rubydragon/metagraphoptimization.jl
1
0
Fork 0
Code Issues 6 Pull Requests Actions Packages Projects Releases Wiki Activity

Add qed bench run script and singularity container

Browse Source
This commit is contained in:
Anton Reinhard
2024-02-02 06:56:16 +01:00
parent 82ed774b7e
commit 7098d1801a
6 changed files with 415 additions and 51 deletions
Download Patch File Download Diff File Expand all files Collapse all files

5
examples/import_bench.jl
View File

@ -34,9 +34,10 @@ function import_bench()
bench_txt("AB->ABBB.txt")
bench_txt("AB->ABBBBB.txt")
bench_txt("AB->ABBBBBBB.txt")
#bench_txt("AB->ABBBBBBBBB.txt")
bench_txt("AB->ABBBBBBBBB.txt")
bench_txt("ABAB->ABAB.txt")
return bench_txt("ABAB->ABC.txt")
bench_txt("ABAB->ABC.txt")
return nothing
end
import_bench()

281
examples/qed_bench.jl
View File

@ -2,6 +2,40 @@ using MetagraphOptimization
using LIKWID
using CUDA
using UUIDs
using DataFrames
using CSV
using Random
DISABLE_GPU = false
results_filename = "results.csv"
df = DataFrame(
process_name = String[],
graph_gen_time = Float64[],
optimization_time = Float64[],
function_generation_time = Float64[],
graph_nodes = Int[],
graph_edges = Int[],
graph_mem = Float64[],
cpu_threads = Int[],
n_inputs = Int[],
nflops_likwid = Int[],
cpu_time = Float64[],
cpu_rate = Float64[],
cpu_gflops = Float64[],
gpu_name = String[],
gpu_time = Float64[],
gpu_rate = Float64[],
gpu_gflops = Float64[],
)
# if they exist, read existing results and append new ones
if isfile(results_filename)
df = CSV.read(results_filename, DataFrame)
end
nInputs = 100_000
function cpu_bench(compute_function, inputs)
compute_function.(inputs[begin:10]) # make sure it's compiled
@ -24,21 +58,38 @@ end
function bench_process(
process::MetagraphOptimization.AbstractProcessDescription,
process_name::String,
graph::DAG,
func,
gen_time::Float64,
opt_time::Float64,
func_time::Float64,
io::IO = stdout;
use_likwid = true,
use_gpu = true,
)
println(io, "\n--- Benchmarking $(process) ---")
println(io, "\n--- Benchmarking $(process_name) ---")
if DISABLE_GPU
use_gpu = false
end
NFLOPs = GraphProperties(graph).computeEffort
graph_props = GraphProperties(graph)
NFLOPs = graph_props.computeEffort
nflops_likwid = 0
if use_likwid
input = gen_process_input(process)
func(input) # compile first
# get rid of annoying output to console
oldstd = stdout
redirect_stdout(devnull)
_, events = @perfmon "FLOPS_DP" func(input)
redirect_stdout(oldstd) # recover original stdout
NFLOPs = first(events["FLOPS_DP"])["RETIRED_SSE_AVX_FLOPS_ALL"]
nflops_likwid = NFLOPs
end
nInputs = 10000000 # ten million
println(io, "Generating $nInputs inputs with $(Threads.nthreads()) threads...")
inputs = Vector{typeof(gen_process_input(process))}()
@ -54,12 +105,21 @@ function bench_process(
println(io, "Benchmarking CPU with $(Threads.nthreads()) threads...")
