Quick benchmark on the bounded_buffer<T> example from Boost Templated Circular Buffer

Boost's Templated Circular Buffer documentation claims that some modest Boost.Thread wrapping of a boost::circular_buffer<T> makes a bounded producer-consumer queue. Happily, they provide a benchmark. I believe 'em, of course, but I'd like to get an idea of the variability behind the claims.

Rusty Russell, of CCAN fame, put out a nice pipe-oriented processor for gathering statistics. This can directly be fed the output of the benchmark as a pretty cute one-liner:

gcc -Wall -O3 -pthread bounded_buffer_comparison.cpp -lboost_thread-mt
for i in `seq 1 100`; do ./a.out; done | ./stats

With a touch of reformatting and sorting the columns by increasing mean, here are the results on a circa-2007 T61 stinkpad of mine:

bounded_buffer<int>                         0.45 -- 0.98  (0.7588 +/- 0.17 ) s
bounded_buffer<std::string>                 0.58 -- 0.91  (0.7678 +/- 0.095) s
bounded_buffer_deque_based<std::string>     0.43 -- 0.94  (0.8291 +/- 0.095) s
bounded_buffer_deque_based<int>             0.48 -- 1.03  (0.8374 +/- 0.12 ) s
bounded_buffer_space_optimized<int>         0.55 -- 1.10  (0.8459 +/- 0.1  ) s
bounded_buffer_space_optimized<std::string> 0.46 -- 1.21  (0.8852 +/- 0.24 ) s
bounded_buffer_list_based<int>              2.84 -- 6.65  (4.2641 +/- 0.74 ) s
bounded_buffer_list_based<std::string>      2.22 -- 5.77  (4.2663 +/- 0.66 ) s

Building -O3 makes a big difference as you might expected and I've not bothered showing -O2. On average, the sample bounded_buffer<T> does slightly outperform the deque-based logic as the Boost Templated Circular Buffer documentation suggests. The variability is, uhhh, more variable than I'd expected.