Trex with mlx5 driver - Multiple streams with different VLAN priority causes high CPU utilization

Rubens Figueiredo rubens.figueiredo at bisdn.de
Fri Apr 19 14:44:44 CEST 2024
Previous message (by thread): Trex with mlx5 driver - Multiple streams with different VLAN priority causes high CPU utilization
Next message (by thread): connection tracking in mlx5 not working or displaying strange results
Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]
Thank you for the help!

I ran the commands and traffic is being generated correctly.

Best regards,

Rubens

On 4/19/24 14:31, Dariusz Sosnowski wrote:
> Thank you for the info.
>
> You might have run into an issue, where NIC generates backpressure to the SW because of very frequent switching between different VLAN priorities on Tx datapath in the NIC.
>
> Could you please apply the following QoS configuration on all interfaces and rerun the test with different VLAN priorities?
>
> sudo mlnx_qos -i <iface> --trust=dscp
> for dscp in {0..63}; do sudo mlnx_qos -i <iface> --dscp2prio set,$dscp,0; sleep 0.001;done
>
> These commands will map internally all priorities to priority 0. This workaround should reduce the backpressure without affecting the generated traffic.
>
> Best regards,
> Dariusz Sosnowski
>
>> -----Original Message-----
>> From: Rubens Figueiredo <rubens.figueiredo at bisdn.de>
>> Sent: Thursday, April 18, 2024 15:21
>> To: Dariusz Sosnowski <dsosnowski at nvidia.com>
>> Cc: users at dpdk.org
>> Subject: Re: Trex with mlx5 driver - Multiple streams with different VLAN
>> priority causes high CPU utilization
>>
>> External email: Use caution opening links or attachments
>>
>>
>> Hi Dariusz,
>>
>> Thank you for the help.
>>
>> If the two parallel stream use the same VLAN priority then the issue is not
>> visible anymore. Trex output visible below.
>>
>> Different priority
>> CPU util.  |            96.97% |              0.0% | Tx bps L2  |         3.59 Gbps |          0.29
>> bps | 3.59 Gbps Tx bps L1  |         3.64 Gbps |          0.37 bps | 3.64 Gbps
>> Tx pps     |       298.49 Kpps |             0 pps | 298.49 Kpps
>> Line Util. |            3.64 % |               0 % |
>>
>> Same priority
>> CPU util.  |             0.54% |              0.0% | Tx bps L2  |        23.98 Gbps |             0
>> bps | 23.98 Gbps Tx bps L1  |         24.3 Gbps |             0 bps | 24.3 Gbps
>> Tx pps     |         1.99 Mpps |             0 pps | 1.99 Mpps
>> Line Util. |            24.3 % |               0 % |
>>
>> I have attached the requested output to the email.
>>
>> Best,
>> Rubens
>>
>> On 4/18/24 14:48, Dariusz Sosnowski wrote:
>>> Hi Rubens,
>>>
>>> Would you be able to provide the output of "ethtool -S <iface>" for both VFs
>> before and after the test?
>>> Does the same issue appear on this system if both parallel streams use the
>> same VLAN priority?
>>> Best regards,
>>> Dariusz Sosnowski
>>>
>>>> From: Rubens Figueiredo <rubens.figueiredo at bisdn.de>
>>>> Sent: Wednesday, April 17, 2024 19:07
>>>> To: users at dpdk.org
>>>> Subject: Trex with mlx5 driver - Multiple streams with different VLAN
>>>> priority causes high CPU utilization
>>>>
>>>> Hello community,
>>>> I am facing a strange issue in the Trex stateless code, version v3.02 and
>> v3.04. I am using the Mellanox Cx-5, and have created two VFs on top of the
>> PF 0. The mlx5_core version I am using is the 5.7-1.0.2, and the ofed version is
>> MLNX_OFED_LINUX-5.7-1.0.2.0 (OFED-5.7-1.0.2).
>>>> I have created the following issue in the trex-core repository
>> [here](https://github.com/cisco-system-traffic-generator/trex-
>> core/issues/1124), and was recommended to post the issue in here. In the
>> github issue you see screenshots of the issue I am facing.
>>>> I am trying to create two parallel streams with different VLAN priorities, but
>> the load generated is not what I expect it to be, and CPU util. seems incredibly
>> high (~99%).
>>>> I have reproduced this issue with the --software and non software version.
>>>> The script I used is below.
