CPU reaching 100%

[gaming master blaster]

Dear Team,

Greetings.

I have created bond by combining 4 NICs each with 10gig speed so having total speed as 40gbps. when I try to run test even with 9 gbps then observing TREX (I m using stateless) process is going 100% and not utilizing all CPUs. Is there any config to tell TREX process to say pls use all CPUs. If I run same test with 3gbps then still it reaches 76% so not distributing its load to all available CPUs.

Test Details :-

1) Started trex server so which started scrappy as well.

2) Executed trex>start -f stl/udp_range_client_to_server.py -m 9gbps

3) Got CPU 100% for CPU number 2 and rest all are sitting idle.

Thanks

Vishal

[Besart Dollma]

How did you start TRex? Did you use all the cores with -c flag?

[hanoh haim]

It is expected. Use “low_end” for traditional Linux behavior 

--
You received this message because you are subscribed to the Google Groups "TRex Traffic Generator" group.
To unsubscribe from this group and stop receiving emails from it, send an email to trex-tgn+u...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/trex-tgn/85edce0d-1d76-4d7d-9787-296b0d94374bn%40googlegroups.com.

--

Hanoh

Sent from my iPhone

[gaming master blaster]

Hello,

I have full blown baremetal server with huge RAM (364 GB) , 48 CPUs, 4 Mallanox cards with speed of 10gbps of each so created bond having total capacity as 40gbps so even with such a huge infra do I need to enable "low_end" because my thought was "low_end" requires when trex running in limited infrastructure.

I have executed test in stateless mode and also tried "-c 4" options but getting error that " STL mode does not support more than one core by dual interfaces" so is it limitation of trex for running traffic in Stateless mode? 

/opt/trex/v2.89# ./t-rex-64 -i -c 4 -v 7 -p  (This is how running server)  

Thanks!

Vishal

To view this discussion on the web visit https://groups.google.com/d/msgid/trex-tgn/CA%2BYxBo%2BDjEgkEpSnhBZRDVXTVqGKyLoeULVjtekqHNkHBfuo%3Dg%40mail.gmail.com.

[gaming master blaster]

Hi Team,

I understand you are having many queries from many people but please get back to me about my query as well.

Thanks!

Vishal

[hanoh haim]

Hi Vishal,

TRex works with DPDK and it uses user-space polling driver that consumes 100% CPU from the kernel perspective, even when you send 1% of the traffic. The delay is not a kernel delay it is spinning in idle mode.

It is a feature not a bug. 

Thanks

Hanoh

[gaming master blaster]

Hi Hanoh,

Many many thanks for the reply.

So what is the workaround which I can use in TREX to utilize all the CPUs and to avoid error "STL mode does not support more than one core by dual interfaces". I have 48 cores so obviously cant create those many interfaces because having one logical bond (bond0.101) to send traffic to DUT and another logical bond (bond0.102) to receive traffic back from DUT.

Thanks!

Vishal

[hanoh haim]

Hi Vishal, 

I was out of context here and forget my answer, please send the trex_cfg.yaml, the command from the console and the TUI output 

Thanks

Hanoh

[Emmanuel García]

Hi Hanoh,

I'm also trying to use a logical bond linux interface and the test runs well, however traffic is going through just one of the two available members of the bonded interface.

Pings from linux (to same destinations used in the test) go through both members, but when I run the test, traffic sticks to one member.

Am I missing something?

Regards,