Shellac A distributed web accelerator. Kalan MacRow Data at Scale ‘13 

Shellac. (noun) Lac bug resin melted into thin flakes, used for making varnish Kalan MacRow Data at Scale ‘13 

Shellac. (verb) To defeat or beat (someone) decisively Kalan MacRow Data at Scale ‘13 

5.2% of the top 10,000 sites. Kalan MacRow Data at Scale ‘13 

Kalan MacRow Data at Scale ‘13 ! Script interpreters 50+ SQL queries Localization Personalization Template rendering …. ! 3 sec. 

• HTTP/1.1 proxy server for Linux 2.5+ • Distributed RAM cache built on Memcached • Level-triggered edge-polling architecture • 1K lines of Python Kalan MacRow Data at Scale ‘13 Design 

Kalan MacRow Data at Scale ‘13 “The key to making programs fast is to make them do practically nothing” — Mike Haertel, original author of GNU grep Design 

Kalan MacRow Data at Scale ‘13 • I/O bound and largely avoids the GIL • Substantial function/method/syscall call overhead • Iteration with while, for expensive • Event dispatch really suffers, but… Design 

Load Balancer Web Server Accelerator Cache Web Server Accelerator Cache Web Server Accelerator Cache Typical HA Kalan MacRow Data at Scale ‘13 

Shellac HA Kalan MacRow Data at Scale ‘13 Load Balancer Web Server Shellac Web Server Shellac Web Server Shellac Cache Dedicated Cache 

Kalan MacRow Data at Scale ‘13 • Interested in performance both • as a proxy / load balancer • as a caching proxy / accelerator • Graphs: best of Apache, Varnish and worst for Shellac over 3 runs Evaluation 

Kalan MacRow Data at Scale ‘13 • 4 x m1.large instances (4 core Xeon, 8GB) • Elastic Load Balancer • Master for running benchmarks • One availability zone, one region 

Kalan MacRow Data at Scale ‘13 Evaluation Apache Benchmark (-n 10k -c 1000) 

Evaluation Kalan MacRow Data at Scale ‘13 Apache Benchmark (-n 10k -c 1000) 

Evaluation Kalan MacRow Data at Scale ‘13 Apache Benchmark (-n 10k -c 1000) 

Kalan MacRow Data at Scale ‘13 Static Dynamic 1 Dynamic 2 Requests / sec. (mean) Evaluation 10K requests, 1K clients 

Static Dynamic 1 Dynamic 2 Transfer Rate (KB/s) Kalan MacRow Data at Scale ‘13 Evaluation 10K requests, 1K clients 

Static Dynamic 1 Dynamic 2 Peak Memory Usage (MB) Kalan MacRow Data at Scale ‘13 Evaluation 

Kalan MacRow Data at Scale ‘13 Conclusions • Shellac 0.1.0a is competitive • as a proxy / load balancer • as a caching proxy / accelerator • A C port might shellac Varnish! 

Future • HTTP/1.1 compliance • Try with PyPy JIT’ing, eventually port to C • Binary format for cache entries • Tighter integration of cache node with proxy Kalan MacRow Data at Scale ‘13 

https://github.com/kmacrow/Shellac Kalan MacRow Data at Scale ‘13