(time_cpu, rate_cpu) = cpu_bench(func, inputs)
flops_cpu = (rate_cpu * NFLOPs) / 1024^3
flops_cpu = (rate_cpu * NFLOPs) / 10^9
println(io, "Benchmarking GPU...")
cuInputs = CuArray(inputs)
(time_gpu, rate_gpu) = gpu_bench(func, cuInputs)
flops_gpu = (rate_gpu * NFLOPs) / 1024^3
time_gpu = 0.0
rate_gpu = 0.0
flops_gpu = 0.0
gpu_name = "none"
if use_gpu
println(io, "Benchmarking GPU...")
gpu_name = "$(name(first(CUDA.devices())))"
cuInputs = CuArray(inputs)
(time_gpu, rate_gpu) = gpu_bench(func, cuInputs)
flops_gpu = (rate_gpu * NFLOPs) / 10^9
else
println(io, "Skipping GPU...")
end
println(io, "\nBenchmark Summary for $(process):")
@ -73,10 +133,39 @@ function bench_process(
println(io, " Time: $time_cpu")
println(io, " Rate: $rate_cpu")
println(io, " GFLOPS: $flops_cpu")
println(io, "GPU, $(name(first(CUDA.devices())))")
println(io, " Time: $time_gpu")
println(io, " Rate: $rate_gpu")
return println(io, " GFLOPS: $flops_gpu")
if use_gpu
println(io, "GPU, $gpu_name")
println(io, " Time: $time_gpu")
println(io, " Rate: $rate_gpu")
println(io, " GFLOPS: $flops_gpu")
end
if (process_name != "warmup")
push!(
df,
Dict(
:process_name => process_name,
:graph_gen_time => gen_time,
:optimization_time => opt_time,
:function_generation_time => func_time,
:graph_nodes => graph_props.noNodes,
:graph_edges => graph_props.noEdges,
:graph_mem => MetagraphOptimization.mem(graph),
:cpu_threads => Threads.nthreads(),
:n_inputs => nInputs,
:nflops_likwid => nflops_likwid,
:cpu_time => time_cpu,
:cpu_rate => rate_cpu,
:cpu_gflops => flops_cpu,
:gpu_name => gpu_name,
:gpu_time => time_gpu,
:gpu_rate => rate_gpu,
:gpu_gflops => flops_gpu,
),
)
end
return nothing
end
# use "mock" machine that only uses cpu
@ -92,57 +181,151 @@ machine = Machine(
],
[-1.0;;],
)
optimizer = ReductionOptimizer()
# sadly cannot put these in functions because the world age must increase after the function is created which happens only in the global scope
# compton
process = parse_process("ke->ke", QEDModel())
graph = gen_graph(process)
optimize_to_fixpoint!(optimizer, graph)
compute_func = get_compute_function(graph, process, machine)
bench_process(process, compute_func)
## -- WARMUP TO COMPILE FUNCTIONS first
optimizer = RandomWalkOptimizer(MersenneTwister(0))
# 2-photon compton
process = parse_process("ke->kke", QEDModel())
graph = gen_graph(process)
optimize_to_fixpoint!(optimizer, graph)
compute_func = get_compute_function(graph, process, machine)
bench_process(process, compute_func)
gen_time = @elapsed graph = gen_graph(process)
opt_time = @elapsed optimize!(optimizer, graph, 200)
func_gen_time = @elapsed compute_func = get_compute_function(graph, process, machine)
bench_process(process, "warmup", graph, compute_func, gen_time, opt_time, func_gen_time)
# 3-photon compton
process = parse_process("ke->kkke", QEDModel())
graph = gen_graph(process)
optimize_to_fixpoint!(optimizer, graph)
compute_func = get_compute_function(graph, process, machine)
bench_process(process, compute_func)
# AB->AB
process = parse_process("AB->AB", ABCModel())
graph = parse_dag("input/AB->AB.txt", ABCModel())
optimize_to_fixpoint!(optimizer, graph)
compute_func = get_compute_function(graph, process, machine)
bench_process(process, compute_func)
optimizer = ReductionOptimizer()
# AB->AB^3
process = parse_process("AB->ABBB", ABCModel())
graph = parse_dag("input/AB->ABBB.txt", ABCModel())
optimize_to_fixpoint!(optimizer, graph)
compute_func = get_compute_function(graph, process, machine)
bench_process(process, compute_func)
gen_time = @elapsed graph = parse_dag("input/AB->ABBB.txt", ABCModel())
opt_time = @elapsed optimize_to_fixpoint!(optimizer, graph)
func_gen_time = @elapsed compute_func = get_compute_function(graph, process, machine)
