Packet-based Network Traffic Monitoring and Analysis with GPUs Wenji Wu, Phil DeMar wenji@fnal.gov, demar@fnal.gov GPU Technology Conference 2014 March 24-27, 2014 SAN JOSE, CALIFORNIA
Background Main uses for network traffic monitoring & analysis tools: Operations & management Capacity planning Performance troubleshooting Levels of network traffic monitoring & analysis: Device counter level (snmp data) Traffic flow level (flow data) Packet level (The Focus of this work) security analysis application performance analysis traffic characterization studies 2
Characteristics of packet-based network monitoring & analysis applications Time constraints on packet processing. Compute and I/O throughput-intensive High levels of data parallelism. Background (cont.) Packet parallelism. Each packet can be processed independently Flow parallelism. Each flow can be processed independently Extremely poor temporal locality for data Typically, data processed once in sequence; rarely reused 3
The Problem Packet-based traffic monitoring & analysis tools face performance & scalability challenges within highperformance networks. High-performance networks: 40GE/100GE link technologies Servers are 10GE-connected by default 400GE backbone links & 40GE host connections loom on the horizon. Millions of packets generated & transmitted per sec 4
Monitoring & Analysis Tool Platforms (I) Requirements on computing platform for high performance network monitoring & analysis applications: High Compute power Ample memory bandwidth Capability of handing data parallelism inherent with network data Easy programmability 5
Monitoring & Analysis Tool Platforms (II) Three types of computing platforms: NPU/ASIC CPU GPU Features NPU/ASIC CPU GPU High compute power Varies High memory bandwidth Varies Easy programmability Data-parallel execution model Architecture Comparison 6
Our Solution Use GPU-based Traffic Monitoring & Analysis Tools Highlights of our work: Demonstrated GPUs can significantly accelerate network traffic monitoring & analysis 11 million+ pkts/s without drops (single Nvidia M2070) A generic I/O architecture to move network traffic from wire into GPU domain Implemented a GPU-accelerated library for network traffic capturing, monitoring, and analysis. Dozens of CUDA kernels, which can be combined in a variety of ways to perform monitoring and analysis tasks 7
Key Technical Issues GPU s relatively small memory size: Nvidia M2070 has 6 GB Memory, K10 has 8 GB Memory Workarounds: Mapping host memory into GPU with zero-copy technique? Partial packet capture approach Need to capture & move packets from wire into GPU domain without packet loss Need to design data structures that are efficient for both CPU and GPU 8
System Architecture Four Types of Logical Entities: Traffic Capture Preprocessing Monitoring & Analysis Output Display 1. Traffic Capture 2. Preprocessing 3. Monitoring & Analysis GPU Domain 4. Output Display Captured Data... Packet Buffer Packet Buffer Output Output Output Capturing Packet Chunks Monitoring & Analysis Kernels Output User Space NICs... Network Packets 9
Packet I/O Engine Processing Data Recycle Capture User Space OS Kernel Packet Buffer Chunk Attach... Descriptor Segments Free Packet Buffer Chunks Recv Descriptor Ring NIC Incoming Packets... Goal: To avoid packet capture loss Key techniques A novel ring-buffer-pool mechanism Pre-allocated large packet buffers Packet-level batch processing Memory mapping based zero-copy Key Operations Open Capture Recycle Close 10
GPU-based Network Traffic Monitoring & Analysis Algorithms A GPU-accelerated library for network traffic capturing, monitoring, and analysis apps. Dozens of CUDA kernels Can be combined in a variety of ways to perform intended monitoring & analysis operations 11
Packet-Filtering Kernel index 0 1 2 3 4 5 6 7 raw_pkts [ ] filtering_buf [ ] p1 p2 p3 p4 p5 p6 p7 p8 1 x x x x 0 1 1 0 0 1 0 1 Filtering We use Berkeley Packet Filter (BPF) as the packet filter scan_buf [ ] filtered_pkts [ ] index 0 1 1 2 3 3 3 4 0 1 2 3 p1 p3 p4 p7 index 0 1 2 3 4 5 6 7 2 Scan 3 Compact A few basic GPU operations, such as sort, prefix-sum, and compact. Advanced packet filtering capabilities at wire speed are necessary so that we only analyze those packets of interest to us. 12
