
4194304 7186 0.00 1437.20 0.000359
4194304 14261 0.00 2852.20 0.000713
4194304 6801 0.00 1360.20 0.000340
Node C output:
4194304 2456 0.00 491.19 0.000123
4194304 162 0.00 32.40 0.000008
4194304 4125 0.00 824.99 0.000206
4194304 6177 0.00 1235.39 0.000309
4194304 498 0.00 99.60 0.000025
4194304 80 0.00 16.00 0.000004
If I either start ib_send_bw on node B or on node C, it can generate very good number:
4194304 14612 0.00 11689.36 0.002922
4194304 14612 0.00 11689.35 0.002922
4194304 14612 0.00 11689.35 0.002922
4194304 14612 0.00 11689.35 0.002922
4194304 14612 0.00 11689.36 0.002922
Or, I exchange the client/server role (reverse the traffic direction), the BW is evenly split:
4194304 7306 0.00 5844.68 0.001461
4194304 7306 0.00 5844.68 0.001461
4194304 7306 0.00 5844.68 0.001461
4194304 7309 0.00 5847.08 0.001462
4194304 7306 0.00 5844.68 0.001461
What I hoped is that the BW is evenly split in the poor performance case.
Want to get confirm if this is problem of ib_send_bw itself, or general RoCE problem.
reporter: [email protected]
int destroy_ctx(struct pingpong_context *ctx,
struct perftest_parameters *user_param)
{
int i, first, dereg_counter, rc;
int test_result = 0;
int num_of_qps = user_param->num_of_qps;
if (user_param->wait_destroy) {
printf(" Waiting %u seconds before releasing resources...\n",
user_param->wait_destroy);
sleep(user_param->wait_destroy);
}
......
}
The field has default value 0, and I want to know why should I wait for x seconds before destroy_ctx()
. Does this have any special significance?
The perftest_communication is written so that the timing between server performing rdma_listen
and client calling rdma_connect is sometimes out of sync.
The connect flow implements a retry attempt if the rdma_connect fails, but the retry fails
on a return code before and the retry attempt eventually does not happen
repro steps:
This was exposed on iWARP, where the rdma_listen takes longer.
Can be easily reproducible by delaying the rdma_listen, or simply running the client before server, no retry occurs. (This also happens on RoCE)
I test ib_send_bw for two case on ARM services and IB card.
Case1:numactl -C 64 ib_send_bw -d mlx5_0 -a --iters=1000
Case2:numactl -C 64 ib_send_bw -d mlx5_0 --size=1024 --iters=1000
Case1 perfermance data:
#bytes #iterations BW peak[MB/sec] BW average[MB/sec] MsgRate[Mpps]
1024 1000 0.00 5544.28 5.677340
Case2 perfermance data:
#bytes #iterations BW peak[MB/sec] BW average[MB/sec] MsgRate[Mpps]
1024 1000 0.00 920.13 0.942218
The performance of the case1 is 6 times of case2 performance. What may cause the performance gap?
I encountered an error when using the following perftest commands to do an environment test:
server side: ib_write_bw -d mlx5_0 -D 10
client side: ib_write_bw -d mlx5_0 -D 10 10.0.13.4
I reinstalled the OFED driver without any modification to the default configuration, but the same error still occurs. I then noticed that these two machines are using different gid index by default. The error is fixed by specifying the gid index as follows:
server side: ib_write_bw -d mlx5_0 -D 10 -x 3
client side: ib_write_bw -d mlx5_0 -D 10 10.0.13.4 -x 3
My question: Why gid index matters? And why the default gid index is different?
Hi all:
I'm very interested in your work. And I personally tested --rate_limit
using a Python script between two machines in an IB network. The network traffic limits were set ranging from 10MBps to 5000MBps+. Whether it is ib_send_bw
,ib_read_bw
or ib_write_bw
, a stable relative error was observed.
The rate_limit_type is SW.:
I wonder what mechanism causes this stable relative error, thanks!
perftest should support benchmarking of these new kinds of memory.
There are two basic variants of CUDA Unified Memory:
On IBM machines based POWER9, where the GPU is attached to the CPU via NVLINK, e.g. AC922 servers, the CUDA runtime supports GPUDirect RDMA on both variants.
For that to work, ODP must be enabled.
On other systems, like x86_64, support is still missing.
Currently lintian complains about missing man pages for following binaries:
Please write man pages for these tools.
The reason is this commit: 239df74. ctx->rx_buffer_addr
isn't initialized for the send_lat test. I'm not sure if the fix is just to add if (user_param->tst == BW)
before it as I'm not clear on what ctx->rx_buffer_addr
is for.
Apologies if this isn't how issues are reported. Let me know what I should do instead if that is the case.
I run ib_send_lat on AWS EFA card, but server desn't recv any packet
client: sudo ./ib_send_lat -d efa_0 -s a 172.31.6.59 -c SRD
server: sudo ./ib_send_lat -d efa_0 -s a -c SRD
alloc_pd_err :0
alloc_ucontext_err :0
cmds_err :0
completed_cmds :105
create_ah_err :0
create_cq_err :0
create_qp_err :0
keep_alive_rcvd :87
lifespan :12
mmap_err :0
no_completion_cmds :0
rdma_read_bytes :0
rdma_read_resp_bytes :0
rdma_read_wr_err :0
rdma_read_wrs :0
recv_bytes :0
recv_wrs :0
reg_mr_err :0
rx_bytes :0
rx_drops :0
rx_pkts :0
send_bytes :2
send_wrs :1
submitted_cmds :108
tx_bytes :2
tx_pkts :1
alloc_pd_err :0
alloc_ucontext_err :0
cmds_err :0
completed_cmds :39
create_ah_err :0
create_cq_err :0
create_qp_err :0
keep_alive_rcvd :91
lifespan :12
mmap_err :0
no_completion_cmds :0
rdma_read_bytes :0
rdma_read_resp_bytes :0
rdma_read_wr_err :0
rdma_read_wrs :0
recv_bytes :0
recv_wrs :0
reg_mr_err :0
rx_bytes :0
rx_drops :0
rx_pkts :0
send_bytes :0
send_wrs :0
submitted_cmds :42
tx_bytes :0
tx_pkts :0
Hello!
I was wondering if the Device Memory API (ibv_alloc_dm
, ibv_memcpy_to_dm
, ibv_reg_dm_mr
, etc.) could be added to the IB perftests as a runtime option? I was trying to implement it myself, but there's a snag about local protection errors that I'm currently debugging (I have an idea, but I just need the time to get to it).
Thank you!
Hi,
I see that with multiple QP support to perftest (-q), single uber CQ is created and shared by all the QPs in an instance (ib_read_lat/ib_write_bw etc.). If more QPs are used per instance say -q 1000, CQ becomes too big and run out of max device supported cqe per CQ and may fail creating CQ or fail polling the CQ with limited size (device supported max size) for more number QPs.
Any reason behind using single large CQ for multiple QPs? Please let me know if creating CQ per QP makes sense.
Thanks
Hi,
I want to run ib_write_bw
with service level 2
when I am running based on rdma_cm
(i.e., with -R
), but it doesn't take effect. When I remove -R
, it takes effect.
So -R
and -S
cannot be used simultaneously? How can I specify service level
with rdma_cm
?
When run with use_cuda, perftests fail with error "Couldn't allocate MR":
root@u1:/dev/shm/pt/bin# /dev/shm/pt/bin/ib_write_bw -d mlx5_3 -a --use_cuda
************************************
* Waiting for client to connect... *
************************************
initializing CUDA
There are 2 devices supporting CUDA, picking first...
