1. Internet and Intranet Protocols and Applications Section V: Network Application Performance Lecture 11: Why the World Wide Wait? 4/11/2000 Arthur P. Goldberg Computer Science Department New York University artg@cs.nyu.edu

2. Performance definitions Latency, delay: how long? –Ie, message transport latency Response time: how long to respond? –Ie, time from request to response Bandwidth Utilization: fraction (%) used Throughput: # processed / time

3. In Application Protocols, Where’s The Time Go? Trace the steps –Client issues request –Network transports request –Server receives request –Server generates and sends response –Network transports response –Client receives and displays response Complications in this model –Multiple concurrent requests (how long until the last finishes?) –Network protocol delays –Server/proxy hierarchy

4. Primary Components of Delay “Critical path” computation Transmission, ie signal travel time =~ (2/3)c Congestion (queueing) –Network Ie, outgoing links on routers –Servers All components: CPU, disks, memory, synchronization, etc. Protocol (and, sometimes, system) waits –Eg in TCP Delayed ACK Slow start

5. Tanenbaum, Section 6.6, Perf. Problems in Networks Problems –Congestion –Overload -> data loss -> retransmission Eg, bandwidth mismatch Eg, broadcast storm (errors, booting, anything synchronous) –Timeout Too short: excessive retransmissions Too long: slow recovery –Underutilization

6. Underutilization –An underutilized pipe has less effective bandwidth –Bandwidth – delay product: capacity of a pipe Bandwidth * delay Eg: cross country T1 line: –70 ms RTT –1.544 Mbps –Bandwidth – delay product = 108,080 bits Consider maximum bits in flight –Eg, TCP window: if 8 KB then max bw only: (T1 speed)* 64,000/108,080 = 0.91 Mbps

7. Delay Estimate Rule-of-Thumb Draw delay vs. load function Load ranges –Light: 0.0 to 0.5 –Moderate: about 0.5 to 0.95 –Heavy: 0.95 to 1.0 –Over: over 1.0 Under moderate load –Delay varies as 1/(1- Utilization) –(From first principles and distribution assumptions and queueing theory) Under heavy –Avoid thrashing Under over load –Shed work: Mogul and Ramakrishnan, Eliminating Receive Livelock in an Interrupt-driven Kernel.

8. Tanenbaum 6.6.3: System Design for Better Performance CPU speed is more important than network speed OS and protocol overhead dominates Reduce packet count to reduce software overhead –Ie, use bigger packets Minimize context switches Minimize copying Pre-compute whenever possible –Header prediction –Pre-compute partial checksum You can buy more bandwidth, but not lower delay Avoiding congestions is better than recovering from it Avoid timeouts

9. Impact of Fiber In the race between computing and communication, communication won. The full implications of essentially infinite bandwidth (although not at zero cost) have not yet sunk in to a generation of computer scientists and engineers taught to think in terms of the low Nyquist and Shannon limits imposed by copper wire. The new conventional wisdom should be that all computers are hopelessly slow, and networks should try to avoid computation at all costs, no matter how much bandwidth that wastes. Andrew S. Tanenbaum, Computer Networks, 1996

10. Padmanabhan & Mogul, Improving HTTP Latency Distribution of document sizes –Show curve Avoid extra round trips Do pipelining –How do sequential requests and pipelining interact with a proxy server?

11. OTHER NYU WebPerf measurements –Live WebPerf Demo? Repair of HTTP 1.0 performance bugs in HTTP 1.1. Multiple concurrent requests –Cool paper on Browser rendering Delayed ACKs Problem –Interactions Between Delayed Acks and Nagle's Algorithm in HTTP and HTTPS: Problems and Solutions In http://WWW.CS.NYU.EDU/artg/http://WWW.CS.NYU.EDU/artg/

12. Admin Final? Due date of Phase II Changes to Phase II

13. IIPA Sections Intro to Application Protocols and Socket Programming Web Vacation Email Network Application Performance