In my blog last month, Internet Bent not Broken, I discussed how the Internet uses a “best effort” weighted fair queuing mechanism for the majority of traffic. The advantages of WFQ are that it is simple to implement and, as its name implies, it is “fair.” While WFQ prioritizes traffic queues, for example based on their length, it does not distinguish between individual frames of packets. It treats them all equally.
However, in our human world we are used to extending differential treatment. Say for people with special needs, or for individuals willing to pay more. As it turns out, the Internet can make similar accommodations with the addition of priority forwarding mechanisms. It relies on traffic being “tagged” into various classes, and a “classifier” to examine these tags and direct the frames into the appropriate forwarding queue.
The aforementioned best effort forwarding using WFQ is overprovisioned by design to a predetermined contention level, so that frame drops and fluctuations of available bandwidth are expected. This is not problematic so long as the bulk of traffic transported is best effort; and virtually all Internet browsing and video streaming traffic falls into this class.
Assured forwarding is like a silver level pass that provides an intermediate level of priority. It supports services with a guaranteed bandwidth, but only at a level that is moderately better than could be achieved using best effort forwarding. AF frames are given regular but not absolute priority over BE frames.
Expedited forwarding is like a VIP pass that jumps you to the front of the line. It supports services like voice and interactive video that function best with low latency delays and low jitter (a smooth, continuous signal). Frames directed by the classifier into the EF buffer queue are almost always selected with strict priority over other traffic classes.
Many Internet engineers would consider this approach of adding EF/AF prioritization on top of BE as good enough to deliver a range of services with QoS guarantees. However, this approach still has several problems. There is no full isolation between flows. EF traffic would have to wait an unspecified time until the current frame is delivered (which can be up to 1500 bytes), which leads to jitter, even on the EF traffic. In addition, strict priority behavior can easily lead to starvation for the non-EF traffic, even if the total traffic is less than 100% of link capacity. This severely limits the portion of EF traffic to about 5% of the total traffic.
These are not minor concerns as we start moving into a 5G services world where there is a growing expectation for users to obtain hard service guarantees. An often-cited example is remote surgery, where bandwidth, latency, and jitter performance guarantees are all a must. While this may be an extreme example, there is no doubt that many businesses and gamers would certainly be willing to pay more for added levels of Internet performance guarantees.
Possible role of TDM
In response, some voices are beginning to advocate for adding deterministic time-division multiplexing (TDM) mechanisms to the packet queuing mix. This could look something like the diagram below. This adds a TDM scheduler in series to the packet queues frame selector, and dictates that all traffic is mapped into fixed time slots. Most time slots would be reserved for traffic coming from the BE/AF/EF queues, and the remainder dedicated to specific services coming in on the TDM stream with hard QoS guarantees.
One downside of this approach is that it reduces the overall efficiency of the egress link, because in some cases if there is no traffic coming in on the TDM stream, then the TDM scheduler is processing empty slots. Of course, there is always the possibility of adding a feedback link from the TDM stream to the frame selector (not shown in the diagram) indicating when it could populate unused slots. This is sort of like moving airline passengers from an overbooked economy class into empty seats in business class.
In conclusion, while “best effort” works fine for most Internet traffic, there are more sophisticated priority queuing and even deterministic scheduling mechanisms available to support a diverse services world with expectations for performance guarantees, for at least some of the services some of the time. We expect that the market will use the entire palette of available mechanisms to find the best balance between investment in increasingly complex solutions, and profits based on maximizing services revenues.
Jonathan Homa, is Senior Director Portfolio Marketing at ECI for the last few years. Jonathan enjoys French beer in Nice and loves to discuss the 3 hottest topics in networking today: softwarization, virtualization and cloudification.