[pid = 4869, dev = 0] device name = [Tesla K80]
creating CUDA Ctx
making it the current CUDA Ctx
cuMemAlloc() of a 16777216 bytes GPU buffer
allocated GPU buffer address at 0000002305040000 pointer=0x2305040000
Couldn't allocate MR
failed to create mr
Failed to create MR
Couldn't create IB resources
Without use_cuda, the test works well:
root@u1:/dev/shm/pt/bin# /dev/shm/pt/bin/ib_write_bw -d mlx5_3 -a
************************************
* Waiting for client to connect... *
************************************
---------------------------------------------------------------------------------------
RDMA_Write BW Test
Dual-port : OFF Device : mlx5_3
Number of qps : 1 Transport type : IB
Connection type : RC Using SRQ : OFF
CQ Moderation : 100
Mtu : 4096[B]
Link type : IB
Max inline data : 0[B]
rdma_cm QPs : OFF
Data ex. method : Ethernet
---------------------------------------------------------------------------------------
local address: LID 0x0e QPN 0x0086 PSN 0xb72682 RKey 0x0056fe VAddr 0x007fb28132d000
remote address: LID 0x10 QPN 0x0086 PSN 0x316d10 RKey 0x003aae VAddr 0x007f895f3a0000
---------------------------------------------------------------------------------------
#bytes #iterations BW peak[MB/sec] BW average[MB/sec] MsgRate[Mpps]
8388608 5000 11682.08 11672.99 0.001459
---------------------------------------------------------------------------------------
===================================================
The code was built by
CUDA_H_PATH=/usr/local/cuda/include/cuda.h ./configure --prefix=/dev/shm/pt/; make -j 20 install
root@u1:/dev/shm/perftest# git show
commit 198473181e0365f97c5840b8fd406ff52af6335b
Author: Zohar Ben Aharon <[email protected]>
Date: Wed Mar 14 10:33:59 2018 +0200
Software stack is as following:
root@u2:/dev/shm/pt/bin# ofed_info -s
MLNX_OFED_LINUX-4.3-1.0.1.0:
root@u2:/dev/shm/pt/bin# uname -a
Linux u2 4.4.0-119-generic #143-Ubuntu SMP Mon Apr 2 16:08:24 UTC 2018 x86_64 x86_64 x86_64 GNU/Linux
root@u2:/dev/shm/pt/bin# /usr/local/cuda/bin/nvcc --version
nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2017 NVIDIA Corporation
Built on Fri_Nov__3_21:07:56_CDT_2017
Cuda compilation tools, release 9.1, V9.1.85
CA 'mlx5_3'
CA type: MT4119
Number of ports: 1
Firmware version: 16.22.1002
Hardware version: 0
Node GUID: 0x248a070300a80a00
System image GUID: 0x248a070300a80a00
Port 1:
State: Active
Physical state: LinkUp
Rate: 100
Base lid: 16
LMC: 0
SM lid: 1
Capability mask: 0x2651e848
Port GUID: 0x248a070300a80a00
Link layer: InfiniBand
Any ideas?
Thanks
Hi,
I'm trying to run a latency test for the first time but I get an error when I try to execute it.
My steps:
Downloaded and unzipped the repo on the box on which I want to run the server for the test
Executed:
./autogen
./configure
make clean && make V=1
./ib_send_lat --duration=30 -H
Port number 1 state is Down
Couldn't set the link layer
Couldn't get context for the device
What am I doing wrong?
Running other tests using RDMA (eg. the ones in Accelio) the system shows no problems.
Thanks
ibv_devices
device node GUID
------ ----------------
mlx5_2 0000000000000000
mlx5_0 506b4b0300285028
mlx5_3 0000000000000000
mlx5_1 506b4b0300285034
address 1.1.1.2/24 defined for mlx5_2
run: ib_read_bw --report_gbits -s 1048576 -D5 -F -R
ibv_devices
device node GUID
------ ----------------
mlx5_0 248a0703004bf0a8
mlx5_1 0000000000000000
address 1.1.1.1/24 defined for mlx5_0
run: ib_read_bw --report_gbits -s 1048576 -D5 -F -R 1.1.1.2
server (running ib_read_bw on VF) failed with segV fault when the client tries to connect.
Program terminated with signal SIGSEGV, Segmentation fault.
#0 0x00007f7c7f0dfe64 in ibv_alloc_pd () from /lib64/libibverbs.so.1
(gdb) bt
#0 0x00007f7c7f0dfe64 in ibv_alloc_pd () from /lib64/libibverbs.so.1
#1 0x0000000000417267 in ctx_init (ctx=0x1c75e10, user_param=0x1c75ad0)
at src/perftest_resources.c:1795
#2 0x0000000000406ca4 in rdma_server_connect (ctx=0x1c75e10, user_param=0x1c75ad0)
at src/perftest_communication.c:1110
#3 0x00000000004074ea in establish_connection (comm=0x7ffc9c28f320)
at src/perftest_communication.c:1243
#4 0x0000000000403230 in main (argc=7, argv=0x7ffc9c28f8d8) at src/read_bw.c:115
Further debugging:
1102 ctx->context = ctx->cm_id->verbs;
(gdb) p ctx->context
$15 = (struct ibv_context *) 0x0
see
perftest/src/perftest_resources.c
Line 4222 in 6369e62
Hi all,
With latest perftest while running ib_write_bw the below errors are seen on client side
Failed to disconnect RDMA CM connection.
ERRNO: Connection reset by peer.
Failed to disconnect RDMA CM nodes.
No errors observed in dmesg
Current behavior:
Server
#ib_write_bw -n 10 -R -s 2048 -D 10 -p 11000
************************************
* Waiting for client to connect... *
************************************
---------------------------------------------------------------------------------------
RDMA_Write BW Test
Dual-port : OFF Device : cxgb4_0
Number of qps : 1 Transport type : IW
Connection type : RC Using SRQ : OFF
PCIe relax order: ON
ibv_wr* API : OFF
CQ Moderation : 100
Mtu : 1024[B]
Link type : Ethernet
GID index : 0
Max inline data : 0[B]
rdma_cm QPs : ON
Data ex. method : rdma_cm
---------------------------------------------------------------------------------------
Waiting for client rdma_cm QP to connect
Please run the same command with the IB/RoCE interface IP
---------------------------------------------------------------------------------------
local address: LID 0000 QPN 0x040a PSN 0x95f56e
GID: 00:07:67:60:01:112:00:00:00:00:00:00:00:00:00:00
remote address: LID 0000 QPN 0x040a PSN 0xc476a6
GID: 00:07:67:62:204:144:00:00:00:00:00:00:00:00:00:00
---------------------------------------------------------------------------------------
#bytes #iterations BW peak[MB/sec] BW average[MB/sec] MsgRate[Mpps]
2048 6143300 0.00 1999.78 1.023887
---------------------------------------------------------------------------------------
Client
------
#ib_write_bw -n 10 -R -s 2048 -D 10 -p 11000 102.1.1.245
---------------------------------------------------------------------------------------
RDMA_Write BW Test
Dual-port : OFF Device : cxgb4_0
Number of qps : 1 Transport type : IW
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 : 0
Max inline data : 0[B]
rdma_cm QPs : ON
Data ex. method : rdma_cm
---------------------------------------------------------------------------------------
local address: LID 0000 QPN 0x040a PSN 0xc476a6
GID: 00:07:67:62:204:144:00:00:00:00:00:00:00:00:00:00
remote address: LID 0000 QPN 0x040a PSN 0x95f56e
GID: 00:07:67:60:01:112:00:00:00:00:00:00:00:00:00:00
---------------------------------------------------------------------------------------
#bytes #iterations BW peak[MB/sec] BW average[MB/sec] MsgRate[Mpps]
Conflicting CPU frequency values detected: 2636.683000 != 2179.229000. CPU Frequency is not max.
2048 6143300 0.00 1999.78 1.023887
---------------------------------------------------------------------------------------
Failed to disconnect RDMA CM connection.
