netrunner wrote:Am I understanding correctly that Cake is an evolutionary "upgrade" to fq_codel (a la UBNT "Smart Queue") AND avoids the heavy performance penalty inherent with disabling HW acceleration? Be still my beating heart!
yes, to the first, or so we hoped. We poured most of our conception of typical options for "smart queuing" into C, and added tweaks and features everywhere.
as for the second, the performance answer is nuanced, and a little muddy. I don't have a grip on how all the offloads work in the erl series. Certainly I'd hoped that we'd peel GRO inbound from other interfaces, at least.
One of our hopes also with cake is that it would be faster than your typical htb + several classes of queuing setup, from a pure cpu perspective, so we could maybe scale it past the 60Mbit/sec limit htb + fq_codel had on earlier edgerouter products and wherever it is now on the current stuff.
However, we have not gone back and tuned cake for any architectures as yet, instead driving for the system that had the lowest possible network latency and met our goals for feature completeness and algorithmic correctness.
Funding for cake expired last year and we've been working mostly on the make-wifi-fast project instead of "finishing" cake. I am glad more people are trying it, it's been "good enough for me" - essentially 98% done - for quite a while now, and who can argue with a single command line to set it up?
Please let us know your experiences on the cake mailing list.