1. Cooperative Caching of Dynamic Content on a Distributed Web Server Vegard Holmedahl, Ben Smith, Tao Yang Speaker: SeungLak Choi, DB Lab., CS Dept.

2. Contents Introduction Access log analysis Design of the Swala Experiments Conclusions Critiques

3. Introduction (1/2) Web caching reduces mainly the network delay Processor bottleneck is also important in sites serving extensive requests to dynamic contents

4. Introduction (2/2) Swala Distributed Web server Cooperatively caches the results of CGI requests Alexandria Digital Library (ADL) system at UCSB Accessing multi-resolution images Spatial database queries

5. Access Log Analysis (1/2) There is significant potential for reducing response time by optimizing CGI Total # of requests69,337 # of CGI requests28,663 (41.3%) Total response time of CGI requests 97% of total response time

6. Access Log Analysis (2/2) By caching a few number of CGI results, achieve significant response time reduction Time threshold #long requests Total # repeats # uniq. repeats Time saved Saved % 0.5 sec.1515371643641595734.6 1.0 sec.758128991891324128.7 2.0 sec.672126731551284027.8 3.0 sec.662726661511281227.8

7. Design of the Swala (1/3) Module design HTTP Module Handling HTTP requests Cache module Updates local directory Send contents to other node Delete expired cache entries

8. Design of the Swala (2/3) Cache table consistency Intra-node consistency Avoiding corruption from simultaneous updates Locking granularity – directory, table, entry Inter-node consistency Weak consistency Broadcast update information to the other nodes False cache miss, false cache hit

9. Design of the Swala (3/3) Content consistency Weak consistency – TTL Reasonable for dynamic contents with infrequent updates Store the cache directory in main memory Store CGI results in disk System administrator must specify uncacheable CGI

10. Experiments (1/6) Environment Sun Ultra 1 143-MHz Sun Ultra 2 167-MHz 64 or 128 MB Connected by 100Mbps Ethernet WebStone

11. Experiments (2/6) Single-node performance and overhead of cache fetch 24 clients request nullcgi that does no work and produces <1000 bytes output

12. Experiments (3/6) Multi-node performance

13. Experiments (4/6) Cache insertion overhead Each unique request generates a cache miss

14. Experiments (5/6) Cache directory maintenance overhead Generate update messages to Swala

15. Experiments (6/6) A comparison between cooperative and stand-alone caching Cache size 2000 Cache size 20

16. Conclusions Caching CGI results on a cluster of workstations Significant potential for time saving through caching CGI results Overhead of cooperative cache management is small Cooperative caching of CGI results substantially improves the response time

17. Critiques Strengths Attack the important problem occurred in common web sites CGI programs don’t have to be rewritten Weaknesses Weak content consistency Remote cache fetch can incur network congestion in heavy network traffic Contents hashing of requests?