tcp high performance.pdfVIP

1? 2004 Cisco Systems, Inc. All rights reserved.Presentation_ID High-performance TCP flows and methods June-2004 Larry Dunn Manager, Advanced Architecture ldunn@ 2? 2004 Cisco Systems, Inc. All rights reserved.Presentation_ID Goals ? Provide insight as to why high-performance networks often have low per-flow throughput ? Discuss diagnostic tools available ? Examine common problems, and methods/proposals for increasing TCP throughput ? Review experimental results, and case studies ? Take-away: understand and be able to improve TCP performance; relate to developers and/or end users who might assert “there’s a problem with your network…” ? Audience assumption: have some knowledge of how TCP works (but we’ll review basics) 3? 2004 Cisco Systems, Inc. All rights reserved.Presentation_ID Agenda ? Problem statement ? Instrumentation: Web100, Net100 ? Path MTU Discovery ? TCP dynamics review(sawtooth) ? Larger MTU (and renewed PathMTU) ? Change TCP Dynamics approaches ? Experimental Results ? Case Studies 4? 2004 Cisco Systems, Inc. All rights reserved.Presentation_ID Problem Statement (1) ? Over the past 15 years, many components impacting end-2-end performance have roughly tracked “Moore’s Law” (doubling each 18-24 months): End-station interface speed(shared 10Mbps - switched 1Gbps) Campus access link speed (1.5Mbps - 600+Mbps) WAN/backbone links (45Mbps - 10Gbps) Processor speed(40Mhz - 2Ghz), memorydisk bandwidth, etc. ? But not end-2-end throughput, for high bandwidth*delay product paths… why? 5? 2004 Cisco Systems, Inc. All rights reserved.Presentation_ID Problem Statement (2) ? 100Mbps end-2-end hosts often get 5-10mbps throughput ? 1st issue: transmit/receive buffers need to accommodate full bandwidth*delay product (web100, autotuning are addressing this) ? 2nd issue: TCP dynamics, behavior in presence of congestion- or BER-induced loss, time required to achieve “full utilization” on high- speed links (related- pathMTU discovery broken; L3/L4 revis