search

utahplt / TrackedFloats.jl

Julia library providing tracking of floating point errors through a program resources
https://juliahub.com/ui/Packages/General/TrackedFloats
MIT License
34 stars 3 forks source link

Improve performance of logging #19

Open ashton314 opened 1 year ago

ashton314 commented 1 year ago

Working on this code in FloatTrackerExamples:

using FloatTracker: TrackedFloat16, TrackedFloat32, write_out_logs, set_inject_nan, set_exclude_stacktrace, set_logger
using ShallowWaters, PyPlot

set_logger(filename="pretty_nan_tf", buffersize=100000, cstg=true, cstgArgs=true, cstgLineNum=true)
set_exclude_stacktrace([:prop,:gen]) # Props and kills take the lion share of the perf hit
# set_exclude_stacktrace([:prop,:kill,:gen])

P = run_model(T=TrackedFloat32,
              nx=100, L_ratio=1, bc="nonperiodic", wind_forcing_x="double_gyre",
              topography="seamount",cfl=2,Ndays=300)

pcolormesh(P.η')
savefig("height_pretty_nan_tf.png")

speed = sqrt.(Ix(P.u.^2)[:,2:end-1] + Iy(P.v.^2)[2:end-1,:])
pcolormesh(speed')
savefig("speed_pretty_nan_tf.png")

Dialing the buffer size up really high reduced runtime from 65s → 50s. Turning off kill logs dropped the runtime to just 14s. This suggests to me that part of where FloatTracker gets slow is writing (or just formatting?) the logs. I bet we could easily improve the performance here.

bennn commented 1 year ago

NB: setting maxLogs (#28 ) is a quick way to improve