As I've(we've) struggled to put down the fundamental concepts about fq_codel in particular in the upcoming lwn piece...
It seemed like a good idea to try to explain codel, and fq_codel, in more detail, with some new methods. So I'm giving a talk about it at Stanford, on thursday - abstract and details as to the talk are below...
Attendees can be networking PhD students, professors, and engineers from the industry.
Look forward to perhaps meeting some of you.
This talk is a successor to the talk I'd given in modena a few month's back (youtube here:
http://www.youtube.com/watch?v=bi-jumVNVGk
)
While *this* talk will be filmed, I'm not sure if they'll let me bring the gatorade into the Gates building.
Kathleen Nichols and Van Jacobson spent 14 years developing a successor to RED, called "Codel". Even though well described in journals such as ACM queue, and with commonly available source code, how the tightly intertwined drop strategy and RTT estimator actually work is not well understood. Furthermore the combination of various forms of Fair Queuing on top of any AQM is not well understood.
This talk goes into detail of the current ns2 and Linux based code, while discussing the design decisions and network traffic types we looked at while designing them, and their predecessor (SFQRED) as well as touring through a large selection of graphs and data from various real-world simulations. It will also look into the known flaws of both codel and fq_codel, and describe the ongoing research into
improvements.