Performance Optimization

[pulsipher]

We need to analyze and optimize InfiniteOpt's structure to enhance performance. This might involve changes such as shifting to an array based structure instead of dictionaries. Also, the memory accessing should be better optimized to effectively leverage locality (i.e., use the cache memory effectively). We want to make it such that JuMP is the bottleneck.

[pulsipher]

It might be worth trying something like MOIU.CleverDicts.jl, but without ordered dictionaries. See #41

[pulsipher]

We can also make a number of improvements in transcribe.jl such as passing in whether an expression has a measure, enabling unordered builds, and more. See #42.

[pulsipher]

With the the complete overhaul brought on by #58 our performance is much better. Below is a comparison using the 3-node flexible design problem.

Version	# Samples	Time (s)	# Allocations	Memory
0.1.1	100	0.22	773.22 k	36.20 MiB
0.2.0	100	0.18	387.50 k	19.548 MiB
JuMP	100	0.57	1.05 M	52.56 MiB
0.1.1	1000	4.77	39.56 M	1.578 GiB
0.2.0	1000	0.36	1.20 M	58.673 MiB
JuMP	1000	0.60	1.38 M	75.36 MiB
0.1.1	5000	105.0	1.09 G	39.881 GiB
0.2.0	5000	1.19	4.85 M	234.74 MiB
JuMP	5000	1.09	2.90 M	179.10 MiB
0.1.1	10000	401.4	4.42 G	160.17 GiB
0.2.0	10000	2.65	9.42 M	445.25 MiB
JuMP	10000	1.877	4.79 M	301.18 MiB

Notably, now 0.2.0 seems to scale linearly with transcription size in stark contrast to the current release that just explodes. Moreover, the new version seems to be comparable to JuMP in terms of speed and memory (at least for this simple example).

[pulsipher]

In light of the above findings, overall performance as currently implemented is sufficient to close this general issue. In the future, we can/should open issues relating to specific performance improvements/problems.