Hi Luca,
I am using better test pcap now and there are no high spikes in the traffic
the load balancing sends to one particular queue, now I can pass through
much higher data rates without packet drop on the NIC. Thanks for the offer
of sending it through your FPGA NIC. By the way, up to what
Hi Luca,
I have done some further testing using pktgen traffic generator, and I can
reproduce pfcount_multichannel reporting 6M pps at 64byte, 0% loss and the
NIC driver reports 0% loss ( rx_missed_errors from ethtool -S or drop
from /proc/net/dev), but again all these packets are the *same size*
Jonathan
I can inject traffic to the 82599 and send you the results. Please mail me a
pcap file (or better put it somewhere) so I can inject it using a FPGA-based
card and send you the results.
Luca
On May 24, 2011, at 12:35 PM, Lynch, Jonathan wrote:
Hi Luca,
I have done some further
As can be seen below - core 0-5 and 12-17 are on cpu 0 and core 6-11 and
18-23 are on cpu 1.
= Placement on packages =
Package Id. Core Id.Processors
0 0,1,2,8,9,10(0,12)(1,13)(2,14)(3,15)(4,16)(5,17)
1 0,1,2,8,9,10
Jonathan
we use the latest ubuntu server with no tuning, using a single Xeon processor.
I have never used a dual CPU machine on my tests, but I suppose that you need
to balance traffic properly. In some cases the cores are interleaved (core 0
goes to CPU 0, core 1 to CPU 1, core 2 to CPU 0...)
Hi Luca,
With regard to the ntop blog entry Packet Capture Performance at 10 Gbit:
PF_RING vs TNAPI, May 10, 2011
Do you have a packet drop % at the NIC driver level for each of the
benchmarks? (from rx_missed_errors in ethtool -S or drop from /proc/net/dev)
Also im just following up on the
Jonathan
I have forgot to say that the CPU was loaded ~50% when capturing at over 11
Mpps. This is also important if you plan to do something with these packets
beside just counting them.
Luca
On May 16, 2011, at 6:45 PM, Luca Deri wrote:
Hi Jonathan
On May 16, 2011, at 3:23 PM, Lynch,
Hi Luca,
Based on your blog entry, do you know up to what rates the TNAPI driver can
perform lossless capture...ie 0% packet loss?
Sounds like the DNA driver is progressing nicely then. Thats nearly wire
rate for 64 byte packets
Regards
Jonathan
On 16 May 2011 17:49, Luca Deri d...@ntop.org
Jonathan
the tests show that with pcount (but I assume that you want to do some extra
processing) you start to loose traffic over 6 Mpps, that is ~almost wire rate
for 128 bytes packets (and of course for larger packets)
Luca
On May 16, 2011, at 7:29 PM, Lynch, Jonathan wrote:
Hi Luca,