ERRNO: Connection reset by peer.
Failed to disconnect RDMA CM nodes.
Expected behavior
Error messages on the client should not be seen
Observation:
Context:
OS: RHEL 7.9 (3.10.0-1160.el7.x86_64)
perftest version : tot(7504ce4)
Thanks
Smit Kothari
Reported by the UNH-IOL validation group: "When testing with OFED-4.17-rc2 on Kernel 4.17.14 with iWARP, RDMA fails with a variety of vendor cards with all having the same error below on "small/large rdma send. To confirm this is an OFED-4.17-rc2 issue I backtracked to OFED-4.17-rc1 on Kernel 4.17.14 and saw a full pass.
Commands used:
server: ib_send_bw -d i40iw1 -i 1 -s 1 -n 25000 -m 2048 --dont_xchg_versions -R -x 0 -r 510 -F
client: ib_send_bw -d qedr0 -i 1 -s 1 -n 25000 -m 2048 aegaeon-iw.ofa.iol.unh.edu -F --dont_xchg_versions -R -x 0 -r 510
Failed to deallocate PD - Device or resource busy
Couldn't destroy all SEND resources"
It has been confirmed that this is a perftest regression with perftest-4.4-0.5 included with RC2 and OFED 4.17 GA. The perftest-4.4-0.3 that was included with RC1 is working and it also works with OFED 4.17 GA.
It happened that we run your tests on RISC-V arch and found out that some metrics on the final output are zero. Delving into the code showed that proc_get_cpu_mhz() always returns 0, just because this function reads CPU MHz information from /proc/cpuinfo but only a few arches print out this information into this file. To be honest it is difficult to add sth into /proc/cpuinfo.
See requests to solve this problem:
Why not to get this information from /sys/devices/system/cpu/cpufreq/ ?
It is a portable and clear way to get it.
I would like to revise the code a little bit to add my feature, and I want to use some containers from C++ STL. So how can I compile the code using C++?
i want to test multicast. 1-to-2
server1:
ib_send_bw -d mlx5_1 -F -x 3 -c UD --run_infinitely --report_gbits -M 225.1.1.20
server2:
ib_send_bw -d mlx5_1 -F -x 3 -c UD --run_infinitely --report_gbits -M 225.1.1.20
client:
ib_send_bw -d mlx5_1 -F -x 3 -c UD --run_infinitely --report_gbits -g -M 225.1.1.20
and in this way, i failed.
i don't know correct parameters to test multicast. can anyone help me?
I see the following errors on running ib_read_bw or ib_write_bw with cxgb4.
It is seen with the commit that adds multiple QP support for rdma-cm 5d1895ae602
.
[root@saptharishi perftest]# ./ib_write_bw -R -F -n10 -q1026 -s32 --report_gbits 102.1.1.11
RDMA CM event error:
Event: RDMA_CM_EVENT_REJECTED; error: -111.
ERRNO: No such file or directory.
Failed to handle RDMA CM event.
ERRNO: No such file or directory.
Failed to connect RDMA CM events.
ERRNO: No such file or directory.
rdma-core-24.0-1.el7.x86_64
# uname -r
4.19.36
Hi,
we tried to test GPUDirect RDMA.
Test pod deployed from https://github.com/Mellanox/k8s-images
we deployed 2 pods:
Server pod:
root@rdma-cuda-test-pod-1:~# ib_write_bw -d mlx5_0 -F -R -q 2 --use_cuda=0
Client pod:
root@rdma-cuda-test-pod-1:~# ib_write_bw -d mlx5_0 -F -R -q 2 --use_cuda=0 192.168.111.1
initializing CUDA
Listing all CUDA devices in system:
CUDA device 0: PCIe address is 02:00
Picking device No. 0
[pid = 56, dev = 0] device name = [NVIDIA A30-8C]
creating CUDA Ctx
making it the current CUDA Ctx
cuMemAlloc() of a 262144 bytes GPU buffer
allocated GPU buffer address at 0000010013000000 pointer=0x10013000000
Couldn't allocate MR
failed to create mr
Failed to create MR
Failed to initialize RDMA contexts.
ERRNO: Bad address.
Failed to handle RDMA CM event.
ERRNO: Bad address.
Failed to connect RDMA CM events.
ERRNO: Bad address.
Segmentation fault (core dumped)
what does "Couldn't allocate MR" mean?
thanks in advance
According to my understanding, rdma read and rdma write is the same except data flow direction. I think latency of ib_read_lat should be equal to that of ib_write_lat.
Test command: numactl --cpunodebind=0 --membind=0 ./ib_read/write_lat -d mlx5_0 -R ip_address
Result:
Are there any other reasons that may cause the difference?
I noticed in perftest_parameters.c:1367: fprintf(stderr," Perftest supports CUDA only in BW tests\n");
It seems to work if I uncomment for read_lat, but I wanted to see if there is a good not to use the test?
In the spec file, we have:
Source: http://www.openfabrics.org/downloads/%{name}-%{version}.tar.gz
Url: http://www.openfabrics.org
But OpenFabrics points to GitHub in https://openfabrics.org/downloads/MAINTAINERS.
The spec file also indicates that this is v4.2-%{release} but it doesn't look like a matching %{name}-%{version}.tar.gz (perftest-4.2.tar.gz) is available as a downloadable artifact from GitHub.
I modified the code to print the device name through the parameters inside the struct ibv_context
.I found the device name printed by ctx_print_test_info
is not the same as the parameters inside the struct ibv_context
, further I found ctx_print_test_info
just print the user parameter... Why is this?
My code is like this:
The execution result is as follows
server:
Looking forward to reply from you when it is convenient, any reply will be helpful for me, thanks!!
This is bad, as people are convinced to be running a build with GPU support, so for example they report unreasonable performance.
Version used
lehi-smoke:~ # ib_send_lat --version
Version: 5.60
lehi-smoke:~ # ib_send_bw --version
Version: 5.60
lehi-smoke:~ #
Test Method
4 hosts with MLX HCA inter-connected with ethernet switches.
Each host launches 12 server processes (3 ib_send_bw, 3 ib_write_bw, 3 ib_send_lat, 3 ib_write_lat).
Each host launches 12 client processes (3 ib_send_bw, 3 ib_write_bw, 3 ib_send_lat, 3 ib_write_lat).
It was a full-cross test, each host act as both server and client for other 3 hosts.
And no rate limitation was used.
Problem
Negative value observed in the log of some bw or lat process.
Following is some of the logs.
Fri, 20 Mar 2020 22:35:36 +0800
Send BW Test
Mon, 23 Mar 2020 07:35:37 +0800
Fri, 20 Mar 2020 22:34:20 +0800
Send BW Test
Mon, 23 Mar 2020 07:35:47 +0800
Fri, 20 Mar 2020 22:34:20 +0800
Send Latency Test
Mon, 23 Mar 2020 07:35:48 +0800
Fri, 20 Mar 2020 22:34:20 +0800
Send Latency Test
Mon, 23 Mar 2020 07:35:57 +0800
I'm using running ib_read_lat between two different HCAs, and modify QP to RTS fails on the side with a lower max_qp_rd_atom because it appears that the code uses the out_reads of the peer when modifying the QP (I think in ctx_modify_qp_to_rts, where we set attr->max_rd_atomic = dest->out_reads;
). Shouldn't this just be the device's max_rd_atomic?
Hi, when running latency tests with "iterations" (-n) flag, the tool reports min/max/avg/std_dev/99/99.9 latencies. The same latency test when run with "duration" (-D) flag reports only avg latency. How to get the same report for "duration" based tests?
