1. Distributed Packet Rewriting and its Application to Scalable Server Architectures The 6 th IEEE International Conference on Network Protocol, Oct. 1998 A. Bestavros, M. Crovella, J. Liu, and D. Martin September 7, 2000 Chang Kyu Lee

2. Chang Kyu Lee, Dept. of EECS, KAIST. 2 Introduction Distributed WWW Server System DNS approach (1994 ~ 1996) Significant load imbalance Not fault tolerant HTTP redirection approach Heavy load on server-side Dispatcher at server-side Finer control of load distribution Berkeley MagicRouter for NOW project (1996) Cisco LocalDirector (1997) IBM’s Network Dispatcher IBM’s TCP Router (1996)

3. Chang Kyu Lee, Dept. of EECS, KAIST. 3 DNS-based approach Round-Robin DNS Service Significant Load Imbalance (DNS Cache) Not Fault Tolerant Server1 Server2 Server3 Client A Client B Client C DNS servers DNS server www.yahoo.com 204.71.200.74 204.71.200.75 204.71.200.67 204.71.200.74 204.71.200.75 204.71.200.67

4. Chang Kyu Lee, Dept. of EECS, KAIST. 4 NAT-based Approach (1) NAT (Network Address Translation)     S = C:A D = P:80 S = C:A D = B:81 S = B:81 D = C:A S = P:80 D = C:A private subnet client C:A Dispatcher P:80 server 1 D:81 server 2 B:81

5. Chang Kyu Lee, Dept. of EECS, KAIST. 5 NAT-based Approach (2) Bottle-neck at dispatcher MagicRouter by Berkeley NOW Local Director by Cisco VS-NAT by Linux Virtual Server Project Server Magic Router Server Client A Client B Client C

6. Chang Kyu Lee, Dept. of EECS, KAIST. 6 More Advanced Architecture Server1 TCP Router Server2 Server3 Client A Client B Client C Little overhead at dispatcher TCP Router by IBM VS-NAT/VS-DR by Linux Virtual Server Project

7. Chang Kyu Lee, Dept. of EECS, KAIST. 7 Distributed Packe t Rewriting Goals Transparency Scalability Efficiency (small overhead solution) Graceful Degradation (Fault Tolerant) Connection Assignment Flexibility

8. Chang Kyu Lee, Dept. of EECS, KAIST. 8 Architecture for DPR Server1 Server2 Server3 Server4 Client A Client B Client C All servers are routers, too. Distributed requests by RR-DNS

9. Chang Kyu Lee, Dept. of EECS, KAIST. 9 Design Issues Stateless vs. Stateful Stateless Approach Based on client’s address Client’s IP/Port  hashing  server selection Stateful Approach Based on the current server loads Address Translation Techniques NAT-based approach MAC-based or IPIP-based approach

10. Chang Kyu Lee, Dept. of EECS, KAIST. 10 Performance (1) Overhead test Environment Hardware : 200 MHz Pentium Pro, 32 MB memory, SCSI HDD, 100 Mbps Ethernet Switch Server : Linux 2.0.30, ext2 filesystem, Apache 1.2.4 Client : Windows NT 4.0, SPECWeb96

11. Chang Kyu Lee, Dept. of EECS, KAIST. 11 Results (1-1)

12. Chang Kyu Lee, Dept. of EECS, KAIST. 12 Results (1-2)

13. Chang Kyu Lee, Dept. of EECS, KAIST. 13 Performance (2) Scalability test (1~5 servers) Environment Hardware for server : 200 MHz Pentium Pro, 64 MB memory, IDE HDD, 100 Mbps Ethernet Switch Hardware for central router : 266 MHz Pentium II Server : Linux 2.0.30, Apache 1.2.1 Client : Surge Load Generator

14. Chang Kyu Lee, Dept. of EECS, KAIST. 14 Results (2-1)

15. Chang Kyu Lee, Dept. of EECS, KAIST. 15 Results (2-2)

16. Chang Kyu Lee, Dept. of EECS, KAIST. 16 Conclusion Critique Is there anything new? Load metric? Fault Tolerant? Server kernel modification