BigFloat improvements #136
Open
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
|
The errors seem to be related to MOI and unrelated to these changes. |
|
You can comment out the failing tests for now: #137 |
|
Locally all tests pass with the changes in #137 |
Use the `BigFloat` dot product from MutableArithmetics in HSD code. Helps with the performance of the `BigFloat` arithmetic. The change shouldn't affect other arithmetics, and it's coded so it'd be easy to extend it to another mutable arithmetic apart from just `BigFloat`, if necessary, and if such a type will support MutableArithmetics. Apart from improving performance, this change could possibly also benefit LP problems with numerical issues (when using `BigFloat`), because the MA dot product uses a summation algorithm that's more accurate than naive summation. A performance experiment is presented in the commit message of the following commit. The conclusion is that this commit improves performance only by a tiny bit, likewise with allocation.
We now do basically this for a logical vector l: `dot(a[l], b[l])`, instead of `dot(a .* l, b .* l) as before. This is as suggested here: ds4dm#122 (comment) Helps with the performance of the BigFloat arithmetic. A Julia script and a Unix shell script were used to conduct an experiment for assesing the impact of this commit and the previous commit on performance. The scripts and the resulting CSV file follow. The benchmark experiment is conducted and the CSV created by removing sources of system load on a computer and running the shell script four times: once with the `init_csv` command and three times with the `run` command, each time checking out a different commit in the Tulip git repo. Unix (Bourne) shell script: ```sh set -u command=$1 julia_opts='-O3 --min-optlevel=3 --heap-size-hint=5G --depwarn=error --warn-overwrite=yes' script=tulip_benchmark.jl case "$command" in init_csv) printf '%s,%s,%s,%s\n' 'Tulip version' estimator 'measurement type' value ;; run) tulip_version=$2 $PATH_TO_JULIA_BIN $julia_opts "$script" "$tulip_version" ;; *) printf '%s\n' error 2>&1 exit 1 ;; esac ``` Julia script: ```julia const benchmark_seconds = 500 const polynomial_degree = 20 setprecision(BigFloat, 12 * 2^7) using BenchmarkTools import FindMinimaxPolynomial, # v0.2.3 Tulip, MathOptInterface const FMP = FindMinimaxPolynomial const MMX = FMP.Minimax const PPTI = FMP.PolynomialPassingThroughIntervals const NE = FMP.NumericalErrorTypes const to_poly = FMP.ToSparsePolynomial.to_sparse_polynomial const mmx = MMX.minimax_polynomial const error_type_relative = NE.RelativeError() const MOI = MathOptInterface const itv_max_err = FMP.ApproximateInfinityNorm.interval_max_err function make_lp() lp = Tulip.Optimizer{BigFloat}() # Remove iteration limit just in case MOI.set(lp, MOI.RawOptimizerAttribute("IPM_IterationsLimit"), 2000) # Disable presolve, speeds things up #MOI.set(lp, MOI.RawOptimizerAttribute("Presolve_Level"), 0) lp end const itv = (-big"2.0"^-3, big"45.0") odd_monomials(n::Int) = 1:2:n sind_mmx(n::Int) = mmx( make_lp, sind, (itv,), odd_monomials(n), # Small factor to have less variance in the results initial_perturb_factor = 1//(2^20), # We're benchmarking LP, so disable other stuff worst_segments_density = 5, worst_segments_breadth_limit = 2, worst_segments_depth_ratio = 1/2, # Exit right after the first step exit_condition = true, ) function report(estimator; benchmark, benchmark_name) b = estimator(benchmark) println("$benchmark_name,$estimator,time,$(b.time)") println("$benchmark_name,$estimator,gctime,$(b.gctime)") println("$benchmark_name,$estimator,memory,$(b.memory)") println("$benchmark_name,$estimator,allocs,$(b.allocs)") end function report(;benchmark, benchmark_name) for quantile in (minimum, median, maximum) report( quantile, benchmark = benchmark, benchmark_name = benchmark_name, ) end end report( benchmark = (@benchmark sind_mmx(polynomial_degree) seconds=benchmark_seconds), benchmark_name = first(ARGS), ) ``` CSV results: ```csv Tulip version,estimator,measurement type,value v0.9.5,minimum,time,2.63759609e8 v0.9.5,minimum,gctime,3.4505713e7 v0.9.5,minimum,memory,495299296 v0.9.5,minimum,allocs,3629874 v0.9.5,median,time,4.2594161e8 v0.9.5,median,gctime,1.3250321e8 v0.9.5,median,memory,495299296 v0.9.5,median,allocs,3629874 v0.9.5,maximum,time,4.55935021e8 v0.9.5,maximum,gctime,1.40426286e8 v0.9.5,maximum,memory,495299296 v0.9.5,maximum,allocs,3629874 MutableArithmetics for IPM/HSD,minimum,time,2.57993117e8 MutableArithmetics for IPM/HSD,minimum,gctime,2.9403466e7 MutableArithmetics for IPM/HSD,minimum,memory,442052896 MutableArithmetics for IPM/HSD,minimum,allocs,3238720 MutableArithmetics for IPM/HSD,median,time,4.22323273e8 MutableArithmetics for IPM/HSD,median,gctime,1.282365305e8 MutableArithmetics for IPM/HSD,median,memory,442052896 MutableArithmetics for IPM/HSD,median,allocs,3238720 MutableArithmetics for IPM/HSD,maximum,time,4.56330849e8 MutableArithmetics for IPM/HSD,maximum,gctime,1.57061172e8 MutableArithmetics for IPM/HSD,maximum,memory,442052896 MutableArithmetics for IPM/HSD,maximum,allocs,3238720 IPM/HSD: use logical slicing ...,minimum,time,2.40996648e8 IPM/HSD: use logical slicing ...,minimum,gctime,2.5335783e7 IPM/HSD: use logical slicing ...,minimum,memory,386588512 IPM/HSD: use logical slicing ...,minimum,allocs,2833356 IPM/HSD: use logical slicing ...,median,time,3.76039574e8 IPM/HSD: use logical slicing ...,median,gctime,1.06930941e8 IPM/HSD: use logical slicing ...,median,memory,386588512 IPM/HSD: use logical slicing ...,median,allocs,2833356 IPM/HSD: use logical slicing ...,maximum,time,4.00347376e8 IPM/HSD: use logical slicing ...,maximum,gctime,1.27260987e8 IPM/HSD: use logical slicing ...,maximum,memory,386588512 IPM/HSD: use logical slicing ...,maximum,allocs,2833356 ``` Fixes ds4dm#122
|
Added the |
Codecov Report
@@ Coverage Diff @@
## master #136 +/- ##
==========================================
+ Coverage 89.02% 89.18% +0.16%
==========================================
Files 43 44 +1
Lines 2751 2783 +32
==========================================
+ Hits 2449 2482 +33
+ Misses 302 301 -1
... and 1 file with indirect coverage changes Help us with your feedback. Take ten seconds to tell us how you rate us. Have a feature suggestion? Share it here. |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
See the commit messages of the two commits