ORIGAMI

To address the unsustainability issues highlighted in Section ‎4.1, ORIGAMI proposes two strategies.
1.      Opportunistic HA offloading. As hinted by the experimental results shown in Figure 11, CPUs alone may handle some 5G PHY workloads in Distributed Units (DUs) without the assistance of Hardware Accelerators (HAs) by exploiting SIMD programming and other optimisations. Additional information can be found in ‎[29] ‎[30]. However, CPUs alone cannot ensure 5-nines reliability for all workloads, as shown in Figure 11, and, consequently, they are usually shunned for this job in industry-grade RANs. Instead, ORIGAMI will show that CPUs can be a valuable complement to HAs in these tasks and that balancing DU workloads between CPUs and HAs can substantially improve the cost- and energy-efficiency of vRANs. The rationale is that minimising processing latency brings no benefit as long as DU processing deadlines are met, hence CPUs may be occasionally exploited to alleviate the HAs’ energy toll. As an example, Figure 11 (left) shows that a CPU core can decode within 1 ms TBs below 100 Kb (which correspond to a large portion of today’s real-world TBs, as reported in the literature) consuming ∼5.7× less energy than a GPU-based HA (right plot).
2.      Processor Pooling. HAs co-located with (and thus exclusively used by) individual DUs suffer from low usage under real workloads. ORIGAMI will seize this opportunity to share HAs among multiple DUs, so as to amortize the cost of these expensive resources, and provide the needed acceleration at an affordable cost per DU. The concept of RAN pooling is not new, though. Indeed, 71% of US operators intend to realise RAN pooling solutions by 2025 ‎[31], and some already implement it ‎[32], but the traditional RAN centralisation approaches only exploit long-term traffic variations, such as day-night ones, which is insufficient for cost-efficient RAN virtualisation.