[Cake] Cake fix: diffserv4 priority quanta

[Cake] Cake fix: diffserv4 priority quanta

[Jonathan Morton]

My preferred cake fix is Death By Chocolate - and it’s been quite a while since I’ve had any.  But this will have to do for now.

While drafting a paper to go with the code, I had to go back and think about the rationale for quite a few things, and this revealed that the quanta used for the priority layer in diffserv4 mode were incorrect.  There are two sets of quanta here:

- The one used when the allocated bandwidth for the class has not been exceeded, which sets a priority balance.  The ratios here need to be at least sufficient to allow any higher-priority class to force any lower-priority class out of the higher class's allocation; in practice it’s only necessary to consider adjacent pairs for this purpose.

- The one used when the allocated bandwidth for the class *has* been exceeded, which sets a bandwidth balance.  The ratios here should correspond to the effective allocations when *all* classes are fully saturated; transitions between the bandwidth and priority balance states take care of any other case.  The effective saturated allocation is the class’s own allocation *minus* the allocation of the class above it.

The basic allocations of the four classes in diffserv4 are:

Background:  100%
Best Effort: 15/16
Video:        3/4
Voice:        1/4

The effective saturated allocations are therefore:

Background:   1/16
Best Effort:  3/16
Video:        1/2
Voice:        1/4

It is now obvious that the quantum ratio between Best Effort and Background needs to be at least 16:1 in the priority balance, and should be 3:1 in the bandwidth balance.  Ratios of 4:1 for the elevated-priority classes in priority balance are sufficient.

The quanta for the diffserv8 and precedence modes are already sane (although the bandwidth-balanced quantum for the highest class turns out low, but this doesn't matter in practice; exercise for the reader is to work out why).  It helped, there, that the bandwidth allocations proceed geometrically (7:8 ratio in each step up) and are thus easier to calculate intuitively.

I’ve just pushed a corresponding patch to the net-next derived repo.  I don’t have push access to the out-of-tree version.

 - Jonathan Morton