Revealing his observations through a mailer, Willy Tarreau, wrote that the performance issues that he has observed may very well be coming from Linux network stack and that these issues may be present in Linux 3.8 rc-2 as well but, he hasn’t seen similar reports.
Tarreau has explained in detail about the technicalities of the issue through this mailing list thread. His setup included three systems having 10GE network interface cards and through his tests he observed that the transfer speed wasn’t going beyond 6.8 Gbps even after he tried tweaking the “IRQ rate, CPU affinity, tcp_rmem/tcp_wmem, disabling flow control, etc.”
According to the Linux developer, some of the reasons behind such a behavior include possibilities of out-of-order issues with the Linux TCP stack because of interrupts being spread across multiple CPUs. Tarreau also believes that because he didn’t receive as many packets in the tcpdump, there could be some issue with the Linux memory allocation as well.
“Note that I’ve been using 3.5.7 as a fallback, but the same lab has run 3.6.6 without any issues 2 months ago, so I’m really confident that this is a 3.7 regression”, notes Willy in the mailing list thread.