Add IO benchmark job

config/gpu_configs/gpu_aquaplanet_diagedmf.yml

@@ -25,5 +25,5 @@ edmfx_sgs_mass_flux: true
 edmfx_sgs_diffusive_flux: true
 precip_model: 0M
 dt: 100secs 
-t_end: 12hours 
+t_end: 1days 
 toml: [toml/diagnostic_edmfx_box.toml]

config/gpu_configs/gpu_aquaplanet_dyamond_ss_1process.yml

@@ -18,6 +18,6 @@ dt_cloud_fraction: 1hours
 surface_setup: DefaultMoninObukhov
 rayleigh_sponge: true
 dt: 100secs
-t_end: 12hours
+t_end: 1days
 job_id: gpu_aquaplanet_dyamond_ss_1process
 toml: [toml/longrun_aquaplanet_dyamond.toml]

config/gpu_configs/gpu_aquaplanet_dyamond_ss_2process.yml

@@ -18,6 +18,6 @@ approximate_linear_solve_iters: 2
 surface_setup: "DefaultMoninObukhov"
 rayleigh_sponge: true
 dt: "100secs"
-t_end: "12hours"
+t_end: "1days"
 job_id: "gpu_aquaplanet_dyamond_ss_2process"
 toml: [toml/longrun_aquaplanet_dyamond.toml]

config/gpu_configs/gpu_aquaplanet_dyamond_ss_4process.yml

@@ -18,6 +18,6 @@ approximate_linear_solve_iters: 2
 surface_setup: "DefaultMoninObukhov"
 rayleigh_sponge: true
 dt: "100secs"
-t_end: "12hours"
+t_end: "1days"
 job_id: "gpu_aquaplanet_dyamond_ss_4process"
 toml: [toml/longrun_aquaplanet_dyamond.toml]

config/gpu_configs/gpu_aquaplanet_dyamond_ss_diag_1process.yml

@@ -0,0 +1,22 @@
+dt_save_state_to_disk: Inf
+dt_save_to_sol: Inf
+h_elem: 30
+z_max: 55000.0
+z_elem: 63
+dz_bottom: 30.0
+dz_top: 3000.0
+moist: equil
+precip_model: 1M
+rad: allskywithclear
+idealized_insolation: false
+dt_rad: 1hours
+vert_diff: FriersonDiffusion
+implicit_diffusion: true
+approximate_linear_solve_iters: 2
+dt_cloud_fraction: 1hours
+surface_setup: DefaultMoninObukhov
+rayleigh_sponge: true
+dt: 100secs
+t_end: 1days
+job_id: gpu_aquaplanet_dyamond_ss_diag_1process
+toml: [toml/longrun_aquaplanet_dyamond.toml]

config/gpu_configs/gpu_aquaplanet_dyamond_ws_1process.yml

@@ -18,6 +18,6 @@ approximate_linear_solve_iters: 2
 surface_setup: "DefaultMoninObukhov"
 rayleigh_sponge: true
 dt: "100secs"
-t_end: "12hours"
+t_end: "1days"
 job_id: "gpu_aquaplanet_dyamond_ws_1process"
 toml: [toml/longrun_aquaplanet_dyamond.toml]

config/gpu_configs/gpu_aquaplanet_dyamond_ws_2process.yml

@@ -18,6 +18,6 @@ approximate_linear_solve_iters: 2
 surface_setup: "DefaultMoninObukhov"
 rayleigh_sponge: true
 dt: "100secs"
-t_end: "12hours"
+t_end: "1days"
 job_id: "gpu_aquaplanet_dyamond_ws_2process"
 toml: [toml/longrun_aquaplanet_dyamond.toml]

config/gpu_configs/gpu_aquaplanet_dyamond_ws_4process.yml

@@ -18,6 +18,6 @@ approximate_linear_solve_iters: 2
 surface_setup: "DefaultMoninObukhov"
 rayleigh_sponge: true
 dt: "100secs"
-t_end: "12hours"
+t_end: "1days"
 job_id: "gpu_aquaplanet_dyamond_ws_4process"
 toml: [toml/longrun_aquaplanet_dyamond.toml]

perf/flame.jl

@@ -1,5 +1,6 @@
 redirect_stderr(IOContext(stderr, :stacktrace_types_limited => Ref(false)))
 import Random
+import PProf
 Random.seed!(1234)
 import ClimaAtmos as CA
 
@@ -28,6 +29,7 @@ mkpath(output_dir)
 Profile.clear()
 prof = Profile.@profile SciMLBase.step!(integrator)
 results = Profile.fetch()
+PProf.pprof(results, web=false)
 Profile.clear()
 
 ProfileCanvas.html_file(joinpath(output_dir, "flame.html"), results)
@@ -79,6 +81,7 @@ Profile.Allocs.@profile sample_rate = sampling_rate SciMLBase.step!(integrator)
 results = Profile.Allocs.fetch()
 Profile.Allocs.clear()
 profile = ProfileCanvas.view_allocs(results)
+PProf.Allocs.pprof(results, web = false)
 ProfileCanvas.html_file(joinpath(output_dir, "allocs.html"), profile)
 
 # We're grouping allocation tests here for convenience.

perf/longer_flame.jl

@@ -0,0 +1,43 @@
+redirect_stderr(IOContext(stderr, :stacktrace_types_limited => Ref(false)))
+import Random
+Random.seed!(1234)
+import ClimaAtmos as CA
+import Profile, ProfileCanvas
+import SciMLBase
+
+length(ARGS) != 1 && error("Usage: longer_flame.jl <config_file>")
+config_file = ARGS[1]
+config = CA.AtmosConfig(config_file)
+FT = eltype(config)
+# 1 day is a good compromise to go through most of the functions and diagnostics with almost
+# realistic cadence. For diagnostics, production runs will probably have a monthly output,
+# so this is a worst case to optimize for.
+config.parsed_args["t_end"] = "1days"
+simulation = CA.get_simulation(config)
+
+# First, we run the simulation for a few timesteps to compile most of the functions
+# (including top level function such as timed_solve!)
+
+simulation_dt = FT(CA.time_to_seconds(config.parsed_args["dt"]))
+# Change the final time to 5dt
+SciMLBase.reinit!(simulation.integrator, tf = 5simulation_dt)
+
+# We force the compilation of all callbacks to also include radiation
+CA.call_all_callbacks!(simulation.integrator)
+# Solve for a little bit
+CA.timed_solve!(simulation.integrator)
+
+# Change the final time to 1day
+SciMLBase.add_tstop!(simulation.integrator, FT(86400))
+
+@info "$simulation"
+
+@info "Collecting profile"
+Profile.init(n = 10^7, delay = 0.1)
+prof = Profile.@profile CA.timed_solve!(simulation.integrator)
+results = Profile.fetch()
+
+flame_path = joinpath(simulation.output_dir, "flame.html")
+
+ProfileCanvas.html_file(flame_path, results)
+@info "Flame saved in $flame_path"