Docker cker Ov Overlay rlay Networks tworks Performance analysis - PowerPoint PPT Presentation

Docker cker Ov Overlay rlay Networks tworks Performance analysis in high-latency environments Students: s: Siem Hermans Patrick de Niet Resea earc rch Project ect 1 Supervi rviso sor: Dr. Paola Grosso System and Network Engineering System and Network Engineering

2 Resea search rch question on “What is the performance of various Docker overlay solutions when implemented in high latency environments and more specifically in the GÉANT Testbeds Services (GTS )?”

3 Relat lated ed Work Internal • Claassen, J. (2015, July). Container Network Solutions. Retrieved January 31, 2016, from http://rp.delaat.net/2014-2015/p45/report.pdf. • Rohprimardho, A. (2015, August). Measuring The Impact of Docker on Network I/O Performance. Retrieved January 31, 2016, from http://rp.delaat.net/2014-2015/p92/report.pdf. External • Kratzke, N. (2015). About Microservices, Containers and their Underestimated Impact on Network Performance. CLOUD COMPUTING 2015, 180. • Barker, S. K., & Shenoy, P. (2010, February). Empirical evaluation of latency-sensitive application performance in the cloud. In Proceedings of the first annual ACM SIGMM conference on Multimedia systems (pp. 35-46). ACM.

4 Docker er - Concepts Basics • Containerization • Gaining traction • Performance increases • Role of Docker Conta taine ner Virt rtual ual Machi hine ne

5 Multi-host st networki working ng • Virtual networks that span underlying hosts • Powered by libnetwork

6 Overlay erlay solutions ns Libnetwork Weave Net Flannel Project Calico (Native overlay driver) • Based on SocketPlane • Previously routing based • Flanneld agent • Technically not an overlay on pcap . Now uses OVS. • Integrating OVS APIs in • No integration with • Routing via BGP Docker • Libnetwork plugin libnetwork • Segmentation via iptables • Subnet per host • VXLAN based forwarding • VXLAN based forwarding • State distribution via BGP • UDP or VXLAN forwarding route reflectors Kratzke, N. (2015). • No tunneling

7 GÉANT NT - Introduction • European research community - Amsterdam - Bratislava - Ljubljana - Milan - Prague GÉANT Testbeds Service (GTS) • OpenStack platform, interconnected by MPLS • KVM for compute nodes • Resembles IaaS providers; Shared infrastructure •

8 Topologies Topologies (1) 1) • Four full mesh instances • • DSL 2.0 grammar (JSON) • Local site; Feasibility evaluation DSL FullMesh { id="FullMesh_Dispersed" host { id= "h1" location= "AMS" port { id="port11"} port { id="port12"} } link { id="l1" port {id="src"} port {id="dst"} } adjacency h1.port14, l1.src adjacency h2.port24, l1.dst } {...}

9 Topologies (2) Topologies 2) • Scaling up from single-site feasibility check • Calico o droppe ropped • Full mesh divided in: 1. 1. Point oint-to to-poi point, synthetic benchmarks 2. 2. Star ar top opol ology gy, real-world scenario Setup • Flannel VXLAN tunneling • Key-value store placement • Storing network state • Separate distributed system

10 Methodology logy - Performance Synthe ntheti tic benchm nchmar ark (PtP tP) Placement of nodes • Netperf • Latency • Jitter Iperf • TCP/UDP throughput • Jitter Laten atency cy sensi nsiti tive ve app pplica cati tion n (Medi dia stream reaming) g) • Darwin Streaming Server, Faban RTSP clients • Jitter (with netperf) • Bitrate Barker, S. K., & Shenoy, P. (2010, February).

11 Resu sults ts - GÉANT Documentatio ion Provis isio ionin ing Acces cess Setup up Support VPN Resources rces

12 Resu sults ts - PtP VM to VM Latency

13 Resu sults ts - PtP Docker to Docker Latency In Milliseconds (ms) 99 th % Latency Circuit Topology Min. Latency Mean Latency LIBNET 36.3 36.5 37.0 AMS – MIL WEAVE 36.2 36.5 37.0 FLANNEL 42.5 42.9 43.0 LIBNET 30.1 30.3 31.0 AMS – LJU WEAVE 29.8 30.3 31.0 FLANNEL 29.8 30.3 31.0 LIBNET 17.6 17.7 18.0 AMS – BRA WEAVE 17.4 17.7 18.0 FLANNEL 17.4 17.7 18.0 LIBNET 61.8 62.1 62.4 MIL – LJU WEAVE 59.6 59.8 60.0 FLANNEL 55.6 55.8 56.0 LIBNET 12.7 13.0 14.0 MIL – BRA WEAVE 12.9 13.1 14.0 FLANNEL 12.9 13.1 14.0 LIBNET 47.1 47.4 48.0 BRA – LJU WEAVE 43.1 59.5 130.0 FLANNEL 43.1 43.4 44.0

14 Resu sults ts - PtP Throughput AMS to BRA TCP Throughput AMS to BRA UDP Throughput VM VM Flannel Flannel Solution Solution Weave Weave Libnet Libnet 0 50 100 150 200 250 0 50 100 150 200 250 300 Mbps Mbps

15 Mean Jitter in ms 0.5 1.5 2.5 0 1 2 Resu 1 Worker sults VM BRA - AMS Concurrency Jitter 3 Worker 9 Worker ts - Streaming Experiment 1 Worker LIBNET 3 Worker BRA – AMS Instance 9 Worker 1 Worker WEAVE 3 Worker 9 Worker 1 Worker FLANNEL 3 Worker 9 Worker Bitrate per stream in Mbps 0.00 0.20 0.40 0.60 0.80 1.00 1.20 1.40 1.60 1.80 2.00 1 Worker BRA - AMS Concurrency Bitrate VM 3 Worker 9 Worker 1 Worker LIBNET 3 Worker BRA – AMS Instance 9 Worker 1 Worker WEAVE 3 Worker 9 Worker Mean 1 Worker FLANNEL Maximum 3 Worker 9 Worker

16 Conclus usion ion & Fu Future e Work Measurements currently only valid within GTS environment; • Reconduct performance analysis in heavily shared environment (e.g. Amazon EC2) – Perform experiments with more compute resources (CPU capping) – Anomalies in throughput performance not identified (UDP, TCP) • Similar behavior discovered in the work of J. Claassen – Ideally more measurements to increase accuracy • No significant performance degradations by implementing Docker overlays within GTS • Use Weave ideally within the GTS environment •

Resea earc rch Project ect 1 System and Network Engineering System and Network Engineering Qu Ques estio ions ns ? Thank you A: Science Park 904, Amsterdam NH github.com/siemhermans/gtsperf W: rp.delaat.net siem.hermans@os3.nl patrick.deniet@os3.nl