A standard Bullet cache benchmark investigates the influence of increasing record size on the cache server performance. As the keys are of the same structure, the performance is largely influenced by the efficiency of the network protocol and internal storage methods.
The benchmark is performed (and thus can be reproduced by interested parties) under the following conditions:
- OS: Xeon 3440, 4 GB RAM, FreeBSD 8.2-stable, AMD64
- Server started with the following arguments: ./mdcached -b4 -t0
- Benchmark client started on the same system (communicating with the server over the Unix sockets mechanism) with the following arguments: ./bench_mdcached_async -t 5 -n 60 -s #, where the "#" symbol stands for the benchmark scale argument which is one of the following, in sequence: 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 15, 20, 30, 40. The scale argument determines the record data sizes used in the benchmark.
The latest results of this benchmark are graphed in the following image:
If you have any more benchmark results which are obtained by closely following steps 2 and 3 in the above description, please send them to the Bullet cache project!