G-NET: Effective GPU Sharing In NFV Systems

Transcription

1 G-NET: Effective Sharing In NFV Systems Kai Zhang*, Bingsheng He^, Jiayu Hu #, Zeke Wang^, Bei Hua #, Jiayi Meng #, Lishan Yang # *Fudan University ^National University of Singapore #University of Science and Technology of China

2 Network Function Virtualization (NFV) Network Functions: nodes on the data path between a source host and a destination host Firewall, NIDS, IPS, Gateway, VPNs, Load Balancers, etc. NFV is a network architecture concept: hardware => software Based on virtualization techniques Easier to manage/deploy, higher flexibility, higher scalability, easier to validate, etc. Construct service chains to provide specific services to meet different demands Firewall NIDS IPsec Router Virtualization Virtualization IPS LB Virtualization

3 s in Accelerating Network Functions s are proven to be a good candidate for accelerating network functions Router - PacketShader [Sigcomm 10] SSL reverse proxy - SSLShader [NSDI 11] IPsec Router Virtualization NIDS - Kargus [CCS 12], MIDeA [CCS 11] NDN Router - [NSDI 13] CPU Intel Xeon E v4: 18 Cores 76.8 GB/s memory bandwidth price: $2702 Nvidia Titan X: 3840 Cores 550 GB/s memory bandwidth price: $ Massive Processing Units Network functions large number of packets thousands of cores for parallel processing 2. Massively Hiding Memory Access Latency Network functions large number of memory accesses in processing packets s can effectively hide memory access latency with massive hardware threads and zero-overhead thread scheduling (a hardware support)

4 -Accelerated Network Functions RX Pre- Processing Processing Post- Processing TX

5 -Accelerated Network Functions RX Packet parsing, batching, etc. Pre- Processing Compute/memory- intensive tasks TX Post- Processing Construct/filter packet, etc. Parallel Processing in s 5

6 Why s Have not Been Utilized in NFV Systems? Current -based NFs - Exclusive Access to The is only accessed by one network function NIDS Virtualization Aho-Corasick algorithm

7 Why s Have not Been Utilized in NFV Systems? Temporal Sharing - Only kernels from one VM can run on the at a time Inefficient Firewall NIDS IPsec Router Virtualization Bit vector linear search Aho-Corasick algorithm AES and SHA1 DIR-24-8-BASIC

8 Current Way of Virtualization is Inefficient Temporal Sharing - Only kernels from one VM can run on the at a time underutilization Input: 20 Mpps Firewall NIDS IPsec Router Virtualization Aho-Corasick algorithm capability: 70 Mpps Idle

9 Current Way of Virtualization is Inefficient Temporal Sharing - Only kernels from one VM can run on the at a time underutilization Higher latency (1) Exclusive access A A A A A Kernel Timeline (2) Temporal sharing A B C A Kernel Timeline

10 Spatial Sharing (1) Exclusive access A A A A A Kernel Timeline (2) Temporal sharing A B C A Kernel Timeline A A A A (3) Spatial sharing B B B B B B Kernel C C C C Timeline Spatial sharing multiple kernels run on the simultaneously Minimize the interference of kernel executions from other NFs (Latency) Enhance utilization - Kernels from VMs can run on the simultaneously (Throughput)

11 Hyper-Q for Spatial Sharing Hyper-Q for spatial sharing A technique that enables kernels from the same context to execute on the simultaneously Firewall NIDS IPsec Router Virtualization Hyper-Q Challenges 1. VMs have different context => Cannot utilize Hyper-Q directly 2. Kernels utilizing Hyper-Q can access the entire memory space => Security issue 3. NFs are not aware of the existence of other NFs; An NF tries to maximize its resources would influence other NFs => Demanding scheduling and resource allocation schemes

12 The Goal of G-NET Easy to manage /deploy Flexibility Scalability Network Function Virtualization Easy validation Agility Spatial Sharing Security Development Scheduling Resource Allocation G-NET: NF-Hypervisor Co-Design

13 G-NET: Virtualization NF1 NFn Use API remoting to launch operations Framework Framework VMs A proxy creates a common context in the hypervisor for spatial sharing Manager Common Context Receive requests Perform ops Send response Hypervisor

14 G-NET: System Workflow NF1 Zero-copy principle applied in system implementation NFn Framework Req. Framework VMs Resp. Switch Shared Memory Manager Common Context Hypervisor NIC

15 Achieve Predictable Performance How to control the performance of an NF? Batch Size How to guarantee the performance of a service chain? NF Perf. Throughput Latency Kernel Performance Quantity of Work Compute Resource? Firewall NIDS IPsec Router Virtualization How to allocate resources? s utilize fast thread switching to enhance hardware utilization s have massive hardware threads (#thread >> #core) Unlike CPUs, a thread is unable to be bond to a specific core

