On Mon, 27 Apr 2009, Yanping Du wrote: > On a single 10GE link, we tuned and got netperf 9.4Gb/s (with 1500 > MTU) and ~9.9Gb/s with jumbo frame.
82598 has only a PCIe GEN 1 (2.5 Gb/s per lane) > But when running netperf tcp traffic to the two 10GE links on the card > at the same time, we still got only ~10Gb/s of throughput on the two > links summed up. The 10GE card is installed on a PCIe gen-1 x8 slot so > we were expecting something close to bus speed of 2GB/s = 16Gbps. use UDP, not TCP, and that will get you closer to your max. Due to PCIe overheads, the bandwidth for data traffic will likely be closer to 12-14 Gb/s > What's the expected aggregated throughput on the two 10GE ports on the > same card? is this bulk receive mostly or transmit? > > Our testing environment: > - Linux 2.6.23 kernel; this is single tx queue only, correct? 2.6.27 will automatically use all available tx queues. > - 5~10 iperf/netperf threads generating TCP traffic from 2 individual > clients; Both direct connect and from a switch were tested. with that many threads and only a single tx queue you're going to be limited somewhat on transmit per port. > - 8 receive queues on 10GE server side; Interrupt affinity masks were > 0xff so interrupts are evenly load-balanced to multiple cpus. Observed > interrupt rate was not high and cpus had enough idle cycles. in this test I bet letting irqbalance do its thing will actually be more efficient as long as you let netperf run for 20-60 seconds, especially on the receive side. you don't have two machines hooked back to back do you? > Please let me know if more info is needed. what type of machine? what is your memory configuration and does it have enough bandwidth to do 10GbE x 2? UDP tests (or better yet pktgen.ko) are the only way to go to test pure hardware throughput. I suggest for udp, do netperf -H remote1 -C -c -t UDP_STREAM -- -m 1460 -S 128K -s 128K or better yet run pktgen (comes with 2.6.23 kernel) and I can email a script I use to configure it if you need it. ------------------------------------------------------------------------------ Register Now & Save for Velocity, the Web Performance & Operations Conference from O'Reilly Media. Velocity features a full day of expert-led, hands-on workshops and two days of sessions from industry leaders in dedicated Performance & Operations tracks. Use code vel09scf and Save an extra 15% before 5/3. http://p.sf.net/sfu/velocityconf _______________________________________________ E1000-devel mailing list E1000-devel@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/e1000-devel