Performance and benchmarking
ConsensusBasedX.jl has been developed with performance in mind. As such, it follows the recommended code patterns, such as avoiding global variables, or keeping memory allocation outside of performance-critical functions.
In order to maintain good performance, it's important that your function f also follows these patterns. We recommend the performance tips section of the Julia documentation. You should benchmark f, paying attention to memory allocations.
You should use the BenchmarkTools.jl package to benchmark f.
Once you have benchmarked f, you might want to test its performance within ConsensusBasedX.jl. You could run
config = (; D = 2)
@time minimise(f, config)but the output of @time will be noisy, as it will include the time and memory allocations due to the caches created by ConsensusBasedX.jl.
If, instead, you want to test only the performance-critical parts of the code, you can run the routine in benchmark mode:
config = (; D = 2, benchmark = true)
minimise(f, config)This will run the beginning of the minimise routine as normal, creating the required caches. However, instead of computing the full particle evolution, it will only calculate a few steps, printing the output of @time to console, and returning the output of @timed.
The benchmark mode reports zero allocations with all the Example objectives provided by ConsensusBasedX.jl. Wherever possible, your function should also lead to zero allocations.
Full example
using ConsensusBasedX
f(x) = ConsensusBasedX.Ackley(x, shift = 1)
config = (; D = 2, benchmark = true)
minimise(f, config)If you are running Consensus-Based Sampling by calling sample, allocations might occur whenever the root = :SymmetricRoot is automatically selected (see Root-covariance types). To have zero allocations, you must run with the option root = :AsymmetricRoot. Nevertheless, despite the allocations, the root = :SymmetricRoot option offers better performance when N is large (roughly if N > 10 * D).