LF_DPDK17_Power Aware Packet Processing
- 1. Chris MacNamara, Dave Hunt DPDK Summit - San Jose – 2017 Power Aware Packet Processing #DPDKSummit
- 2. 2 Why We Are Here u Drive for data and always on networks u Opportunity for green DPDK u Based on continued polling and varying traffic rates u Achieve electricity cost saving & increase performance u Updates to the power management scheme in 17.11
- 3. 3 Mapping Power Usage To Network Traffic Intel RSNDA Example 24-hour period System%Busy Network Traffic
- 4. 4 Moving To Green DPDK u Evaluation considerations Latency Loss Energy Power State/Draw u Benefits Matching Platform Energy to Network Load
- 5. 5 DPDK View Of Energy Efficiency & Performance u Out of the box frequency management u Freq Up / Freq Down u Freq Min / Freq Max u Enable Turbo: Enable Intel® Turbo Boost Technology on the specific lcore u Core frequency will go to whatever frequency is allowed for that core based on number of active cores on the packet, thermal limits, etc. u Disable Turbo: Disable Intel® Turbo Boost Technology on the specific lcore u Core frequency will return to the maximum non- turbo frequency, if lower freq required, a further library call is required to scale down, go to minimum, etc. PowerConsumption L3Fwd Power L3Fwd Low (W) High (W) Time
- 6. 6 Meeting The Needs Of An On Demand Network u Scale always on DPDK performance with the network demand u Common challenges u Always On u Adjust PMD cores frequency to adjust to packet demand u Potential to save power drawn per core using frequency scaling u Additional savings from sleeping u Speed of (re)action u Challenge: Fast Scale Up to react to increases in n/w traffic u Time = queueing/buffering u Challenge: fast monitor & reaction time u Closer to hardware gives faster reaction time u Move to policy-based control (Re)Action Monitor Measure Apply Power Where and When it’s needed
- 7. 7 Elements Of An Ideal Scheme u A system with a penalty-free reconfiguration capability u DPDK: Be deliberate & control change, only change when you need to change u Fast detection at the micro-burst level u DPDK: Decide on key performance indicators (KPIs) u Mechanism to determine the compute configuration u DPDK: Use DPDK APIs to manage u Power/energy savings with minimum impact to network performance u DPDK: Toolbox allows fine grained control, network load
- 8. 8 In-band Policy Control For Power Mgmt u Patch Set for 17.11 u Power governor on host u Takes profiles from Guest u Scale up/down based on: u Traffic Rates u Time of Day u Workload (next) u Match compute to network/CPU load
- 9. 9 Acknowledgement u Nemanja Marjanovic u Rory Sexton u Konstantin Ananyev u John Browne