Iteration test : ib_write_lat -F -d mlx5_0 -I 64 -n 100000 -s 1 20.2.1.2
Duration test : ib_write_lat -F -d mlx5_0 -I 64 -D 20 -s 1 20.2.1.2
We are testing ibverbs on newly developed IB driver on our microblaze(32 bit) platform with xilinx rdma nic HW. While running ib_read_bw (server on microblaze and client on mlnx connectx-4), I found the client feeding a swapped VA in RETH. I narrowed it down to struct pingpong_dest. The size of this structure is 52bytes on 32bit system and 56bytes on 64bit x86 system with connect-x4.
I did a quick test with attribute((packed)) macro on struct pingpong_dest and I now see correct VA coming back from client via RETH.
ib_read_bw --version shows 5.5. In config.h I have HAVE_ENDIAN 1.
hello, I failed gpu direct rdma test over infiniband after upgrading OFED version.
I tested several times, and
got completion with error:
00000000 00000000 00000000 00000000
00000000 00000000 00000000 00000000
00000000 00000000 00000000 00000000
00000000 01005104 08001a78 000013d2
Completion with error at client
Failed status 4: wr_id 0 syndrom 0x51
scnt=128, ccnt=0
Failed to complete run_iter_bw function successfully
status 4 indicates IBV_WC_LOC_PROT_ERR, and it looks like
"This event is generated when a user attempts to access an address outside of the registered memory region."
from https://www.ibm.com/docs/en/sdk-java-technology/8?topic=jverbs-workcompletionstatus
without GPU direct, by deleting the --use_cuda=0 argument, it worked well.
I would really appreciate any help.
The full log is given as
$./runme remote_host numactl -N 3 ib_write_bw -a -b --use_cuda=0 -d mlx5_1
initializing CUDA
Listing all CUDA devices in system:
CUDA device 0: PCIe address is 07:00
CUDA device 1: PCIe address is 0B:00
CUDA device 2: PCIe address is 48:00
CUDA device 3: PCIe address is 4C:00
CUDA device 4: PCIe address is 88:00
CUDA device 5: PCIe address is 8B:00
CUDA device 6: PCIe address is C8:00
CUDA device 7: PCIe address is CB:00
Picking device No. 0
[pid = 574, dev = 0] device name = [A100 Graphics Device]
creating CUDA Ctx
making it the current CUDA Ctx
cuMemAlloc() of a 16777216 bytes GPU buffer
allocated GPU buffer address at 00007f7b4e000000 pointer=0x7f7b4e000000
---------------------------------------------------------------------------------------
RDMA_Write Bidirectional BW Test
Dual-port : OFF Device : mlx5_1
Number of qps : 1 Transport type : IB
Connection type : RC Using SRQ : OFF
PCIe relax order: ON
ibv_wr* API : ON
TX depth : 128
CQ Moderation : 100
Mtu : 4096[B]
Link type : IB
Max inline data : 0[B]
rdma_cm QPs : OFF
Data ex. method : Ethernet
---------------------------------------------------------------------------------------
local address: LID 0x07 QPN 0x1a78 PSN 0xbe6767 RKey 0x00cf39 VAddr 0x007f7b4e800000
remote address: LID 0x3f QPN 0x1544 PSN 0xebc222 RKey 0x00ba05 VAddr 0x007f670a800000
---------------------------------------------------------------------------------------
#bytes #iterations BW peak[MB/sec] BW average[MB/sec] MsgRate[Mpps]
mlx5: test-sleep-worker-0: got completion with error:
00000000 00000000 00000000 00000000
00000000 00000000 00000000 00000000
00000000 00000000 00000000 00000000
00000000 01005104 08001544 000020d2
Completion with error at client
Failed status 4: wr_id 0 syndrom 0x51
scnt=128, ccnt=0
Failed to complete run_iter_bw function successfully
************************************
* Waiting for client to connect... *
************************************
initializing CUDA
Listing all CUDA devices in system:
CUDA device 0: PCIe address is 07:00
CUDA device 1: PCIe address is 0B:00
CUDA device 2: PCIe address is 48:00
CUDA device 3: PCIe address is 4C:00
CUDA device 4: PCIe address is 88:00
CUDA device 5: PCIe address is 8B:00
CUDA device 6: PCIe address is C8:00
CUDA device 7: PCIe address is CB:00
Picking device No. 0
[pid = 872, dev = 0] device name = [A100 Graphics Device]
creating CUDA Ctx
making it the current CUDA Ctx
cuMemAlloc() of a 16777216 bytes GPU buffer
allocated GPU buffer address at 00007f670a000000 pointer=0x7f670a000000
---------------------------------------------------------------------------------------
RDMA_Write Bidirectional BW Test
Dual-port : OFF Device : mlx5_1
Number of qps : 1 Transport type : IB
Connection type : RC Using SRQ : OFF
PCIe relax order: ON
ibv_wr* API : ON
TX depth : 128
CQ Moderation : 100
Mtu : 4096[B]
Link type : IB
Max inline data : 0[B]
rdma_cm QPs : OFF
Data ex. method : Ethernet
---------------------------------------------------------------------------------------
local address: LID 0x3f QPN 0x1544 PSN 0xebc222 RKey 0x00ba05 VAddr 0x007f670a800000
remote address: LID 0x07 QPN 0x1a78 PSN 0xbe6767 RKey 0x00cf39 VAddr 0x007f7b4e800000
---------------------------------------------------------------------------------------
#bytes #iterations BW peak[MB/sec] BW average[MB/sec] MsgRate[Mpps]
mlx5: test-sleep-master-0: got completion with error:
00000000 00000000 00000000 00000000
00000000 00000000 00000000 00000000
00000000 00000000 00000000 00000000
00000000 01005104 08001a78 000013d2
Completion with error at client
Failed status 4: wr_id 0 syndrom 0x51
scnt=128, ccnt=0
Failed to complete run_iter_bw function successfully
And the below is ofed information
air@cl-rndcgpu-a12:~$ ofed_info
MLNX_OFED_LINUX-5.3-1.0.0.1 (OFED-5.3-1.0.0):
ar_mgr:
osm_plugins/ar_mgr/ar_mgr-1.0-5.8.2.MLNX20210321.g58d33bf.tar.gz
clusterkit:
/sw/release/mlnx_ofed/IBHPC/MLNX_OFED_LINUX-5.2-1.0.4.0/SRPMS/clusterkit-1.0.36-1.52104.src.rpm
dapl:
/sw/release/mlnx_ofed/IBHPC/MLNX_OFED_LINUX-5.2-1.0.4.0/SRPMS/dapl-2.1.10.1.mlnx-OFED.4.9.0.1.4.52104.src.rpm
dpcp:
/sw/release/sw_acceleration/dpcp/dpcp-1.1.2-1.src.rpm
dump_pr:
osm_plugins/dump_pr//dump_pr-1.0-5.8.2.MLNX20210321.g58d33bf.tar.gz
fabric-collector:
/sw/release/mlnx_ofed/IBHPC/MLNX_OFED_LINUX-5.2-1.0.4.0/SRPMS/fabric-collector-1.1.0.MLNX20170103.89bb2aa-0.1.52104.src.rpm
hcoll:
/sw/release/mlnx_ofed/IBHPC/MLNX_OFED_LINUX-5.2-1.0.4.0/SRPMS/hcoll-4.7.3189-1.52104.src.rpm
ibdump:
/sw/release/mlnx_ofed/IBHPC/MLNX_OFED_LINUX-5.2-1.0.4.0/SRPMS/ibdump-6.0.0-1.52104.src.rpm
ibsim:
/sw/release/mlnx_ofed/IBHPC/MLNX_OFED_LINUX-5.2-1.0.4.0/SRPMS/ibsim-0.10-1.52104.src.rpm
ibutils2:
ibutils2/ibutils2-2.1.1-0.132.MLNX20210321.g84ba964.tar.gz
iser:
mlnx_ofed/mlnx-ofa_kernel-4.0.git mlnx_ofed_5_3
commit 19d48787e91c989a7a0a856d04d0c80bca648423
isert:
mlnx_ofed/mlnx-ofa_kernel-4.0.git mlnx_ofed_5_3
commit 19d48787e91c989a7a0a856d04d0c80bca648423
kernel-mft:
mlnx_ofed_mft/kernel-mft-4.16.3-12.src.rpm
