UNIVERSITY OF CALIFORNIA, SAN DIEGO Network Performance Improvements For Web Services – An End-to-End View A dissertation submitted in partial satisfaction of the requirements for the degree of Doctor of Philosophy in Computer Science by Sivasankar Radhakrishnan Committee in charge: Amin Vahdat, Chair George Papen George Porter Stefan Savage Geoffrey M. Voelker 2014 Copyright Sivasankar Radhakrishnan, 2014 All rights reserved. The Dissertation of Sivasankar Radhakrishnan is approved and is accept- able in quality and form for publication on microfilm and electronically: Chair University of California, San Diego 2014 iii TABLE OF CONTENTS Signature Page . iii Table of Contents . iv List of Figures . vii List of Tables . ix Acknowledgements . x Vita................................................................. xiii Abstract of the Dissertation . xv Chapter 1 Introduction . 1 1.1 Challenges . 2 1.1.1 Wide Area Network Latency . 2 1.1.2 Data Center Network Infrastructure . 3 1.1.3 Co-tenancy and Network Performance Isolation . 3 1.2 Efficient Networking for Modern Web Services . 4 1.2.1 TCP Fast Open . 4 1.2.2 Dahu . 6 1.2.3 SENIC . 7 1.3 Organization . 7 Chapter 2 Design and Implementation Principles . 8 2.1 Principles . 8 2.1.1 State Management Principle . 8 2.1.2 Short Circuiting Principle . 9 2.1.3 Inheritance Principle . 9 2.2 Implications of the Principles . 10 2.2.1 State Management Principle . 10 2.2.2 Short Circuiting Principle . 12 2.2.3 Inheritance Principle . 14 2.3 Summary . 15 Chapter 3 TCP Fast Open . 16 3.1 Introduction . 16 3.2 Motivation . 18 3.2.1 Google Server Logs Analysis . 19 3.2.2 Chrome Browser Statistics . 20 3.3 Design . 22 iv 3.3.1 Context and Assumptions . 22 3.3.2 Design Overview . 24 3.3.3 Cookie Design . 26 3.3.4 Security Considerations . 26 3.3.5 Handling Duplicate SYN Segments . 29 3.3.6 API Changes . 30 3.4 Deployability . 32 3.4.1 New TCP Options / Data in SYN . 32 3.4.2 Server Farms . 33 3.4.3 Network Address Translation (NAT) . 34 3.4.4 TCP Option Space . 34 3.5 Implementation . 34 3.5.1 Kernel Support . 35 3.5.2 Application Support . 36 3.6 Evaluation . 36 3.6.1 Whole Page Download Performance . 36 3.6.2 Server Performance . 39 3.7 Discussion . 40 3.7.1 One Time Cookies . 40 3.7.2 Data After SYN . 42 3.7.3 Server-side TFO Cache . 43 3.7.4 TCP Fast Open in Low Latency Networks . 43 3.7.5 Cookie-less TCP Fast Open . 44 3.8 Related Work . 45 3.9 Summary . 47 3.10 Acknowledgments . 48 Chapter 4 Dahu: Commodity Switches for Direct Connect Data Center Networks 50 4.1 Introduction . 51 4.2 Motivation and Requirements . 53 4.2.1 Challenges . 54 4.2.2 Dahu Requirements and Design Decisions . 56 4.3 Switch Hardware Primitives . 57 4.3.1 Port Groups With Virtual Ports . 58 4.3.2 Allowed Port Bitmaps . 60 4.3.3 Eliminating Forwarding Loops . 61 4.4 Switch software . 62 4.4.1 Background on HyperX Topology . 63 4.4.2 Non-Minimal Routing . 64 4.4.3 Traffic Load Balancing . 68 4.4.4 Load Balancing Algorithm . 71 4.4.5 Load Balancing Heuristic . 72 v 4.4.6 Fault Tolerance . 73 4.5 Deployability . 73 4.6 Evaluation . 75 4.6.1 Simulator . 75 4.6.2 HyperX Networks . 77 4.6.3 Fat-Tree Networks . 82 4.6.4 MPTCP in HyperX Networks . 83 4.7 Discussion . 85 4.8 Related Work . ..