>>>> import stl_path
>>>> from trex.stl.api import *
>>>>
>>>> import time
>>>> import pprint
>>>> from ipaddress import ip_address, ip_network
>>>>
>>>> import argparse
>>>> import configparser
>>>> import os
>>>> import json
>>>>
>>>>
>>>> def get_packet(tos, mac_dst, ip_src, size):
>>>>       # pkt = Ether(src="02:00:00:00:00:01",dst="00:00:00:01:00:01") /
>>>> IP(src="10.0.0.2", tos=tos) / UDP(sport=4444, dport=4444)
>>>>
>>>>       pkt = (
>>>>           Ether(src="00:01:00:00:00:02", dst=mac_dst)
>>>>           # Ether(dst="11:11:11:11:11:11")
>>>>           # / Dot1AD(vlan=0)
>>>>           / Dot1Q(vlan=0, prio=tos)
>>>>           / IP(src=ip_src)
>>>>           / UDP(sport=4444, dport=4444)
>>>>       )
>>>>       pad = max(0, size - len(pkt)) * "x"
>>>>
>>>>       return pkt / pad
>>>>
>>>> def main():
>>>>       """ """
>>>>       tx_port = 0
>>>>       rx_port = 1
>>>>
>>>>       c = STLClient()
>>>>
>>>>       # connect to server
>>>>       c.connect()
>>>>
>>>>       # prepare our ports
>>>>       c.reset(ports=[tx_port, rx_port])
>>>>
>>>>       streams = []
>>>>       s = STLStream(
>>>>           packet=STLPktBuilder(
>>>>               pkt=get_packet(4,"00:11:22:33:44:55", "10.1.0.2",512),
>>>>               # vm = vm,
>>>>           ),
>>>>           isg=0 * 1000000,
>>>>           mode=STLTXCont(pps=1.2*10**6),
>>>>           # flow_stats = STLFlowLatencyStats(pg_id = 0)
>>>>           flow_stats = STLFlowStats(pg_id=0),
>>>>       )
>>>>
>>>>       streams.append(s)
>>>>
>>>>       s2 = STLStream(
>>>>           packet=STLPktBuilder(
>>>>               pkt=get_packet(2,"00:11:22:33:44:55", "10.1.0.2",512),
>>>>               # vm = vm,
>>>>           ),
>>>>           isg=0 * 1000000,
>>>>           mode=STLTXCont(pps=1.2*10**6),
>>>>           # flow_stats = STLFlowLatencyStats(pg_id = 0)
>>>>           flow_stats = STLFlowStats(pg_id=1),
>>>>       )
>>>>
>>>>       streams.append(s2)
>>>>
>>>>       c.add_streams(streams, ports=[tx_port])
>>>>
>>>>       c.clear_stats()
>>>>
>>>>       c.start(ports=[tx_port], duration=60, mult="25gbpsl1")
>>>>
>>>>       c.wait_on_traffic(ports=[tx_port, rx_port])
>>>>
>>>>       stats = c.get_stats()
>>>>       print(stats)
>>>>
>>>> if __name__ == "__main__":
>>>>       main()
>>>>
>>>>
>>>> And the configuration is
>>>> - port_limit: 2
>>>>     version: 2
>>>>     port_bandwidth_gb: 100
>>>>     interfaces: ["3b:00.2", "3b:00.3"]
>>>>     port_info:
>>>>       - dest_mac: 00:00:00:00:00:01
>>>>         src_mac: 00:01:00:00:00:01
>>>>       - dest_mac: 00:00:00:00:00:02
>>>>         src_mac: 00:01:00:00:00:02
>>>>     c: 14
>>>>     platform:
>>>>       master_thread_id: 8
>>>>       latency_thread_id: 27
>>>>       dual_if:
>>>>         - socket: 0
>>>>           threads:
>>>> [9,10,11,12,13,14,15,16,17,18,19,20,21,22,23,24,25,26]
>>>>
>>>>
>>>> BISDN GmbH
>>>> Körnerstraße 7-10
>>>> 10785 Berlin
>>>> Germany
>>>>
>>>> Phone: +49-30-6108-1-6100
>>>>
>>>> Managing Directors:
>>>> Dr.-Ing. Hagen Woesner, Andreas Köpsel
>>>>
>>>> Commercial register:
>>>> Amtsgericht Berlin-Charlottenburg HRB 141569 B VAT ID No:
>> DE283257294
>>>> ________________________________________
>> --
>> BISDN GmbH
>> Körnerstraße 7-10
>> 10785 Berlin
>> Germany
>>
>>
>> Phone:
>> +49-30-6108-1-6100
>>
>>
>> Managing Directors:
>> Dr.-Ing. Hagen Woesner, Andreas
>> Köpsel
>>
>>
>> Commercial register:
>> Amtsgericht Berlin-Charlottenburg HRB 141569 B VAT ID No: DE283257294

-- 
BISDN GmbH
Körnerstraße 7-10
10785 Berlin
Germany


Phone: 
+49-30-6108-1-6100


Managing Directors: 
Dr.-Ing. Hagen Woesner, Andreas 
Köpsel


Commercial register: 
Amtsgericht Berlin-Charlottenburg HRB 141569 
B
VAT ID No: DE283257294
Previous message (by thread): Trex with mlx5 driver - Multiple streams with different VLAN priority causes high CPU utilization
Next message (by thread): connection tracking in mlx5 not working or displaying strange results
Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]
More information about the users mailing list