One lock-free queue is just simpler to work with than multiple queues. And the one queue nicely decouples the producers and the consumers, where they don't need to know each other's load and scheduling.
This implements all the work units consumes the same load for the consumers. Otherwise, the producer needs more logic to figure out which queue to enqueue. Accessing the consumers' statistics again becomes the problem of accessing shared data.
> multi-producer becomes trivial (atomic doorbell + round robin)
This implements all the work units consumes the same load for the consumers. Otherwise, the producer needs more logic to figure out which queue to enqueue. Accessing the consumers' statistics again becomes the problem of accessing shared data.