knem:
/sw/release/mlnx_ofed/IBHPC/MLNX_OFED_LINUX-5.2-1.0.4.0/SRPMS/knem-1.1.4.90mlnx1-OFED.5.1.2.5.0.1.src.rpm
libvma:
/sw/release/mlnx_ofed/IBHPC/MLNX_OFED_LINUX-5.2-1.0.4.0/SRPMS/libvma-9.2.2-0.src.rpm
mlnx-dpdk:
https://github.com/Mellanox/dpdk.org mlnx_dpdk_20.11_last_stable
commit f05f7352848c8906a9d8dfd607687d52b5f7fcaf
mlnx-en:
mlnx_ofed/mlnx-ofa_kernel-4.0.git mlnx_ofed_5_3
commit 19d48787e91c989a7a0a856d04d0c80bca648423
mlnx-ethtool:
/sw/release/mlnx_ofed/IBHPC/MLNX_OFED_LINUX-5.2-1.0.4.0/SRPMS/mlnx-ethtool-5.8-1.52104.src.rpm
mlnx-iproute2:
mlnx_ofed/iproute2.git mlnx_ofed_5_3
commit 130d9fd2320a4ece4a6fc330624be2bc341f569a
mlnx-nfsrdma:
mlnx_ofed/mlnx-ofa_kernel-4.0.git mlnx_ofed_5_3
commit 19d48787e91c989a7a0a856d04d0c80bca648423
mlnx-nvme:
mlnx_ofed/mlnx-ofa_kernel-4.0.git mlnx_ofed_5_3
commit 19d48787e91c989a7a0a856d04d0c80bca648423
mlnx-ofa_kernel:
mlnx_ofed/mlnx-ofa_kernel-4.0.git mlnx_ofed_5_3
commit 19d48787e91c989a7a0a856d04d0c80bca648423
mpi-selector:
/sw/release/mlnx_ofed/IBHPC/MLNX_OFED_LINUX-5.2-1.0.4.0/SRPMS/mpi-selector-1.0.3-1.52104.src.rpm
mpitests:
/sw/release/mlnx_ofed/IBHPC/MLNX_OFED_LINUX-5.2-1.0.4.0/SRPMS/mpitests-3.2.20-5d20b49.52104.src.rpm
mstflint:
mlnx_ofed_mstflint/mstflint-4.16.0-1.tar.gz
multiperf:
/sw/release/mlnx_ofed/IBHPC/MLNX_OFED_LINUX-5.2-1.0.4.0/SRPMS/multiperf-3.0-0.14.g5f0fd0e.52104.src.rpm
ofed-docs:
docs.git mlnx_ofed-4.0
commit 3d1b0afb7bc190ae5f362223043f76b2b45971cc
openmpi:
/sw/release/mlnx_ofed/IBHPC/MLNX_OFED_LINUX-5.2-1.0.4.0/SRPMS/openmpi-4.1.0rc5-1.52104.src.rpm
opensm:
mlnx_ofed_opensm/opensm-5.8.2.MLNX20210321.2958ab8.tar.gz
openvswitch:
openvswitch.git mlnx_ofed_5_3
commit b5c772fad3a4b3345805fb4dc19b9e181931b23d
perftest:
mlnx_ofed_perftest/perftest-4.5-0.1.g23b8f9c.tar.gz
rdma-core:
mlnx_ofed/rdma-core.git mlnx_ofed_5_3
commit 696ecf5f1be2fbc7a3b72787c6a4d455d8f47904
rshim:
mlnx_ofed_soc/rshim-2.0.5-10.g0ae03b4.src.rpm
sharp:
mlnx_ofed_sharp/sharp-2.4.5.MLNX20210302.7c3c223.tar.gz
sockperf:
/sw/release/mlnx_ofed/IBHPC/MLNX_OFED_LINUX-5.2-1.0.4.0/SRPMS/sockperf-3.7-0.gita1e8e835a689.52104.src.rpm
srp:
mlnx_ofed/mlnx-ofa_kernel-4.0.git mlnx_ofed_5_3
commit 19d48787e91c989a7a0a856d04d0c80bca648423
ucx:
mlnx_ofed_ucx/ucx-1.10.0-1.src.rpm
xpmem:
xpmem.git mellanox-master
commit 740afed1b7647b241a5a3a57b13ce11ec882d9d5
Installed Packages:
-------------------
ii ar-mgr 1.0-5.8.2.MLNX20210321.g58d33bf.53100 amd64 Adaptive Routing Manager
ii dapl2-utils 2.1.10.1.mlnx-OFED.4.9.0.1.4.53100 amd64 Utilities for use with the DAPL libraries
ii dpcp 1.1.2-1.53100 amd64 Direct Packet Control Plane (DPCP) is a library to use Devx
ii dump-pr 1.0-5.8.2.MLNX20210321.g58d33bf.53100 amd64 Dump PathRecord Plugin
ii hcoll 4.7.3189-1.53100 amd64 Hierarchical collectives (HCOLL)
ii ibacm 52mlnx1-1.53100 amd64 InfiniBand Communication Manager Assistant (ACM)
ii ibdump 6.0.0-1.53100 amd64 Mellanox packets sniffer tool
ii ibsim 0.10-1.53100 amd64 InfiniBand fabric simulator for management
ii ibsim-doc 0.10-1.53100 all documentation for ibsim
ii ibutils2 2.1.1-0.132.MLNX20210321.g84ba964.53100 amd64 OpenIB Mellanox InfiniBand Diagnostic Tools
ii ibverbs-providers:amd64 52mlnx1-1.53100 amd64 User space provider drivers for libibverbs
ii ibverbs-utils 52mlnx1-1.53100 amd64 Examples for the libibverbs library
ii infiniband-diags 52mlnx1-1.53100 amd64 InfiniBand diagnostic programs
ii iser-dkms 5.3-OFED.5.3.0.3.8.1 all DKMS support fo iser kernel modules
ii isert-dkms 5.3-OFED.5.3.0.3.8.1 all DKMS support fo isert kernel modules
ii kernel-mft-dkms 4.16.3-12 all DKMS support for kernel-mft kernel modules
ii knem 1.1.4.90mlnx1-OFED.5.1.2.5.0.1 amd64 userspace tools for the KNEM kernel module
ii knem-dkms 1.1.4.90mlnx1-OFED.5.1.2.5.0.1 all DKMS support for mlnx-ofed kernel modules
ii libdapl-dev 2.1.10.1.mlnx-OFED.4.9.0.1.4.53100 amd64 Development files for the DAPL libraries
ii libdapl2 2.1.10.1.mlnx-OFED.4.9.0.1.4.53100 amd64 The Direct Access Programming Library (DAPL)
ii libibmad-dev:amd64 52mlnx1-1.53100 amd64 Development files for libibmad
ii libibmad5:amd64 52mlnx1-1.53100 amd64 Infiniband Management Datagram (MAD) library
ii libibnetdisc5:amd64 52mlnx1-1.53100 amd64 InfiniBand diagnostics library
ii libibumad-dev:amd64 52mlnx1-1.53100 amd64 Development files for libibumad
ii libibumad3:amd64 52mlnx1-1.53100 amd64 InfiniBand Userspace Management Datagram (uMAD) library
ii libibverbs-dev:amd64 52mlnx1-1.53100 amd64 Development files for the libibverbs library
ii libibverbs1:amd64 52mlnx1-1.53100 amd64 Library for direct userspace use of RDMA (InfiniBand/iWARP)
ii libibverbs1-dbg:amd64 52mlnx1-1.53100 amd64 Debug symbols for the libibverbs library
ii libopensm 5.8.2.MLNX20210321.2958ab8-0.1.53100 amd64 Infiniband subnet manager libraries
ii libopensm-devel 5.8.2.MLNX20210321.2958ab8-0.1.53100 amd64 Developement files for OpenSM
ii librdmacm-dev:amd64 52mlnx1-1.53100 amd64 Development files for the librdmacm library
ii librdmacm1:amd64 52mlnx1-1.53100 amd64 Library for managing RDMA connections
ii mlnx-ethtool 5.8-1.53100 amd64 This utility allows querying and changing settings such as speed,
ii mlnx-iproute2 5.10.0-1.53100 amd64 This utility allows querying and changing settings such as speed,
ii mlnx-ofed-kernel-dkms 5.3-OFED.5.3.1.0.0.1 all DKMS support for mlnx-ofed kernel modules
ii mlnx-ofed-kernel-utils 5.3-OFED.5.3.1.0.0.1 amd64 Userspace tools to restart and tune mlnx-ofed kernel modules
ii mpitests 3.2.20-5d20b49.53100 amd64 Set of popular MPI benchmarks and tools IMB 2018 OSU benchmarks ver 4.0.1 mpiP-3.3 IPM-2.0.6
ii mstflint 4.16.0-1.53100 amd64 Mellanox firmware burning application
ii openmpi 4.1.0rc5-1.53100 all Open MPI
ii opensm 5.8.2.MLNX20210321.2958ab8-0.1.53100 amd64 An Infiniband subnet manager
ii opensm-doc 5.8.2.MLNX20210321.2958ab8-0.1.53100 amd64 Documentation for opensm
ii perftest 4.5-0.1.g23b8f9c.53100 amd64 Infiniband verbs performance tests
ii rdma-core 52mlnx1-1.53100 amd64 RDMA core userspace infrastructure and documentation
ii rdmacm-utils 52mlnx1-1.53100 amd64 Examples for the librdmacm library
ii rshim 2.0.5-10.g0ae03b4.53100 amd64 driver for Mellanox BlueField SoC
ii sharp 2.4.5.MLNX20210302.7c3c223-1.53100 amd64 SHArP switch collectives
ii srp-dkms 5.3-OFED.5.3.0.3.8.1 all DKMS support fo srp kernel modules
ii srptools 52mlnx1-1.53100 amd64 Tools for Infiniband attached storage (SRP)
ii ucx 1.10.0-1.53100 amd64 Unified Communication X
In perftest, default WRITE_INLINE is 220:
perftest/src/perftest_parameters.h
Line 137 in ed18d32
