Does your NIC connect directly to your Socket? If not, the packet might go through QPI, which will cause additional latency. Check your motherboard.
Wang, Shawn On 1/22/14, 6:52 AM, "Dmitry Vyal" <dmitryvyal at gmail.com> wrote: >Hello MIchael, > >I suggest you to check average burst sizes on receive queues. Looks like >I stumbled upon a similar issue several times. If you are calling >rte_eth_rx_burst too frequently, NIC begins losing packets no matter how >many CPU horse power you have (more you have, more it loses, actually). >In my case this situation occured when average burst size is less than >20 packets or so. I'm not sure what's the reason for this behavior, but >I observed it on several applications on Intel 82599 10Gb cards. > >Regards, Dmitry > > >On 01/09/2014 11:28 PM, Michael Quicquaro wrote: >> Hello, >> My hardware is a Dell PowerEdge R820: >> 4x Intel Xeon E5-4620 2.20GHz 8 core >> 16GB RDIMM 1333 MHz Dual Rank, x4 - Quantity 16 >> Intel X520 DP 10Gb DA/SFP+ >> >> So in summary 32 cores @ 2.20GHz and 256GB RAM >> >> ... plenty of horsepower. >> >> I've reserved 16 1GB Hugepages >> >> I am configuring only one interface and using testpmd in rx_only mode to >> first see if I can receive at line rate. >> >> I am generating traffic on a different system which is running the >>netmap >> pkt-gen program - generating 64 byte packets at close to line rate. >> >> I am only able to receive approx. 75% of line rate and I see the >>Rx-errors >> in the port stats going up proportionally. >> I have verified that all receive queues are being used, but strangely >> enough, it doesn't matter how many queues more than 2 that I use, the >> throughput is the same. I have verified with 'mpstat -P ALL' that all >> specified cores are used. The utilization of each core is only roughly >>25%. >> >> Here is my command line: >> testpmd -c 0xffffffff -n 4 -- --nb-ports=1 --coremask=0xfffffffe >> --nb-cores=8 --rxd=2048 --txd=2048 --mbcache=512 --burst=512 --rxq=8 >> --txq=8 --interactive >> >> What can I do to trace down this problem? It seems very similar to a >> thread on this list back in May titled "Best example for showing >> throughput?" where no resolution was ever mentioned in the thread. >> >> Thanks for any help. >> - Michael >
