search

vsbuffalo / granges

A Rust library and command line tool for working with genomic ranges and their data.
90 stars 5 forks source link

Slow numerics in operations #4

Open vsbuffalo opened 4 months ago

vsbuffalo commented 4 months ago

On large (1M ranges) datasets, bedtools map is beating GRanges. It look like it comes down to numerics:

# using inline: 
command       bedtools time    granges time      granges speedup (%)
------------  ---------------  --------------  ---------------------
map_multiple  26.89 min        688.60 s                     57.3189
map_max       21.34 min        21.54 min                    -0.94889
adjust        21.35 min        499.54 s                     60.9959
filter        27.01 min        20.38 min                    24.5583
map_min       935.69 s         17.48 min                   -12.1008
flank         30.26 min        943.36 s                     48.0353
map_mean      20.13 min        931.21 s                     22.913
map_sum       17.99 min        18.44 min                    -2.5225
windows       507.84 s         84.79 s                      83.304
map_median    16.90 min        17.81 min                    -5.36811

# not using inline
command       bedtools time    granges time      granges speedup (%)
------------  ---------------  --------------  ---------------------
map_multiple  26.89 min        688.60 s                     57.3189
map_max       21.34 min        21.54 min                    -0.94889
adjust        21.35 min        499.54 s                     60.9959
filter        27.01 min        20.38 min                    24.5583
map_min       935.69 s         17.48 min                   -12.1008
flank         30.26 min        943.36 s                     48.0353
map_mean      20.13 min        931.21 s                     22.913
map_sum       17.99 min        18.44 min                    -2.5225
windows       507.84 s         84.79 s                      83.304
map_median    16.90 min        17.81 min                    -5.36811

where inline here refers to #[inline(always)] above the Operations::run() (suggested by @molpopgen).

vsbuffalo commented 4 months ago

So interestingly this isn't an issue on Apple silicon: 

command       bedtools time    granges time      granges speedup (%)
------------  ---------------  --------------  ---------------------
map_multiple  133.10 s         67.61 s                       49.2089
adjust        63.30 s          33.18 s                       47.5765
map_min       414.12 s         291.08 s                      29.7102
map_mean      410.42 s         276.49 s                      32.6327
map_max       417.44 s         281.10 s                      32.6611
map_sum       425.98 s         276.15 s                      35.1739
map_median    419.66 s         294.14 s                      29.9112
flank         86.00 s          51.95 s                       39.5932
filter        78.36 s          43.85 s                       44.0442
windows       287.13 s         49.90 s                       82.6225

vs Ubuntu x86_64

command     bedtools time    granges time      granges speedup (%)
----------  ---------------  --------------  ---------------------
map_max     530.72 s         571.12 s                     -7.61291
map_min     537.00 s         611.83 s                    -13.9352
map_mean    534.22 s         572.59 s                     -7.18139
map_sum     561.18 s         601.37 s                     -7.16121
map_median  570.75 s         549.84 s                      3.6649
molpopgen commented 4 months ago

So interestingly this isn't an issue on Apple silicon:

command       bedtools time    granges time      granges speedup (%)
------------  ---------------  --------------  ---------------------
map_multiple  133.10 s         67.61 s                       49.2089
adjust        63.30 s          33.18 s                       47.5765
map_min       414.12 s         291.08 s                      29.7102
map_mean      410.42 s         276.49 s                      32.6327
map_max       417.44 s         281.10 s                      32.6611
map_sum       425.98 s         276.15 s                      35.1739
map_median    419.66 s         294.14 s                      29.9112
flank         86.00 s          51.95 s                       39.5932
filter        78.36 s          43.85 s                       44.0442
windows       287.13 s         49.90 s                       82.6225

vs Ubuntu x86_64

command     bedtools time    granges time      granges speedup (%)
----------  ---------------  --------------  ---------------------
map_max     530.72 s         571.12 s                     -7.61291
map_min     537.00 s         611.83 s                    -13.9352
map_mean    534.22 s         572.59 s                     -7.18139
map_sum     561.18 s         601.37 s                     -7.16121
map_median  570.75 s         549.84 s                      3.6649

Is this "apples to apples" ? (Pun intended.) The raw times reported for Apple Silicon are 10X less than for Ubuntu/x86. Are you sure that the Apple numbers reflect "large data" tests?

molpopgen commented 4 months ago

On a second read, maybe the Apple Silicon numbers are okay? They're not 10x smaller but more like 2x?