Hi Alfredo,
Thanks for the info !
A few follow up questions, if I may:
1. Out of curiosity, if interrupts aren't a part of the game in ZC, how are packets dequeued from the shared memory ? on-going polling / NAPI (hope I'm using the right technical term) ?
2. Thanks for the reference ! What should I do in case I'm using a 3rd party IDS app, whose source code isn't available ? should I use taskset to pin the process to a specific core ?
Your help is much appreciated !
On Nov 3, 2016 19:39, Alfredo Cardigliano wrote:On 3 Nov 2016, at 18:07, Moshe Danielli <[email protected]> wrote:Hi Alfredo,1. There's no point in distributing the interrupts as the user threads read directly from the NIC's memory ?Yes, we avoid using poll/interrupts, and the application is directly accessing NIC’s memory without going through the kernel with ZC, thus you should not worry much about interrupts.2. Can you give an example / reference to bounding the threads to different cpu cores ?Sure, take a look at https://github.com/ntop/PF_RING/blob/dev/userland/examples_zc/zbalance.c-r and -g options, or search for bind2core() inside the code.Alfredo_______________________________________________Thanks again !
On Nov 3, 2016 17:54, Alfredo Cardigliano wrote:Hi Mosheplease note you can just ignore interrupts with ZC, just make sure you application threads are bound to the correct CPU cores in order to correctly spread the load.AlfredoOn 2 Nov 2016, at 15:38, Moshe Danielli <[email protected]> wrote:_______________________________________________I understand... However, can we control the irq / core distribution of flows as we can with RSS (using smp_affinity)?
On Nov 2, 2016 15:56, Alfredo Cardigliano wrote:Hi MosheRSS is an IP-based (IP+port for TCP) hw hash, if you need something more flexibly (e.g. a custom hash function) then you need zbalance_ipc.AlfredoOn 2 Nov 2016, at 14:06, Moshe Danielli <[email protected]> wrote:_______________________________________________Hi Alfredo,1. You were abosutely right - that was the exact reason. Once I switched to single queue mode, everything worked as expected.2. Given Intel's RSS functionality - what is zbalance_irq doing different ? Why should I use it instead of making use of RSS ?Thanks !
On Nov 2, 2016 12:03, Alfredo Cardigliano wrote:Hi Moshedo you have RSS enabled perhaps?Please read the documents below:AlfredoOn 2 Nov 2016, at 11:00, Moshe Danielli <[email protected]> wrote:_______________________________________________ Hi Everyone,I'm experiencing a weird situation where I'm running zcount twice:1. zcount -i zc:em12. zcount -i em1the first execution (zc mode) produces a count value of about 0.5 gbps while the 2nd execution (non zc mode) produces a value of about 4.5 gbps.How is that possible ? what am I doing wrong ?i'm using version 6.4.1-stable with 4096 ring slots and an intel 10g device (intel ixgbe 82599) on ubuntu 14.04.Thanks in advance
Ntop-misc mailing list
[email protected]
http://listgateway.unipi.it/mailman/listinfo/ntop-misc_______________________________________________
Ntop-misc mailing list
Ntop-[email protected]
http://listgateway.unipi.it/mailman/listinfo/ntop-misc
Ntop-misc mailing list
[email protected]
http://listgateway.unipi.it/mailman/listinfo/ntop-misc_______________________________________________
Ntop-misc mailing list
Ntop-[email protected]
http://listgateway.unipi.it/mailman/listinfo/ntop-misc
Ntop-misc mailing list
[email protected]
http://listgateway.unipi.it/mailman/listinfo/ntop-misc_______________________________________________
Ntop-misc mailing list
Ntop-[email protected]
http://listgateway.unipi.it/mailman/listinfo/ntop-misc
Ntop-misc mailing list
[email protected]
http://listgateway.unipi.it/mailman/listinfo/ntop-misc_______________________________________________
Ntop-misc mailing list
Ntop-[email protected]
http://listgateway.unipi.it/mailman/listinfo/ntop-misc
_______________________________________________ Ntop-misc mailing list [email protected] http://listgateway.unipi.it/mailman/listinfo/ntop-misc
