1. On the Performance of TCP Splicing for URL-aware Redirection Ariel Cohen, Sampath Rangarajan, and Hamilton Slye The 2 nd USENIX Symposium on Internet Technologies & Systems Network Computing Lab, EECS, KAIST Hyun-ju Lee

2. 2001-12-11Network Computing Lab.2 Content Introduction Solution approach Implementation detail Performance Conclusion

3. 2001-12-11Network Computing Lab.3 Introduction(1) Overview  Describe the design, implementation and performance of a layer-7 switch which support URL-aware redirection of HTTP traffic Solution approach  URL-aware switch in the Linux kernel that uses TCP splicing Contribution  Discuss in detail the design and implementation of URL- aware switch  Show the benefit of TCP splicing for small TCP sessions  Re-affirm the benefit of TCP splicing for transform of large amount of data

4. 2001-12-11Network Computing Lab.4 Introduction(2) Basic layer-4 switch  Redirect traffic based on transport-layer(TCP) and network- layer(IP) information  Uses : redirecting web traffic to caches, server load balancing, fault tolerance URL-aware redirection  Capability of a switch located in front of clients or servers to redirect HTTP requests to servers based on the URL specified by the client in its GET request  Extend the scope of information to layer-7 information  Benefit : direct requests to optimized server, direct requests for dynamic content to live servers, reduce the need for replication

5. 2001-12-11Network Computing Lab.5 Introduction(3) Layer-4 switch  Redirect the initial SYN packet from the client to the choosen destination  Redirect all subsequent packets on the connection to the same destination  Use mechanisms such as NAT, PAT URL-aware redirection  Connection must be established between a client and a switch before application-level information can be received  Approaches –TCP gateway –TCP Connection hop

6. 2001-12-11Network Computing Lab.6 Introduction(4) TCP Connection hop  Move the switch side end point of client-switch TCP connection to the server  Establish a direct TCP connection between the client and the server  TCP/IP stack at the back-end server needs to be extended TCP gateway  Two TCP connections –Establish client-switch TCP connection –Receive GET request from client, Determine destination –Establish switch-server TCP connection –Forward GET request to the destination, receive the response –Transfer the response to the client

7. 2001-12-11Network Computing Lab.7 Solution approach TCP gateway  Overhead due to the use of two TCP connection  Go up the protocol stack to the application layer and down the protocol stack TCP splicing  Once the two TCP connections are established, they are spliced together  IP packets are forwarded from one end point to the other at the network layer  Address translation and sequence number modification is required

8. 2001-12-11Network Computing Lab.8 Implementation detail(1) Application-level proxy(proxy-s) Splice module(sp-mod) NEPPI(Network Element for Programmable Packet Injector) Linux ipchains firewall

9. 2001-12-11Network Computing Lab.9 Implementation detail(2)  Establish client-URL- aware proxy TCP connection  URL-aware proxy :Receive GET request from client  sp-mod : record seq, ack_seq  Determine destination  Splice commend to sp-mod  Send GET request to server  Sp-mod : record seq, ack_seq  Splice instruction to NEPPI  NEPPI : address, sequence # translation  All packet are redirected at network layer

10. 2001-12-11Network Computing Lab.10 Performance(1) Experiments  run for a period of three minutes with a concurrent HTTP request of 75  Once all the data is received, a new batch of requests is issued  proxy-s : TCP splicing proxy-ns : do not utilize TCP splicing Internal / external workload  Proxy-s : significantly higher number of connections(47% higher) & lower CPU utilization

11. 2001-12-11Network Computing Lab.11 Performance(2) Impact of using TCP splicing on different file size  The biggest performance gain for large transfers  One might expect to obtain larger number of connections for smaller file (CPU was not fully utilized)

12. 2001-12-11Network Computing Lab.12 Conclusion URL-aware redirection switch  Loadable kernel modules for the Linux OS along with a user-level proxy TCP splicing  Significant performance improvement