Hi Cliff
Please see inline

Il giorno 28/apr/2013, alle ore 18:06, Cliff Burdick <[email protected]> ha 
scritto:

> Hi, we have a fairly simple requirement for capturing packets, and I'm trying 
> to figure out the most efficient way to do this. Our setup is as follows:
> 
> - Intel 82598, Xeon 6-core 2.2GHz (12 with hyperthreading), and 32GB of RAM
> - Traffic coming from up to 12 known IPs, that have a single thread for each 
> IP to do parse and analyze the packets
> - Aggregate of up to 2Gbps of traffic
> 
> Previously we were using libpcap with pf_ring in transparent mode 1, and it 
> seemed to be dropping packets. Alfredo suggested we switch to DNA clustering, 
> so we're in the process of doing that. However, I have a couple of questions:
> 
> 1) It appears that in the DNA load script it's suggested that RSS is 
> completely turned off in the intel driver so that the master cluster thread 
> can do something similar. Since RSS is done in hardware, won't this have a 
> performance impact?

No, specially at 2 gbit. However, if you need a simple IP-based balancing, hw 
RSS is also an option (but please note that you don t have full control on the 
hash function).

> 
> 2) The master would be a single thread whose only job is to pull a packet 
> from the one RX queue on the card, run through the custom hash function, and 
> forward the packet using zero-copy to one of the twelve threads. Will one 
> core be sufficient enough for a single thread to hash every incoming packet?

Yes

> 
> 3) In this paper:
> 
> http://luca.ntop.org/imc2010.pdf
> 
> it is recommended to have the "capture thread" and "polling thread" on a 
> single core so that they can share the same cache. In my case the master 
> cluster would be the "capture thread" (which there is only 1?), and the 
> "polling thread" will be one of the other 12 that capture/analyze data from 
> the unique IPs. Since we only have a 6-core processor (12 with HT), what is 
> the recommended way of distributing the 12 poller threads to minimize cache 
> coherency issues? Since the DNA master is using one part of the core, we 
> would have to share the rest of them somehow, so I wasn't sure if the answer 
> was to have multiple DNA cluster masters.

This paper describes the kernel-space cluster, those considerations do not 
apply to the Libzero DNA Cluster. 
For fine tuning cpu affinity have a look at the load on the cores with your 
traffic and processing applications.

Best Regards
Alfredo

> 
> Any suggestions would be appreciated. Thanks.
> _______________________________________________
> Ntop-misc mailing list
> [email protected]
> http://listgateway.unipi.it/mailman/listinfo/ntop-misc
_______________________________________________
Ntop-misc mailing list
[email protected]
http://listgateway.unipi.it/mailman/listinfo/ntop-misc

Reply via email to