Research On Graphic Processing Units-Based Energy-Efficient High-Performance Packet Cassification

Packet classification is a key technique for filtering network packets in many network infrastructures,such as Routers,Firewall,SDN switch and so on.Packet classification re-quires matching packet headers against a series of pre-defined rules,and the rules determine which action is taken for each packet.The performance of the matching process determines,to some extent,how fast packets can be processed.In recent years,new software and hardware techniques have brought about a trust-worthy packet classification with a performance that is high enough,using off-the-shelf hardware.However,most of the focus has been on improving performance,few of them take power consumption into account.Actually,in the large scale of data center,power consumption will become a constraint.Therefore,the trade-off between high-performance and energy-efficiency is worth discussing.In this thesis we focused on the new software and hardware platform that can sig-nificantly accelerate rule matching process,studied on reducing power consumption while keeping a high throughput.The main contributions are outlined as follows.1.Toward Energy-Efficiency Optimization of Pktgen-DPDK Based on CPU AffunityWe start from a widely used high performance schema and make a deep study on its use on the multi-core platform,especially in terms of parallelism,core allocation and frequency controlling.On this basis,we then propose an AFfinity-oriented Fine-grained CONtrolling(AFFCON)mechanism to improve energy efficiency with desirable performance as well.As clearly demonstrated through a series evaluation experiments,our proposal can reduce CPU power consumption by up-to 11%,while keeping throughput maintains the line rate.2.Packet Classification using Community DetectionPacket classification suffers from a degradation of performance when man-made rules contain some overlap,useless,or redundancy rules.When we implement a packet classi-fication system in a real network,we find that the rules have the characteristics of a social community.On the basis of community detection,some rules can be clustered by similarity and share a common action.Therefore,the rules that affect performance can be optimized,which will be beneficial for matching time,memory usage and rule updating.We present a COMmunity detection CUTtings(ComCuts)algorithm for packet classification based on a counting bloom filter,and a rule similarity algorithm for clustering.Experimental results show that our algorithm reduces matching time by 8%and decreases memory usage by 50%compared to a HiCuts algorithm.Furthermore,our clustering scheme uses elasticity scope to adopt to a frequently updated system,especially in a SDN network.3.Energy-Efficient Fuzzy Control Model for GPU-Accelerated Packet ClassificationIn recent years,the Graphic Processing Unit(GPU)is adopted to accelerate packet classification.However,the high performance of GPU comes at the cost of high power.Then,inspired by the frequency-variable energy-consuming model for air conditioners,a Fuzzy Control-based Energy Efficiency Optimizing(FCEEO)model is proposed for GPU-accelerated packet classification.As demonstrated in the evaluation experiments,when the GPU is in the idle status,the proposed model can save 10 W.In running status,the FCEEO model can avoid GPU shutdown issue caused by GPU self-protection mechanism when the GPU temperature rises to 95~?C.Furthermore,by improving the resource configuration of GPU kernels according to the model,the overall energy-efficiency is enhanced by up to15.5%,while simultaneously keeping throughput at the same level.4.Packet Classification Algorithm Based on CPU-GPU Heterogeneous PlatformAs an accelerator platform,GPU needs to cooperate with the CPU.If an algorithm does not take into account the characteristics of the CPU-GPU heterogeneous platform,it will not be able to maximize the hardware performance.In this paper,an energy-efficient algorithm based on heterogeneous platforms is proposed.Based on CPU affinity and GPU zero copy optimization,not only can improve the throughput performance,but also reduce the overall power consumption of the system.