Performance

Performance measurements were taken using std::chrono::highresolution_clock, with overhead corrections. The code was compiled using the build options: variant = release, optimization = speed ^[7].

The columns labeled fiber (atomics) were compiled with default fiber synchronization, capable of synchronizing fibers running on different threads. The columns labeled fiber (raw) were compiled with BOOST_FIBERS_NO_ATOMICS.

Table 1.1. Overhead of join (contains fiber-context destruction, fiber-stack deallocation)

thread	fiber (atomics)	fiber (raw)	tbb	qthread
18 µs	950 ns	900 ns	570 ns	620 ns

(from overhead_join.cpp)

Table 1.2. Overhead of detach

thread	fiber (atomics)	fiber (raw)
126 ns	21 ns	20 ns

(from overhead_detach.cpp)

Table 1.3. Overhead of yield

thread	fiber (atomics)	fiber (raw)
1.5 µs	310 ns	330 ns

(from overhead_yield.cpp)

Table 1.4. Overhead of waiting on a future

thread	fiber (atomics)	fiber (raw)
16 µs	1.40 µs	1.38 µs

(from overhead_future.cpp)

Table 1.5. Overhead of fiber creation (contains fiber-stack allocation and preparation, fiber-context construction, scheduler handling)

thread	fiber (atomics)	fiber (raw)
18 µs	450 ns	445 ns

(from overhead_create.cpp)

Table 1.6. Scaling of creating and joining

average of	thread	fiber (atomics)	fiber (raw)
10	8.21 µs	1.96 µs	1.85 µs
50	6.67 µs	1.40 µs	1.27 µs
100	6.79 µs	1.84 µs	1.81 µs
500	8.25 µs	1.13 µs	1.10 µs
1000	7.71 µs	1.46 µs	1.26 µs
5000	5.67 µs	2.11 µs	1.90 µs
10000	5.25 µs	2.36 µs	1.89 µs

(from scale_join.cpp)

Numbers of the microbenchmark syknet from Alexander Temerev ^[8]:

Table 1.7. performance of N=100000 actors/goroutines/fibers

Haskell \| stack-1.0.4	fiber (single threaded/raw) \| gcc-5.2.1	fiber (single threaded/atomics) \| gcc-5.2.1	Erlang \| erts-7.0	Go \| go1.4.2
58ms - 108ms	205ms - 263ms	221ms - 278ms	237ms- 470ms	614ms - 883ms

^[7] Intel Core2 Q6700, x86_64, 3GHz

^[8] Intel Core2 Q6700, x86_64, 3GHz

Boost C++ Libraries

Performance