Sample Use Case -1 Using our GPU-accelerated library, we developed a sample use case to monitor network status: Monitor networks at different levels: from aggregate of entire network down to one node Monitor network traffic by protocol Monitor network traffic information per node: Determine who is sending/receiving the most traffic For both local and remote addresses Monitor IP conversations: Characterizing by volume, or other traits. Packet Parallelism 13
Sample Use Case 1 Data Structures Three key data structures were created at GPU: protocol_stat[] an array that is used to store protocol statistics for network traffic, with each entry associated with a specific protocol. ip_snd[] and ip_rcv[] arrays that are used to store traffic statistics for IP conversations in the send and receive directions, respectively. ip_table a hash table that is used to keep track of network traffic information of each IP address node. These data structures are designed to reference themselves and each other with relative offsets such as array indexes. 14
Sample Use Case 1 Algorithm 1. Call Packet-filtering kernel to filter packets of interest 2. Call Unique-IP-addresses kernel to obtain IP addresses 3. Build the ip_table with a parallel hashing algorithm 4. Collect traffic statistics for each protocol and each IP node 5. Call Traffic-Aggregation kernel to build IP conversations 15
Sample use case 2 Using our GPU-accelerated library, we developed a sample use case to monitor bulk data transfer: Monitor bulk data transfer rate from aggregate of entire network down to one node Analyze data transfer performance bottlenecks Sender-limited, network-limited, or receiver-limited? Network-limited? Packet drops, or packet reordering? Packet Parallelism + Flow Parallelism 16
Sample Use Case 2 Algorithm Monitoring Mode: 1. Call Packet-filtering kernel to filter packets of interest 2. Call TCP-throughput kernel to calculate data transfer rates Analyzing Mode: 1. Call Packet-filtering kernel to filter packets of interest 2. Classify packets into flows 3. Call TCP-analysis kernel to analyze performance bottlenecks 17
Node 0 Node 1 Prototyped System Mem QPI M2070 PCI-E 10G-NIC PCI-E 10G-NIC P0 I O H QPI QPI P1 I O H QPI PCI-E Mem Prototyped System Our application is developed on Linux. CUDA 5.0 programming environment. The packet I/O engine is implemented on Intel 82599-based 10GigE NIC A two-node NUMA system Two 8-core 2.67GHz Intel X5650 processors. Two Intel 82599-based 10GigE NICs One Nvidia M2070 GPU. 18
Performance Evaluation Packet I/O Engine Packet drop rate 0.5 0.4 psioe-1.6 psioe-2.0 psioe-2.4 netmap-1.6 netmap-2.0 netmap-2.4 gpu-i/o-1.6 gpu-i/o-2.0 gpu-i/o-2.4 0.3 0.2 0.1 Packet drop rate one-nic-exp 0 1000 10000 100000 1000000 10000000 Number of packets Packet drop rate two-nic-exp Our Packet I/O engine (GPU-I/O) achieves better performance than NETMAP and PSIOE in avoiding packet capture loss. 19
Throughput (Million p/s) Performance Evaluation GPU-based Packet Filtering Algorithm 180 160 140 120 100 80 60 40 20 0 Data-set-1 Data-set-2 Data-set-3 Data-set-4 mmap-gpu s-cpu-1.6 s-cpu-2.0 s-cpu-2.4 m-cpu-1.6 m-cpu-2.0 m-cpu-2.4 st-gpu 20
Performance Evaluation GPU-based Sample Use Case - 1 Throughput (Million p/s) 100 90 80 70 60 50 40 30 20 10 0 s-cpu-1.6g s-cpu-2.0g s-cpu-2.4g m-cpu-1.6 m-cpu-2.0 m-cpu-2.4 st-gpu IP udp net 129.15.30 net 131.225.107 21
Performance Evaluation Overall System Performance Packet Ratio Packet Ratio 1 0.8 0.6 0.4 One NIC Experiments 1.6GHz 2.0GHz 2.4GHz OnDemand 50 100 150 200 250 300 Packet Size (Byte) 1 0.8 0.6 0.4 Two NIC Experiments 1.6GHz 2.0GHz 2.4GHz OnDemand 50 100 150 200 250 300 Packet Size (Byte) In the experiments: Generators transmitted packets at wire rate. Packet sizes are varied across the experiments. Prototype system run in full operation with sample use case 1. 22
Key Take Aways GPUs accelerate network traffic monitoring & analysis. A generic I/O architecture to move network traffic from wire into GPU domain. A GPU-accelerated library for network traffic capturing, monitoring, and analysis. Dozens of CUDA kernels, which can be combined in various ways to perform monitoring and analysis tasks. 23
Question? Thank You! Email: wenji@fnal.gov and demar@fnal.gov 24