Harald Welte writes: > On Mon, Sep 27, 2004 at 10:51:59AM -0600, Stephen J Smoogen wrote: > > > Is there more publishing on Rober Olsson's benchmarks and findings. We > > are doing various testing here and while the numbers look similar, I > > would like to get an established testing methodology so that we can > > confirm/help results from baseline tests. > > http://robur.slu.se/Linux/net-development/experiments/2004/ > http://robur.slu.se/Linux/net-development/experiments/10G.html > http://robur.slu.se/Linux/net-development/experiments/ Yeep I keep notes of some my test here. Not alwayes to well organized or documented I suffer from this myself. > > Our next line of testing is to find out which network cards will give us > > the best bang for the buck on 10 gig testing. We currently have 300kpps > > at 1 gig (E1000 + 2.4.2x kernel) but not sure if we are reaching the > > limitations of the E1000 fibre optics or the PCI-X bus. > 300kpps is bad, even for a 32bit/33MHz PCI bus, and definitely way too > bad for any kind of PCI-X I think Harald and I booth seen ~800 kpps pure forwarding with Opteron and e1000 w. UP-kernel and I've seen aggregated performance of 1.3 Mpps from dual Opteron in a dual flow experiment. Yes in the long run method/benchmark would be needed but simple tests just as adding fw-rules is enough to meditate about improvements. The lookup code is getting cleaned up now to get pluggable lookups. With faster lookup and w/o route hash at least some filters could be very fast but it probably break the currents semantics. Maybe replacing all lookup with general classifier bitvector>/hi-pac as discussed at the workshop could be possible when the times comes... Anyway have the LC-trie todo prefix lookup up now but not interfaced to the kernel yet. A lot of work happens in the hash chains which makes it hard to replace current route lookup. Cheers. --ro
