ok i have some routers running MT
i have some soekris 4501 that do a 30Mbit thruput
a 4801 does a 70mbit
i have a custom embedded router based on a via 700Mhz cpu
and that one does 95Mbit thruput
so now i made a new router with 2 broadcom Fiber gigabit adaptors.
with a Intel P4 2,4ghz cpu.
so i would think if my thruput is linked to cpu speed then this box must rock…
but for some reason the bandwith test just flatlines at 99Mbit.
and the strange thing is both boxes run at 100%
even stranger if i set the bandwith test to max 1Mbit, then even both boxes are running 100% cpu…
this isn’t normal…
i have a normal 24H test license on the boxes. and use the standard MT bandwith test…
we ran (i am posting this the 5th time already, i think) a test through two intel xeon devices with intel gigabit cards. we achieved 3Gbps total throughput, so there is some problem with your setup.
we installed the boxes with fedora and did a bandwith test.
and we got a 800Mbit average.
what i think is verry acceptable.
but with MT on it (with a smaller footprint and less big kernel)
it runs slower???
i think this is strange…
networkcard is a tigon3 broadcom based FX 1000 adaptor.
i did the same test with intel 10/100/1000 TX adaptors …
fedora is fine again…
MT no go beyond 104Mbit
Don’t test “To” the router.
Test “Thru” the router. The test routines take up too much CPU time.
Also the end points will need to be able to produce traffic greater than what you expect the router to forward.
IE you cant use PII boxes to produce traffic across a multi Gb Xeon..
Your “Generators” linitations would show and not the routers.
I run Gig (1000t and Fiber) and have no issues.
I ONLY run Intel NICs. (Compaq Mfg.) in 32/64 bit flavors.
I don’t think it’s only bandwidth that is maxing out cpu load, as often being talked about. the focus should be on pps! the packet forwarding rate is the key, that should be observed. intel gigabit nics do have the feature of “offloading” the main cpu on some tasks, but the pps rate only gets really boosted, if the underlying kernel supports NAPI. this technique “bundles” multiple ingress packets to one single interrupt, the more packets coming in the more it bundles, with the effect that the intr/s value drops down the higher load there is.
so this feature should be built into MT kernel to improve CPU utilization when there is high load on gigabit interfaces. especially (d)DOS attacks can be survived when the kernel doesn’t trigger a single interrupt per packet anymore.