Performance Evaluation of Multipath Performance Evaluation of Multipath TCP Linux Implementations Euroview 2011 Euroview 2011 Amanpreet Singh C. Görg, A. Timm-Giel, M. Scharf, T.-R. Banniza 2 nd August 2011 Outline � Motivation � TCP-based Multipath Protocol Implementations � MPTCP � MCTCP � Experimental Setup � Experimental Setup � G-Lab Experimental Facitility - TOMATO � P � Performance Evaluation f E l ti � Conclusion & Outlook Euroview 2011 2 2
Motivation � End-hosts are often equipped with multiple interfaces � allows for deploying multipath transport to increase throughput, improve resilience and balance congestion in the network [RFC 6182] resilience and balance congestion in the network [RFC 6182] Internet Internet Access Network ce ce vi End Devi vice Device End De d De End En d Device vice nd De Access Networ Access Netw ork k En Access Network � This has led to the design and development of Multipath TCP solutions � a set of extensions for TCP that allows spreading of a single TCP flow across multiple subflows Euroview 2011 3 3 Multipath TCP Variants � Different solutions were published in the Multipath TCP working group in IETF � Different solutions were published in the Multipath TCP working group in IETF as drafts � MultiPath TCP (MPTCP) � MultiPath TCP (MPTCP) user application space socket API [draft-ietf-mptcp-multiaddressed-03] MPTCP kernel � MPTCP’s Linux kernel implementation is open source space TCP TCP IP IP IP IP [https://scm info ucl ac be/trac/mptcp/downloads] [https://scm.info.ucl.ac.be/trac/mptcp/downloads] � each multipath TCP subflow looks to the network as a normal TCP flow � TCP option field is used for signalling information exchange � PayLoad Multi-connection Transport (PLMT) application user socket API [draft-singh-mptcp-plmt-00] space PLMT control socket API � � TCP TCP TCP TCP TCP TCP user space solution, avoids TCP stack modification user space solution avoids TCP stack modification kernel kernel space IP IP IP � encodes signalling information in the payload � Multiple Connection TCP (MCTCP) p ( ) user application space [draft-scharf-mptcp-mctcp-01] socket API user or MCTCP kernel space socket API � A hybrid variant transparent in the single-path case TCP TCP kernel (extended) space � � TCP option field is used only for connection setup TCP option field is used only for connection setup IP IP IP IP Euroview 2011 4 4
Performance Evaluation Setup � Use Topology Management Tool (ToMaTo) - a virtual networking testbed to design and use virtual networking topologies (consisting of devices and connectors) connectors) � End-hosts (KVM Devices) - KVM virtualization technology with hardware emulated by qemu � Network connectivity is through Tinc VPN - Configurable link delay and bandwidth y g g y � Linux Vanilla Kernel 2.6.32 patched with MCTCP and 2.6.35 patched with MPTCP are used to boot KVM devices MPTCP are used to boot KVM devices � Test Strategy � Client and Server programs in C are used to shape application traffic (10Mbps, 20Mbps or higher) � Socket buffer size is set to 262,144B to avoid TCP flow control � Li k b Link bandwidth is set to 10Mbps and delay is varied between 1ms to 100ms d idth i t t 10Mb d d l i i d b t 1 t 100 � Numerical results are averaged over 5 runs Euroview 2011 5 5 G-Lab Experimental Facility http://toma http://tomato.ge http://tomato german http://toma german-la erman-la lab de lab.de de Directly Connected Scenario – Path per subflow Euroview 2011 6 6
Path per Subflow – Load Distribution Server Rate 10Mbps, Link delay 10ms, Cubic Congestion Control Server Rate 10Mbps Link delay 10ms Cubic Congestion Control MCTCP MPTCP oss -0% EV3 Event 3 (EV3): 50-60s Hi h Higher delay of 100ms d l f 100 Event Lo on one (red) link EV2 % Event Loss -2% Event 2 (EV2): 30-40s Packet loss only on one (black) links vent Loss -5% EV1 Event 1 (EV1): 10-20s Packet loss on both links Ev oss -10% Event L MCTCP is robust to link dynamics (packet loss/e2e Delay) and performs fair scheduling over its subflows Euroview 2011 7 7 Path per Subflow - Background Traffic - 2Mbps % vent Loss -0% stion Control MCTCP Traffic Server Server Client Client Cubic Conges E k Delay 10ms, t Loss -10% MCTCP + UDP Traffic Link Even MCTCP can adapt to asymmetric path quality of its subflows and can aggregate the available bandwidth Euroview 2011 8 8
Path per Subflow – Packet Loss on both Links MCTCP may use coupled* # congestion control optionally * draft-ietf-mptcp-congestion-03 draft ietf mptcp congestion 03 # see also MCTCP: A Multipath Transport Shim Layer Euroview 2011 9 9 Conclusion � Multipath TCP solutions adhere to the requirement goals of the multipath TCP architecture � increase throughput, more resilient g p � offer reliable, in-order transport being transparent to applications � MCTCP � is optimized for bulk data transfer (multipath operation is initiated after a short delay – can introduce overhead) � is robust to dynamic changes in the network such as variations in the packet loss rate, end-to-end delay and available bandwidth � more MCTCP performance results and implementation details are published in “MCTCP: A Multipath Transport Shim Layer,” M. Scharf and T.-R. Banniza, Gl b Globecom 2011 2011 � MPTCP � � is a kernel-based solution that uses TCP options field for signalling is a kernel-based solution that uses TCP options field for signalling � the current scheduling strategy doesn't seem to work well in all cases � more elaborate results are presented in „ MultiPath TCP: From Theory to Practice,“ S. Barre´, C. Paasch, and O. Bonaventure, IFIP Networking, 2011 Practice, S. Barre , C. Paasch, and O. Bonaventure, IFIP Networking, 2011 Euroview 2011 10 10
Outlook � Multipath TCP tests � with larger network topologies � in heterogeneous environment � for different application types � Design Issues � How many MxTCP subflows? � Criteria to close an Criteria to close an underperforming subflow nderperforming s bflo � Scheduler to minimize reordering between subflows and resulting jitter � Implementation Issues � Implementation Issues � APIs for multipath-aware applications � Both MPTCP and MCTCP implementation work is in progress and hence evolving to provide a better solution for Future Internet demands Euroview 2011 11 11
Recommend
More recommend
