Current Network Research on UKLIGHT Saleem Bhatti Networked and Distributed Systems (NDS) Research Group Computer Science, University of St. Andrews http://www.dcs.st-andrews.ac.uk/~saleem/ St. Andrews, Computer Science 1 46PaQ & MASTS • Overview • Current Status and Results • Future Work • Questions St. Andrews, Computer Science 2 • Overview • Current Status and Results • Future Work • Questions St. Andrews, Computer Science 3 Technology Innovation Network Science Applications UKLIGHT Network projects ESLEA Applications and the Network MASTS Analysis at All Scales in Time & Space Real-time Monitoring Towards Next Generation Networks Protocol Innovation Behaviour And Performance For QoS and Control Data Acquisition, Storage & Archiving Compression (Model Free & Model Dependant) St. Andrews, Computer Science 46PaQ Data Analysis and Traffic Modelling Front End Enhancements: Advanced Protocols SuperJANET and UKLIGHT 4 46PaQ overview [1] • IPv4 & IPv6 Performance and QoS EPSRC 01 Dec 2004 - 30 Nov 2006 (June 2007) • Sites: Computer Science, St. Andrews Computer Science, UCL Computer Science, QMUL St. Andrews, Computer Science 5 46PaQ overview [2] • Network & transport protocols at multi-Gb/s: UCL & St. Andrews lead New protocols (e.g. DCCP, TCP ‘extensions’): » Collaboration with ESLEA Engineering of the protocols and end-systems • Traffic monitoring at mulit-Gb/s: QMUL leads Multi-protocol analysis: » Collaboration with MASTS Traffic modelling St. Andrews, Computer Science 6 MASTS overview [1] • Meausrement at All Scales in Time and Space EPSRC 01 Sep 2004 - 31 Aug 2007 • Sites: E&E Eng, UCL Computer Science & E&E Eng, Loughborough Computer Science, QMUL Computer Science, St. Andrews St. Andrews, Computer Science 7 MASTS overview [2] • Traffic flow statistics • Growth and scaling of network: Traffic Topology • Monitoring systems: Data capture Data processing • Traffic modelling St. Andrews, Computer Science 8 • Overview • Current Status and Results • Future Work • Questions St. Andrews, Computer Science 9 46PaQ current experiments • Connectivity to UKLIGHT: UCL: 2 × 1Gb/s to Chicago, remote loopback • 1Gb/s loopback experiments on UKLIGHT from Q1/2005 (London-Chicago) • High-speed tests within the lab: bi-directional data flows (not just TCP/FTP) multiple data flows (not just single flows) • Examine end-to-end data path: Interaction between application, hosts & network Performance tuning for high-speed operation St. Andrews, Computer Science 10 TCP Throughput, Chicago-loop [1] St. Andrews, Computer Science 11 TCP Throughput, Chicago-loop [2] Multiple-flows, aggregate throughput St. Andrews, Computer Science 12 Observations • Verification of known results • Protocol performance very much dependant on several parameters (UDP buffer sizes, TCP window size and sender recv buffer size, CPU/HW speed) For TCP, window size is important for paths with large bandwidth-delay products • Tuning of protocol stack and/or application is therefore the most important performance factor St. Andrews, Computer Science 13 TCP ‘Fairness’ - Very Early Results Matched bit-rate fairness (MBF) Chicago-loop, 16MB window St. Andrews, Computer Science Chicago-loop, 32MB window # & % ( Nb (t) Fk (t) = 10log10 % N k (, k = 1,K,N % ( % " bn (t) ( $ n=1 ' 14 MASTS current work • Set up network monitoring systems • UKERNA - AUP for monitoring • Storage of data: Data has to be moved to processing site Lots of data - summarisation and compression • Data processing and analysis: How to process? What are we looking for? • Monitoring of ESLEA links St. Andrews, Computer Science 15 Prototype MASTS monitor Loughborogh University: David Parish Iain Phillips Konstantinos Kyriakopoulos Mark Whithall St. Andrews, Computer Science 16 Example: No problems St. Andrews, Computer Science 17 Example: Some Problems St. Andrews, Computer Science 18 Example: Serious Problems St. Andrews, Computer Science 19 • Overview • Current Status and Results • Future Work • Questions St. Andrews, Computer Science 20 46PaQ • • • • Starting DCCP experiments Interaction with ECN and DIFFSERV Performance on IPv6 Continued co-operation with other projects: MASTS (monitoring) ESLEA (protocol work, e.g. DCCP) St. Andrews, Computer Science 21 MASTS • Monitoring systems at UKLIGHT PoPs: Monitor 1Gb/s and 10Gb/s Data storage and archiving • Traffic and performance analysis: How to monitor performnace at such high speeds? Traffic models at different timescales • Access network vs. Core network St. Andrews, Computer Science 22 • Overview • Current Status and Results • Future Work • Questions St. Andrews, Computer Science 23 Additional slides St. Andrews, Computer Science 24 Equipment setup • End hosts (one as source, one as sink): Gentoo Linux (AMD64 compiled), kernel 2.6.11 2x Opteron 246, 1MB L2 Cache, 2.4GHz 2GB PC3200 DDR RAM on dual banks Intel PCI-X 133MHz 64Bit dual 1Gb/s NIC • Local switch: Cisco Catalyst 3750 2x VLAN, one for each of the 1Gb/s links to Chicago • Tests run for two configurations: Local-switch loopback 1Gb/s (RTT fraction of ms) Chicago loopback 1Gb/s (RTT ~177ms) St. Andrews, Computer Science 25 TCP Baseline Tests • The following TCP settings were used: net.ipv4.tcp_window_scaling = 1 net.ipv4.tcp_sack = 1 net.ipv4.tcp_rmem = 4096 87380 67108864 net.ipv4.tcp_wmem = 4096 87380 67108864 net.ipv4.tcp_mem = 196608 262144 67108864 net.core.rmem_default = 126976 net.core.wmem_default = 126976 net.core.rmem_max = 33554432 net.core.wmem_max = 33554432 • All experiments conducted using iperf St. Andrews, Computer Science 26 TCP Throughput, local-loop [1] St. Andrews, Computer Science 27 TCP Throughput, local-loop [2] Multiple-flows, aggregate throughput St. Andrews, Computer Science 28