16 Achieve Predictable Performance How to control the performance of an NF? Batch Size How to guarantee the performance of a service chain? NF Perf. Throughput Latency Kernel Performance Quantity of Work Compute Resource? Firewall NIDS IPsec Router Virtualization Thread block Our Approach Streaming Multiprocessor (SM) as the basic unit for resource allocation One thread block can only run on one SM; A thread block would be scheduled to run on an idle SM when there are available ones A thread block is allocated with one SM when Total #thread blocks <= Total #SMs Core SM

17 Service Chain Based Scheduling How to optimize the performance of a service chain with limited compute resources NFs have different processing tasks Firewall NIDS IPsec Router #SM Throughput (Mpps) Fairness Service chain based scheduling and resource allocation Locate the bottleneck NF (the NF with the lowest throughput T) Allocate resources for all NFs in the service chain to achieve throughput T * (1+P) (0<P<1) Firewall NIDS IPsec Router SMs Throughput improves by P in each round

18 Service Chain Based Scheduling NF1 NF2 NFn B1 B2 Bn Framework Framework Framework VMs Switch Scheduler Manager Traffic Speed 1. Batch size 2. #Thread blk Common Context Hypervisor NIC Streaming Multiprocessor NF1:5 NF2:4 NF3:7

19 IsoPointer for Memory Isolation NF1 NF2 NFn *Pa *Pb Framework Framework Framework *P > Memory Region Base Address (B) Memory Size (S) Pointer access checking: B <= P < B+S Switch Scheduler Manager Common Context NIC Ma Mb IsoPointer: guarantee memory isolation

20 NF Development NF1 NF2 NFn Repetitive development efforts CPU- pipeline Framework Framework Framework Manage CPU threads Communicate with Manager Packet I/O with Switch Switch Scheduler Manager Common Context Framework handles all common operations NIC

21 NF Development NF1 NF2 NFn NF Spec. NF Spec. Abstraction Abstraction NF Spec. Abstraction Framework Framework Framework Repetitive development efforts CPU- pipeline Manage CPU threads Communicate with Manager Packet I/O with Switch Switch Scheduler Manager Common Context Framework handles all common operations Abstraction to simplify NF development NIC

22 NF Development NF1 Pre- Processing Processing Post- Processing NF Spec. Abstraction Framework memcpy_htod pre_pkt_handler set_kernel_args post_pkt_handler memcpy_dtoh called for each pkt called for each kernel called for each pkt CPU code (Router) Implementation = Significantly reduces development efforts + Kernel Core functions

23 Evaluation Hardware CPU: Intel Xeon E v4 : NVIDIA GTX TITAN X NIC: Intel XL710 40Gbps Software Virtualization: Docker ce NIC Driver & Library: DPDK OS: CentOS , Linux kernel Service Chains 2 NFs: IPsec NIDS 3 NFs: { Firewall IPsec NIDS IPsec NIDS Router 4 NFs: Firewall IPsec NIDS Router

24 Throughput Comparison with Temporal Sharing Normalized Throughput up to 23.8% up to 25.9% Temporal Share G-NET Packet Size (Byte) (a) IPSec+NIDS Packet Size (Byte) (b) Firewall+IPSec+NIDS Normalized Throughput up to 21.5% up to 70.8% More Resource Competition with four NFs Packet Size (Byte) Packet Size (Byte) (c) IPSec+NIDS+Router (d) Firewall+IPSec+NIDS+Router

25 Scheduling Service chain scheduling scheme comparison G-NET: optimize the performance of the service chain FairShare: Evenly partition compute resources among NFs UncoShare: Each NF tries to maximize its resource allocation 1 Firewall + IPSec + NIDS + Router FairShare UncoShare G-NET Throughput Improvement % average improvement over FairShare 80.5% average improvement over UncoShare Packet Size (Byte)

26 Latency NIDS IPSec+NIDS Firewall+IPSec+NIDS Firewall+IPSec+NIDS+Router 1 95th 0.75 CDF th ,000 2,000 3,000 4,000 5,000 6,000 7,000 Round Trip Time (microsecond) 1.5 G-NET Temporal Sharing 1.5 Normalized Latency th: ~20% 95th: 25% - 44% IPSec+NIDS Firewall+ IPSec+NIDS Firewall+IPSec+ NIDS+Router 0.5 IPSec+NIDS Firewall+ IPSec+NIDS Firewall+IPSec+ NIDS+Router (a) 50th percentile latency (b) 95th percentile latency

27 Conclusion G-NET: An NFV system that efficiently utilizes s with spatial sharing Service chain based scheduling and resource allocation scheme Memory isolation with IsoPointer Abstraction that simplifies building -accelerated NFs Experimental Results (Compare with temporal sharing) Enhances throughput by up to 70.8% Reduces latency by up to 44.3% G-NET High-efficient platform for building -based NFV systems