However, 220 max_inline_data could be too large for some devices (e.g., Intel E810 only support max_inline_data less than 102).
Pertest has option (-I
) to change the value, but users may not be aware of the option, and output is a bit unfriendly.
Output of ./ib_write_lat -s 64 -c RC -d irdma0 -n 1000 -i 1
:
************************************
* Waiting for client to connect... *
************************************
Unable to create QP.
Failed to create QP.
Couldn't create IB resources
I got this configure: error: pciutils header files not found, consider installing pciutils-devel when compiling the perftest.
~/perftest# ./autogen.sh && ./configure CUDA_H_PATH=/usr/local/cuda/include/cuda.h && make -j
libtoolize: putting auxiliary files in AC_CONFIG_AUX_DIR, 'config'.
libtoolize: copying file 'config/ltmain.sh'
libtoolize: putting macros in AC_CONFIG_MACRO_DIRS, 'm4'.
libtoolize: copying file 'm4/libtool.m4'
libtoolize: copying file 'm4/ltoptions.m4'
libtoolize: copying file 'm4/ltsugar.m4'
libtoolize: copying file 'm4/ltversion.m4'
libtoolize: copying file 'm4/lt~obsolete.m4'
libtoolize: 'AC_PROG_RANLIB' is rendered obsolete by 'LT_INIT'
configure.ac:55: installing 'config/compile'
configure.ac:36: installing 'config/missing'
Makefile.am: installing 'config/depcomp'
checking for a BSD-compatible install... /usr/bin/install -c
checking whether build environment is sane... yes
checking for a thread-safe mkdir -p... /bin/mkdir -p
checking for gawk... no
checking for mawk... mawk
checking whether make sets $(MAKE)... yes
checking whether make supports nested variables... yes
checking whether make supports nested variables... (cached) yes
checking for gcc... gcc
checking whether the C compiler works... yes
checking for C compiler default output file name... a.out
checking for suffix of executables...
checking whether we are cross compiling... no
checking for suffix of object files... o
checking whether we are using the GNU C compiler... yes
checking whether gcc accepts -g... yes
checking for gcc option to accept ISO C89... none needed
checking whether gcc understands -c and -o together... yes
checking for style of include used by make... GNU
checking dependency style of gcc... gcc3
checking for g++... g++
checking whether we are using the GNU C++ compiler... yes
checking whether g++ accepts -g... yes
checking dependency style of g++... gcc3
checking dependency style of gcc... gcc3
checking build system type... x86_64-unknown-linux-gnu
checking host system type... x86_64-unknown-linux-gnu
checking how to print strings... printf
checking for a sed that does not truncate output... /bin/sed
checking for grep that handles long lines and -e... /bin/grep
checking for egrep... /bin/grep -E
checking for fgrep... /bin/grep -F
checking for ld used by gcc... /usr/bin/ld
checking if the linker (/usr/bin/ld) is GNU ld... yes
checking for BSD- or MS-compatible name lister (nm)... /usr/bin/nm -B
checking the name lister (/usr/bin/nm -B) interface... BSD nm
checking whether ln -s works... yes
checking the maximum length of command line arguments... 1572864
checking how to convert x86_64-unknown-linux-gnu file names to x86_64-unknown-linux-gnu format... func_convert_file_noop
checking how to convert x86_64-unknown-linux-gnu file names to toolchain format... func_convert_file_noop
checking for /usr/bin/ld option to reload object files... -r
checking for objdump... objdump
checking how to recognize dependent libraries... pass_all
checking for dlltool... no
checking how to associate runtime and link libraries... printf %s\n
checking for ar... ar
checking for archiver @file support... @
checking for strip... strip
checking for ranlib... ranlib
checking command to parse /usr/bin/nm -B output from gcc object... ok
checking for sysroot... no
checking for a working dd... /bin/dd
checking how to truncate binary pipes... /bin/dd bs=4096 count=1
checking for mt... mt
checking if mt is a manifest tool... no
checking how to run the C preprocessor... gcc -E
checking for ANSI C header files... yes
checking for sys/types.h... yes
checking for sys/stat.h... yes
checking for stdlib.h... yes
checking for string.h... yes
checking for memory.h... yes
checking for strings.h... yes
checking for inttypes.h... yes
checking for stdint.h... yes
checking for unistd.h... yes
checking for dlfcn.h... yes
checking for objdir... .libs
checking if gcc supports -fno-rtti -fno-exceptions... no
checking for gcc option to produce PIC... -fPIC -DPIC
checking if gcc PIC flag -fPIC -DPIC works... yes
checking if gcc static flag -static works... yes
checking if gcc supports -c -o file.o... yes
checking if gcc supports -c -o file.o... (cached) yes
checking whether the gcc linker (/usr/bin/ld -m elf_x86_64) supports shared libraries... yes
checking whether -lc should be explicitly linked in... no
checking dynamic linker characteristics... GNU/Linux ld.so
checking how to hardcode library paths into programs... immediate
checking whether stripping libraries is possible... yes
checking if libtool supports shared libraries... yes
checking whether to build shared libraries... yes
checking whether to build static libraries... yes
checking how to run the C++ preprocessor... g++ -E
checking for ld used by g++... /usr/bin/ld -m elf_x86_64
checking if the linker (/usr/bin/ld -m elf_x86_64) is GNU ld... yes
checking whether the g++ linker (/usr/bin/ld -m elf_x86_64) supports shared libraries... yes
checking for g++ option to produce PIC... -fPIC -DPIC
checking if g++ PIC flag -fPIC -DPIC works... yes
checking if g++ static flag -static works... yes
checking if g++ supports -c -o file.o... yes
checking if g++ supports -c -o file.o... (cached) yes
checking whether the g++ linker (/usr/bin/ld -m elf_x86_64) supports shared libraries... yes
checking dynamic linker characteristics... (cached) GNU/Linux ld.so
checking how to hardcode library paths into programs... immediate
checking for ranlib... (cached) ranlib
checking for ANSI C header files... (cached) yes
checking infiniband/verbs.h usability... yes
checking infiniband/verbs.h presence... yes
checking for infiniband/verbs.h... yes
checking for ibv_get_device_list in -libverbs... yes
checking for rdma_create_event_channel in -lrdmacm... yes
checking for umad_init in -libumad... yes
checking for log in -lm... yes
checking pci/pci.h usability... no
checking pci/pci.h presence... no
checking for pci/pci.h... no
configure: error: pciutils header files not found, consider installing pciutils-devel
I am using ubuntu 18.04.
