I think that adding a biasing factor to low priority work would be nearly as effective as complex queuing models with just about as much performance overhead. A simple implementation would be causing the producer to sleep for a few milliseconds before submitting low priority requests to the consumer.