bench_process(process, "warmup", graph, compute_func, gen_time, opt_time, func_gen_time)
exit(0)
## -- WARMUP END
optimizer = ReductionOptimizer()
# compton
process = parse_process("ke->ke", QEDModel())
gen_time = @elapsed graph = gen_graph(process)
func_gen_time = @elapsed compute_func = get_compute_function(graph, process, machine)
bench_process(process, "$process no optimization", graph, compute_func, gen_time, 0.0, func_gen_time)
opt_time = @elapsed optimize_to_fixpoint!(optimizer, graph)
func_gen_time = @elapsed compute_func = get_compute_function(graph, process, machine)
bench_process(process, "$process reduced", graph, compute_func, gen_time, opt_time, func_gen_time)
# 2-photon compton
process = parse_process("ke->kke", QEDModel())
gen_time = @elapsed graph = gen_graph(process)
func_gen_time = @elapsed compute_func = get_compute_function(graph, process, machine)
bench_process(process, "$process no optimization", graph, compute_func, gen_time, 0.0, func_gen_time)
opt_time = @elapsed optimize_to_fixpoint!(optimizer, graph)
func_gen_time = @elapsed compute_func = get_compute_function(graph, process, machine)
bench_process(process, "$process reduced", graph, compute_func, gen_time, opt_time, func_gen_time)
# 3-photon compton
process = parse_process("ke->kkke", QEDModel())
gen_time = @elapsed graph = gen_graph(process)
func_gen_time = @elapsed compute_func = get_compute_function(graph, process, machine)
bench_process(process, "$process no optimization", graph, compute_func, gen_time, 0.0, func_gen_time)
opt_time = @elapsed optimize_to_fixpoint!(optimizer, graph)
func_gen_time = @elapsed compute_func = get_compute_function(graph, process, machine)
bench_process(process, "$process reduced", graph, compute_func, gen_time, opt_time, func_gen_time)
# 4-photon compton
process = parse_process("ke->kkkke", QEDModel())
graph = gen_graph(process)
optimize_to_fixpoint!(optimizer, graph)
compute_func = get_compute_function(graph, process, machine)
bench_process(process, compute_func)
gen_time = @elapsed graph = gen_graph(process)
func_gen_time = @elapsed compute_func = get_compute_function(graph, process, machine)
bench_process(process, "$process no optimization", graph, compute_func, gen_time, 0.0, func_gen_time, use_gpu = false)
opt_time = @elapsed optimize_to_fixpoint!(optimizer, graph)
func_gen_time = @elapsed compute_func = get_compute_function(graph, process, machine)
bench_process(process, "$process reduced", graph, compute_func, gen_time, opt_time, func_gen_time)
# 5-photon compton
process = parse_process("ke->kkkkke", QEDModel())
gen_time = @elapsed graph = gen_graph(process)
func_gen_time = @elapsed compute_func = get_compute_function(graph, process, machine)
bench_process(process, "$process no optimization", graph, compute_func, gen_time, 0.0, func_gen_time, use_gpu = false)
opt_time = @elapsed optimize_to_fixpoint!(optimizer, graph)
func_gen_time = @elapsed compute_func = get_compute_function(graph, process, machine)
bench_process(process, "$process reduced", graph, compute_func, gen_time, opt_time, func_gen_time, use_gpu = false)
# 6-photon compton
process = parse_process("ke->kkkkkke", QEDModel())
gen_time = @elapsed graph = gen_graph(process)
func_gen_time = @elapsed compute_func = get_compute_function(graph, process, machine)
bench_process(process, "$process no optimization", graph, compute_func, gen_time, 0.0, func_gen_time, use_gpu = false)
opt_time = @elapsed optimize_to_fixpoint!(optimizer, graph)
func_gen_time = @elapsed compute_func = get_compute_function(graph, process, machine)
bench_process(process, "$process reduced", graph, compute_func, gen_time, opt_time, func_gen_time, use_gpu = false)
