Some bandwidth and CPU power thoughts on gbe pure routing



  • Hi all!
    In the future I wish to have a mixed 10/1gbe fiber/copper network in my home using pfsense. The main reason behind is second hand switches there cost way too much, enought to make a porpusely bilt router somewhat interesting. Right now, I did some runs on gbe to see if even weak low power CPU can handle multi gigabit speeds. The (bare metal) system has a Athlon 5350 (very light overclock), Asus AM1M-A, 4GB DDR3L 12800, i350t4, 3 pcie rtk and a VIA pcie nic. The power usage is around 20w while idling. The system as all the usual feature plus suricata on wan, pfsense was a 2.3 release. LAN to LAN (pc with ssd to raid 10 nas all with Intel nics) was able to delivery data at wire speed even without jumbo frames. Pfsense CPU usage with powerD hiadaptive enable was well lower than 50% and the load was nowhere near 4.
    So my conclusions are:
    -a simmetrical gbe is very easly obtainable even with slow, cheap, not bleeding edge ipc and single channel cpus;
    -there is still a lot of headroom, it should be able to route at least two times as much data across the lan, even to reach 10gbit or more aggregate;
    -power consumption is still acceptable, a few watts more than idling, manly due to not enought efficient psu.
    More studies are needed. Of course, I discourage you to create a setup like this if you don't want to go beyond gibabit speed as fast 8 ports switches are way cheaper and easier to set up. I want to see what happen with 2 pcs to 2 nas, so 4 gigabits aggregate.



    • a router (pfsense) is not a switch. switches are much cheaper ($80) & perform much better
    • realtek nics are known for their bad behaviour
    • 2Gbit/s wirespeed requires serverclass hardware, not low power (at this point in time)