4003-541-70/4005-741-70 Data Communications and Networks II Module 1. High Performance Network Programming -- Lecture Notes Prof. Alan Kaminsky -- Spring Quarter 2006 Rochester Institute of Technology -- Department of Computer Science Echo Server Version 1 Whenever a client sets up a socket connection, read bytes from the socket and write (echo) them back to the socket   Any number of sockets may be open at the same time   Design "Old" I/O Class java.net.ServerSocket Class java.net.Socket Class java.io.InputStream Class java.io.OutputStream Multi-threaded design Main program thread accepts connections Each connection processed in its own separate thread Threads obtained from an executor, a.k.a. thread pool (JDK 1.5) Interface java.util.concurrent.ExecutorService Class java.util.concurrent.Executors   Package edu.rit.datacomm2.echo Class EchoServer1 -- source code Echo Server Version 2 Design New I/O (JDK 1.4) Buffers Class java.nio.Buffer Class java.nio.ByteBuffer Channels Class java.nio.channels.ServerSocketChannel Class java.nio.channels.SocketChannel Non-blocking I/O Class java.nio.channels.Selector Class java.nio.channels.SelectionKey Single-threaded design Main program thread does it all -- accepts connections, reads sockets, writes sockets   Package edu.rit.datacomm2.echo Class EchoServer2 -- source code Echo Server Version 3 Java New I/O is supposed to give better performance and scalability for server applications   Is this really the case?   Design Same as EchoServer1, plus it measures throughput -- bytes/sec read and written "Old" I/O Multi-threaded design with a thread pool Throughput measurement Class java.util.concurrent.atomic.AtomicInteger Interface java.util.concurrent.ScheduledExecutorService Class java.util.concurrent.Executors   Package edu.rit.datacomm2.echo Class EchoServer3 -- source code Echo Server Version 4 Design Same as EchoServer2, plus it measures throughput -- bytes/sec read and written New I/O Single-threaded design   Package edu.rit.datacomm2.echo Class EchoServer4 -- source code Echo Server Version 5 Design Similar to EchoServer1, plus it measures throughput -- bytes/sec read and written "Old" I/O Multi-threaded design, with a new thread created for each connection (no thread pool)   Package edu.rit.datacomm2.echo Class EchoServer5 -- source code Throughput Measurements "Hose" traffic generator Like streaming media Few connections, a large quantity of data per connection Traffic generator sets up one connection to the Echo Server, then reads and writes bytes as fast as it can -- like a fire hose   "Cannon" traffic generator Like a web browser Many connections, a small quantity of data per connection Traffic generator repeatedly sets up a connection to the Echo Server, writes a small block, reads the block, and closes the connection, as fast as it can -- like shooting cannonballs   Package edu.rit.datacomm2.echo Class Hose -- source code Class Cannon -- source code   Testbed I used the CS Department's paranoia cluster, normally used for parallel computing Echo Server ran on one backend processor 1 to 8 traffic generator clients ran on other backend processors Traffic went over the cluster's backend network, a 100 Mbps Ethernet   Throughput measurements (bytes/sec) Hose Client Cannon Client No. of ------------------------------------- ------------------------------------- Clients EchoServer3 EchoServer4 EchoServer5 EchoServer3 EchoServer4 EchoServer5 ------- ----------- ----------- ----------- ----------- ----------- ----------- 1 9.01e6 5.98e6 9.11e6 9.66e5 9.14e5 9.13e5 2 1.16e7 7.10e6 1.20e7 1.62e6 1.42e6 1.31e6 3 1.15e7 7.73e6 1.21e7 1.78e6 1.65e6 1.43e6 4 1.17e7 7.88e6 1.20e7 1.81e6 1.68e6 1.45e6 8 1.17e7 8.50e6 1.17e7 1.81e6 1.76e6 1.42e6 Conclusions? The RIT Overlay Network (RON) Purpose To further illustrate high performance network software design using Java New I/O To serve as a testbed for studying data communications in the rest of the course   RON Architecture Layers Protocols +--------------------------+ | Application Layer | +--------------------------+ | RON Transport Layer | Coming soon +--------------------------+ | RON Network Layer | RONP +--------------------------+ | Internet Transport Layer | UDP +--------------------------+ | Internet Network Layer | IP +--------------------------+ | Data Link Layer | Ethernet +--------------------------+ | Physical Layer | +--------------------------+ The RIT Overlay Network Protocol (RONP) RON's network layer protocol is the RIT Overlay Network Protocol (RONP). RONP provides an unreliable packet (datagram) delivery service. RONP addressing. Each RON endpoint has a unique RONP address, a 32-bit number. RONP packet format. A RONP packet consists of a 12-byte header followed by a payload of 0-536 bytes. The header fields are: Bytes 0-3 -- Destination RONP address, most significant byte first. This is the address of the RON endpoint to which the packet must be delivered.   Bytes 4-7 -- Source RONP address, most significant byte first. This is the address of the RON endpoint that generated the packet.   Byte 8 -- RON transport protocol ID. This designates the transport protocol at the next higher layer that must process the packet.   Byte 9 -- Time to live (TTL) in the range 0-255. The source endpoint initializes the TTL field. Each time a RON router receives a packet, the router decrements the TTL field. If the TTL field is 0 after the decrement, the router drops the packet rather than forwarding it.   Byte 10 -- Reserved for future expansion.   Byte 11 -- Reserved for future expansion.   The maximum RONP payload size, 536 bytes, is chosen so that a RONP packet will not be fragmented when transferred over the Internet, assuming the minimum Internet MTU of 576 bytes (536 bytes RONP payload + 12 bytes RONP header + 8 bytes UDP header + 20 bytes IP header = 576 bytes MTU).   Package edu.rit.ron Interface Constants -- source code Interface PacketListener -- source code Interface RouterHooks -- source code Interface RoutingAlgorithm -- source code Class ManualRoutingAlgorithm -- source code Class Packet -- source code Class Router -- source code Class RouterFactory -- source code Class RouterInterface -- source code   Package edu.rit.ron.test Class Test01 -- source code   RON router configuration   RON router operation   Design of the RON software Class RouterPort encapsulates a datagram channel Class Router encapsulates an I/O thread Single-threaded design; I/O thread reads and writes all channels using a selector Slow data links simulated using timers Send a packet, wait x milliseconds before sending the next packet Receive a packet, wait x milliseconds before forwarding or delivering the packet x depends on the packet length and the port's data rate (bits/sec)   More on class java.nio.ByteBuffer Capacity, limit, position Byte reading and writing operations Primitive data type reading and writing operations Clear, flip operations   Demonstrations The RON Control Message Protocol (RCMP) RCMP is to RON as the Internet Control Message Protocol (ICMP) is to the Internet Protocol (IP)   RFC 792 -- Internet Control Message Protocol (Postel, 1981)   RCMP messages (patterned after ICMP) Echo request Echo reply   Package edu.rit.ron.rcmp Interface Constants -- source code Class RcmpEntity -- source code   Demonstrations Data Communications and Networks II 4003-541-70/4005-741-70 Spring Quarter 2006 Course Page Alan Kaminsky • Department of Computer Science • Rochester Institute of Technology • 4486 + 2220 = 6706 Home Page Copyright © 2006 Alan Kaminsky. All rights reserved. Last updated 23-Mar-2006. Please send comments to ark­@­cs.rit.edu.