# 7-photon compton
process = parse_process("ke->kkkkkkke", QEDModel())
gen_time = @elapsed graph = gen_graph(process)
func_gen_time = @elapsed compute_func = get_compute_function(graph, process, machine)
bench_process(process, "$process no optimization", graph, compute_func, gen_time, 0.0, func_gen_time, use_gpu = false)
opt_time = @elapsed optimize_to_fixpoint!(optimizer, graph)
func_gen_time = @elapsed compute_func = get_compute_function(graph, process, machine)
bench_process(process, "$process reduced", graph, compute_func, gen_time, opt_time, func_gen_time, use_gpu = false)
# 8-photon compton
process = parse_process("ke->kkkkkkkke", QEDModel())
gen_time = @elapsed graph = gen_graph(process)
func_gen_time = @elapsed compute_func = get_compute_function(graph, process, machine)
bench_process(process, "$process no optimization", graph, compute_func, gen_time, 0.0, func_gen_time, use_gpu = false)
opt_time = @elapsed optimize_to_fixpoint!(optimizer, graph)
func_gen_time = @elapsed compute_func = get_compute_function(graph, process, machine)
bench_process(process, "$process reduced", graph, compute_func, gen_time, opt_time, func_gen_time, use_gpu = false)
# AB->AB
process = parse_process("AB->AB", ABCModel())
gen_time = @elapsed graph = parse_dag("input/AB->AB.txt", ABCModel())
func_gen_time = @elapsed compute_func = get_compute_function(graph, process, machine)
bench_process(process, "$process no optimization", graph, compute_func, gen_time, 0.0, func_gen_time)
opt_time = @elapsed optimize_to_fixpoint!(optimizer, graph)
func_gen_time = @elapsed compute_func = get_compute_function(graph, process, machine)
bench_process(process, "$process reduced", graph, compute_func, gen_time, opt_time, func_gen_time)
# AB->AB^3
process = parse_process("AB->ABBB", ABCModel())
gen_time = @elapsed graph = parse_dag("input/AB->ABBB.txt", ABCModel())
func_gen_time = @elapsed compute_func = get_compute_function(graph, process, machine)
bench_process(process, "$process no optimization", graph, compute_func, gen_time, 0.0, func_gen_time)
opt_time = @elapsed optimize_to_fixpoint!(optimizer, graph)
func_gen_time = @elapsed compute_func = get_compute_function(graph, process, machine)
bench_process(process, "$process reduced", graph, compute_func, gen_time, opt_time, func_gen_time)
# AB->AB^5
process = parse_process("AB->ABBBBB", ABCModel())
graph = parse_dag("input/AB->ABBBBB.txt", ABCModel())
optimize_to_fixpoint!(optimizer, graph)
compute_func = get_compute_function(graph, process, machine)
bench_process(process, compute_func)
gen_time = @elapsed graph = parse_dag("input/AB->ABBBBB.txt", ABCModel())
func_gen_time = @elapsed compute_func = get_compute_function(graph, process, machine)
bench_process(process, "$process no optimization", graph, compute_func, gen_time, 0.0, func_gen_time, use_gpu = false)
opt_time = @elapsed optimize_to_fixpoint!(optimizer, graph)
func_gen_time = @elapsed compute_func = get_compute_function(graph, process, machine)
bench_process(process, "$process reduced", graph, compute_func, gen_time, opt_time, func_gen_time, use_gpu = false)
# AB->AB^7
process = parse_process("AB->ABBBBBBB", ABCModel())
gen_time = @elapsed graph = parse_dag("input/AB->ABBBBBBB.txt", ABCModel())
func_gen_time = @elapsed compute_func = get_compute_function(graph, process, machine)
bench_process(process, "$process no optimization", graph, compute_func, gen_time, 0.0, func_gen_time, use_gpu = false)
opt_time = @elapsed optimize_to_fixpoint!(optimizer, graph)
func_gen_time = @elapsed compute_func = get_compute_function(graph, process, machine)
bench_process(process, "$process reduced", graph, compute_func, gen_time, opt_time, func_gen_time, use_gpu = false)
CSV.write(results_filename, df)