Internet infrastructures have a large environmental impact. We need more and more machines, implying more manufacturing, more shipping, more powering and more cooling. With DPDK, every CPU cycle matters in order to achieve a high throughput. It helps to reduce the number of required machines for maximum load. Using machine capabilities at best is a case of power efficiency. There in another case to consider regarding our energy footprint. When the machines are not under heavy load, we can reduce powering and cooling. This case requires power management to be well integrated with DPDK systems. Last but not the least, the power efficiency must be considered for every hardwares used in datacenters and telecom equipments.
I think reducing workload on cores would help having a better power management. Having a unified interrupt support is the first step for a better adoption from applications. VFIO / igb_uio / others do not aggregate / demux interrupts in the same way, meaning that for a single application to have a NAPI-like behavior requires significant device-specific code, dependent upon the current IO driver used. Instead of proposing only rxq-interrupt, a possible evolution could be to propose NAPI-like bursts, that would abstract away the rx-interrupt handling and allow all PMDs to behave similarly.
Is it a bug? looks like the open discussion.
It is the bug #1. When it will be fixed, we can all take big vacations. Please let's keep it open to remind us this major goal.