How do I get pciutils-devel installed properly?
Debian uses dashes to separate the upstream version from the Debian revision. Please avoid dashes in version. Is there a reason why you use the dash? Why do you use 4.1-0.2 as version instead of just 4.1 or 4.1.2 or 4.1.0.2?
When running the following, for example:
ib_send_bw -b -e -q 16
I think that the root cause is wrong usage of notifications - the cq isn't emptied when the application is woken up. The argument passed to ibv_poll_cq is user_param->rx_depth, but we might have more completions than that, since we use multiple QPs.
There are two options I can think of for a fix:
During MPA negotiation, when passive side gets MPA receive packet, the RDMA-CM stack
creates connection resources (PD, MR, CQ, QP) and initiates MPA offload.
The active side, who initiated the MPA connection sets a timer that times out in case of huge buffers
(initiating those buffers takes more than reasonable MPA timeout).
I think that in this case, perftest should support an option to skip buffer initialization -
for (i = 0; i < ctx->buff_size; i++) {
((char*)ctx->buf[qp_index])[i] = (char)rand();
send_lat could easily support CUDA device memory as source/sink.
write_lat cannot do the same as easily, as it relies on direct memory polling from the CPU.
If perftest was built with a machine which PCIe Relaxed Ordering compliant, and run over two machines which are NOT PCIe Relaxed Ordering compliant, we will see significantly performance degradation for large size message.
https://bugzilla.redhat.com/show_bug.cgi?id=1902855
The perftest of "ib_send_lat RC" test on data size of 4194304 bytes and 8388608 bytes had test time increase of about 250 fold when compared with the same test on other MLX4 ROCE device, like MLX5 CX-3.
So, should we disable PCIe RO by default, and let the user to enable it or not when run the perftest benchmarks?
Thanks
When we use ib_send_bw or ib_write_bw with about 96 QP q=96 and n=1000, we notice out of sequence or out of buffers or packet sequence errors while running these tests between 2 different machines using mellanox ofed 5.
If we run the client and server tests both on the same machine then we observe no errors.the issue is seen only when we go across machines.
Is that an issue with the code in ibsendbw or ib_write_bw or is the issue some where else?
Thanks
the problem is in run_iter_bw_infinitely(), where the call to pthread_sigmask happens too late, when the CUDA driver has been initialized and its worker thread launched.
the solution is to move the call to pthread_sigmask somewhere before ctx_init().
In ctx_alloc_credit function in perftest_resources.c file, ctrl_buf is allocated a size of user_param->num_of_qps however it is being memset to a size of buf_size:
ALLOCATE(ctx->ctrl_buf,uint32_t,user_param->num_of_qps);
memset(&ctx->ctrl_buf[0],0,buf_size);
Fix this by allocating the right size ctrl_buf:
ALLOCATE(ctx->ctrl_buf,uint32_t,buf_size);
This is output of 'ib_write_bw -a -d mlx5_0 --report_gbits node1', seems to work fine:
[root@node3 bin]# ib_write_bw -a -d mlx5_0 --report_gbits node1
---------------------------------------------------------------------------------------
RDMA_Write BW Test
Dual-port : OFF Device : mlx5_0
Number of qps : 1 Transport type : IB
Connection type : RC Using SRQ : OFF
PCIe relax order: ON
ibv_wr* API : ON
TX depth : 128
CQ Moderation : 100
Mtu : 4096[B]
Link type : IB
Max inline data : 0[B]
rdma_cm QPs : OFF
Data ex. method : Ethernet
---------------------------------------------------------------------------------------
local address: LID 0x02 QPN 0x0032 PSN 0x5f2841 RKey 0x002440 VAddr 0x007f2e5f64b000
remote address: LID 0x01 QPN 0x003a PSN 0xc0fd7e RKey 0x002442 VAddr 0x007f443b2f8000
---------------------------------------------------------------------------------------
#bytes #iterations BW peak[Gb/sec] BW average[Gb/sec] MsgRate[Mpps]
2 5000 0.042350 0.042185 2.636533
4 5000 0.084507 0.084457 2.639276
8 5000 0.17 0.17 2.641399
16 5000 0.34 0.34 2.640952
32 5000 0.68 0.68 2.638957
64 5000 1.35 1.35 2.638629
128 5000 2.71 2.71 2.643606
256 5000 5.42 5.42 2.644112
512 5000 10.78 10.77 2.629186
1024 5000 21.38 21.37 2.608802
2048 5000 42.13 42.09 2.568967
4096 5000 83.97 83.91 2.560721
8192 5000 186.89 149.84 2.286319
16384 5000 195.18 169.98 1.296822
32768 5000 196.21 185.39 0.707209
65536 5000 196.25 190.26 0.362886
131072 5000 196.33 193.93 0.184945
262144 5000 195.49 195.03 0.092996
524288 5000 196.25 196.25 0.046789
1048576 5000 196.48 196.48 0.023422
2097152 5000 196.62 196.59 0.011718
4194304 5000 196.67 196.63 0.005860
8388608 5000 196.63 196.58 0.002929
---------------------------------------------------------------------------------------
But it would fail if I plus '-R', like:
[root@node3 bin]# ib_write_bw -a -d mlx5_0 --report_gbits -R node1
Received 10 times ADDR_ERROR
Unable to perform rdma_client function
Unable to init the socket connection
And I read source code and have known it's caused by RDMA_CM_EVENT_ADDR_ERROR, but I don't known why.
