Rupert Gallagher [[email protected]] wrote: > > You seem to say that handling larger packets is a feature of having limited > CPU. I disagree. >
Rupert, I'm saying that a slower CPU can process less packets per second. The important measurement is packets-per-second. The APU has plenty of memory bandwidth to handle large volumes of data. For adequate CPU power, you have to either lower the cost of processing (make software better/more efficient) or you have to distribute the cost across the 4 cores of the APU2 (make software execution parallel). > > The same traffic level, with 1500 byte packets generates 6 times more > > packets per second than that traffic level with 9000 bytes packets. > > You divided 9000 by 1500 without mistakes. Congratulations. > The point was clearly lost on you. > > There is ongoing work to improve the network stack performance on boxes > > like the APU2 (which have 4 cores). You will see improvements. If you want > > it better today, you need a faster box. Chris > > The apu2c4 is fast enough to saturate its Intel 1Gbits/sec link. It has three > of those. If you connect all three to the switch, you get 3Gbps shy. No need > for a faster box. You rather need a faster switch, class 7 S-FTP wires > (better than class 6), and 2.5Gbps lan cards for clients. No, you don't need any of that. You have no idea what you are talking about. The APU requires software crafted to evenly distribute PER-PACKET PROCESSING cost across multiple cores. That is what is happening in OpenBSD today. It has been happening for years, and it is getting closer to becoming a reality with OpenBSD + APU2, as well as other chipsets/platforms. For a couple years now, we've had interrupts processed by one core, PF on another, and other parts of the kernel on a third core. But to accelerate packet processing alone, we need interrupts handled on multiple cores, PF processing handled on multiple cores. This is hard work. By the way, what I'm describing is the general-purpose OS approach towads this problem. If you want to turn computer hardware into routers with little other concern, the go-to platform is DPDK + VPP. It is something like an order of magnitude faster than any general purpose OS (OpenBSD, Linux) at packet pushing. https://www.reddit.com/r/networking/comments/6upchy/can_a_bsd_system_replicate_the_performance_of/dlvdq2e/ Chris