This is output about 'lscpi -vvv':
[root@node3 bin]# lspci -vvv | grep Mellanox -A 65
41:00.0 Infiniband controller: Mellanox Technologies MT28908 Family [ConnectX-6]
Subsystem: Mellanox Technologies Device 0007
Control: I/O- Mem+ BusMaster+ SpecCycle- MemWINV- VGASnoop- ParErr- Stepping- SERR- FastB2B- DisINTx+
Status: Cap+ 66MHz- UDF- FastB2B- ParErr- DEVSEL=fast >TAbort- <TAbort- <MAbort- >SERR- <PERR- INTx-
Latency: 0, Cache Line Size: 64 bytes
Interrupt: pin A routed to IRQ 1125
NUMA node: 0
Region 0: Memory at 2807e000000 (64-bit, prefetchable) [size=32M]
Expansion ROM at b4400000 [disabled] [size=1M]
Capabilities: [60] Express (v2) Endpoint, MSI 00
DevCap: MaxPayload 512 bytes, PhantFunc 0, Latency L0s unlimited, L1 unlimited
ExtTag+ AttnBtn- AttnInd- PwrInd- RBE+ FLReset+ SlotPowerLimit 75.000W
DevCtl: Report errors: Correctable+ Non-Fatal+ Fatal+ Unsupported-
RlxdOrd+ ExtTag+ PhantFunc- AuxPwr- NoSnoop+ FLReset-
MaxPayload 512 bytes, MaxReadReq 512 bytes
DevSta: CorrErr+ UncorrErr- FatalErr- UnsuppReq+ AuxPwr- TransPend-
LnkCap: Port #0, Speed 16GT/s, Width x16, ASPM not supported, Exit Latency L0s unlimited, L1 <4us
ClockPM- Surprise- LLActRep- BwNot- ASPMOptComp+
LnkCtl: ASPM Disabled; RCB 64 bytes Disabled- CommClk+
ExtSynch- ClockPM- AutWidDis- BWInt- AutBWInt-
LnkSta: Speed 16GT/s, Width x16, TrErr- Train- SlotClk+ DLActive- BWMgmt- ABWMgmt-
DevCap2: Completion Timeout: Range ABC, TimeoutDis+, LTR-, OBFF Not Supported
DevCtl2: Completion Timeout: 50us to 50ms, TimeoutDis-, LTR-, OBFF Disabled
LnkCtl2: Target Link Speed: 16GT/s, EnterCompliance- SpeedDis-
Transmit Margin: Normal Operating Range, EnterModifiedCompliance- ComplianceSOS-
Compliance De-emphasis: -6dB
LnkSta2: Current De-emphasis Level: -6dB, EqualizationComplete+, EqualizationPhase1+
EqualizationPhase2+, EqualizationPhase3+, LinkEqualizationRequest-
Capabilities: [48] Vital Product Data
Product Name: ConnectX-6 VPI adapter card, HDR IB (200Gb/s) and 200GbE, single-port QSFP56
Read-only fields:
[PN] Part number: MCX653105A-HDAT
[EC] Engineering changes: AE
[V2] Vendor specific: MCX653105A-HDAT
[SN] Serial number: MT2130T07644
[V3] Vendor specific: 92a87ffbcbeaeb118000b8cef6f7f1c0
[VA] Vendor specific: MLX:MN=MLNX:CSKU=V2:UUID=V3:PCI=V0:MODL=CX653105A
[V0] Vendor specific: PCIeGen4 x16
[VU] Vendor specific: MT2130T07644MLNXS0D0F0
[RV] Reserved: checksum good, 1 byte(s) reserved
End
Capabilities: [9c] MSI-X: Enable+ Count=64 Masked-
Vector table: BAR=0 offset=00002000
PBA: BAR=0 offset=00003000
Capabilities: [c0] Vendor Specific Information: Len=18 <?>
Capabilities: [40] Power Management version 3
Flags: PMEClk- DSI- D1- D2- AuxCurrent=375mA PME(D0-,D1-,D2-,D3hot-,D3cold+)
Status: D0 NoSoftRst+ PME-Enable- DSel=0 DScale=0 PME-
Capabilities: [100 v1] Advanced Error Reporting
UESta: DLP- SDES- TLP- FCP- CmpltTO- CmpltAbrt- UnxCmplt- RxOF- MalfTLP- ECRC- UnsupReq+ ACSViol-
UEMsk: DLP- SDES- TLP- FCP- CmpltTO- CmpltAbrt- UnxCmplt- RxOF- MalfTLP- ECRC- UnsupReq+ ACSViol-
UESvrt: DLP+ SDES- TLP- FCP+ CmpltTO+ CmpltAbrt- UnxCmplt+ RxOF+ MalfTLP+ ECRC+ UnsupReq- ACSViol-
CESta: RxErr- BadTLP- BadDLLP- Rollover- Timeout- NonFatalErr+
CEMsk: RxErr- BadTLP- BadDLLP- Rollover- Timeout- NonFatalErr-
AERCap: First Error Pointer: 04, GenCap+ CGenEn+ ChkCap+ ChkEn+
Capabilities: [150 v1] Alternative Routing-ID Interpretation (ARI)
ARICap: MFVC- ACS-, Next Function: 0
ARICtl: MFVC- ACS-, Function Group: 0
Capabilities: [1c0 v1] #19
Capabilities: [320 v1] #27
Capabilities: [370 v1] #26
Capabilities: [420 v1] #25
Kernel driver in use: mlx5_core
Kernel modules: mlx5_core
42:00.0 Non-Volatile memory controller: Intel Corporation NVMe DC SSD [3DNAND, Beta Rock Controller] (prog-if 02 [NVM Express])
Subsystem: Intel Corporation Device 8008
Any clue about what happened? look forward to your reply, thanks!
client
mlx5 nic
./ib_write_bw -d mlx5_0 -i 1 --use_cuda=0 server_ip_address -a
server
mlx5 NIC
run: ./ib_write_bw -d mlx5_0 -i 1 --use_cuda=0
when pressing ctrl+c to kill the process, the hole system will crash and report system deadlock.
it will not happened if we don't use the param --use_cuda;
perftest/src/perftest_communication.c
Line 519 in 13f7177
The line should read:
memcpy(rem_dest, comm->rdma_ctx->buf[0], sizeof(struct pingpong_dest));
Command that I have tried:
Server:
./ib_send_lat -x 0 -c UD -F -I 0 -a
Client:
./ib_send_lat -x 0 -c UD -F -I 0 -a
Server output:
************************************
* Waiting for client to connect... *
************************************
Unable to create QP.
Failed to create QP.
Couldn't create IB resources
Client output:
Unable to create QP.
Failed to create QP.
Couldn't create IB resources.
I ran “lspci” on the instances where I ran the test and it seems that I have the efa_device installed:
00:00.0 Host bridge: Intel Corporation 440FX - 82441FX PMC [Natoma]
00:01.0 ISA bridge: Intel Corporation 82371SB PIIX3 ISA [Natoma/Triton II]
00:01.3 Non-VGA unclassified device: Intel Corporation 82371AB/EB/MB PIIX4 ACPI (rev 08)
00:03.0 VGA compatible controller: Amazon.com, Inc. Device 1111
00:04.0 Non-Volatile memory controller: Amazon.com, Inc. Device 8061
00:05.0 Ethernet controller: Amazon.com, Inc. Elastic Network Adapter (ENA)
00:06.0 Ethernet controller: Amazon.com, Inc. Elastic Fabric Adapter (EFA)
Then I hacked into the code, and found that ibv_create_qp()
(https://github.com/linux-rdma/perftest/blob/master/src/perftest_resources.c#L1848) is not called.
This is because the HAVE_IBV_WR_API
flag was not set(https://github.com/linux-rdma/perftest/blob/master/src/perftest_resources.c#L1818)
A declarative, efficient, and flexible JavaScript library for building user interfaces.
🖖 Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.
TypeScript is a superset of JavaScript that compiles to clean JavaScript output.
An Open Source Machine Learning Framework for Everyone
The Web framework for perfectionists with deadlines.
A PHP framework for web artisans
Bring data to life with SVG, Canvas and HTML. 📊📈🎉
JavaScript (JS) is a lightweight interpreted programming language with first-class functions.
Some thing interesting about web. New door for the world.
A server is a program made to process requests and deliver data to clients.
Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.
Some thing interesting about visualization, use data art
Some thing interesting about game, make everyone happy.
We are working to build community through open source technology. NB: members must have two-factor auth.
Open source projects and samples from Microsoft.
Google ❤️ Open Source for everyone.
Alibaba Open Source for everyone
Data-Driven Documents codes.
China tencent open source team.