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

Add evaluation script, run script, csv data and first plots

Browse Source
This commit is contained in:
Anton Reinhard
2024-02-06 09:35:04 +01:00
parent 7098d1801a
commit 3ac9954d32
42 changed files with 6725 additions and 36 deletions
Download Patch File Download Diff File Expand all files Collapse all files

35
examples/qed_bench.jl
View File

@ -6,7 +6,7 @@ using DataFrames
using CSV
using Random
DISABLE_GPU = false
DISABLE_GPU = true
results_filename = "results.csv"
@ -261,29 +261,6 @@ bench_process(process, "$process reduced", graph, compute_func, gen_time, opt_ti
# 6-photon compton
process = parse_process("ke->kkkkkke", QEDModel())
gen_time = @elapsed graph = gen_graph(process)
func_gen_time = @elapsed compute_func = get_compute_function(graph, process, machine)
bench_process(process, "$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, "$process reduced", graph, compute_func, gen_time, opt_time, func_gen_time, use_gpu = false)
@ -318,14 +295,4 @@ opt_time = @elapsed optimize_to_fixpoint!(optimizer, graph)
func_gen_time = @elapsed compute_func = get_compute_function(graph, process, machine)
bench_process(process, "$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)

209
examples/qed_bench_tape.jl Normal file
View File

@ -0,0 +1,209 @@
using MetagraphOptimization
using LIKWID
using UUIDs
using DataFrames
using CSV
using Random
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
# use "mock" machine that only uses cpu
machine = Machine(
[
MetagraphOptimization.NumaNode(
0,
1,
MetagraphOptimization.default_strategy(MetagraphOptimization.NumaNode),
-1.0,
UUIDs.uuid1(),
),
],
[-1.0;;],
)
function cpu_bench(tape, inputs)
time = @elapsed Threads.@threads for i in eachindex(inputs)
execute_tape(tape, inputs[i])
end
rate = length(inputs) / time
return (time, rate)
end
function bench_process(
process::MetagraphOptimization.AbstractProcessDescription,
process_name::String,
graph::DAG,
gen_time::Float64,
opt_time::Float64,
io::IO = stdout;
use_likwid = true,
)
println(io, "\n--- Benchmarking $(process_name) ---")
func_time = @elapsed tape = gen_tape(graph, process, machine)
graph_props = GraphProperties(graph)
NFLOPs = graph_props.computeEffort
nflops_likwid = 0
if use_likwid
input = gen_process_input(process)
# get rid of annoying output to console
oldstd = stdout
redirect_stdout(devnull)
_, events = @perfmon "FLOPS_DP" execute_tape(tape, input)
redirect_stdout(oldstd) # recover original stdout
NFLOPs = first(events["FLOPS_DP"])["RETIRED_SSE_AVX_FLOPS_ALL"]
nflops_likwid = NFLOPs
end
println(io, "Generating $nInputs inputs with $(Threads.nthreads()) threads...")
inputs = Vector{typeof(gen_process_input(process))}()
resize!(inputs, nInputs)
processes = Vector{typeof(process)}()
for i in 1:Threads.nthreads()
push!(processes, copy(process))
end
Threads.@threads for i in eachindex(inputs)
inputs[i] = gen_process_input(processes[Threads.nthreads()])
end
println(io, "Benchmarking CPU with $(Threads.nthreads()) threads...")
(time_cpu, rate_cpu) = cpu_bench(tape, inputs)
flops_cpu = (rate_cpu * NFLOPs) / 10^9
println(io, "\nBenchmark Summary for $(process):")
if use_likwid
println(io, "Measured FLOPS by LIKWID: $NFLOPs")
else
println(io, "Total graph compute effort: $NFLOPs")
end
println(io, "Total input size: $(bytes_to_human_readable(Base.summarysize(inputs)))")
println(io, "CPU, $(Threads.nthreads()) threads")
println(io, " Time: $time_cpu")
println(io, " Rate: $rate_cpu")
println(io, " GFLOPS: $flops_cpu")
if (process_name != "warmup")
push!(
df,
Dict(
:process_name => process_name,
:graph_gen_time => gen_time,
:optimization_time => opt_time,
:function_generation_time => func_time,
:graph_nodes => graph_props.noNodes,
:graph_edges => graph_props.noEdges,
:graph_mem => MetagraphOptimization.mem(graph),
:cpu_threads => Threads.nthreads(),
:n_inputs => nInputs,
:nflops_likwid => nflops_likwid,
:cpu_time => time_cpu,
:cpu_rate => rate_cpu,
:cpu_gflops => flops_cpu,
:gpu_name => "none",
:gpu_time => 0.0,
:gpu_rate => 0.0,
:gpu_gflops => 0.0,
),
)
end
return nothing
end
function bench_qed(process_string::String, skip_unoptimized = false)
optimizer = ReductionOptimizer()
process = parse_process(process_string, QEDModel())
gen_time = @elapsed graph = gen_graph(process)
opt_time = 0.0
if !skip_unoptimized
bench_process(process, "$process not optimized tape", graph, gen_time, opt_time)
end
opt_time = @elapsed optimize_to_fixpoint!(optimizer, graph)
bench_process(process, "$process reduced tape", graph, gen_time, opt_time)
return nothing
end
function bench_abc(process_string::String)
optimizer = ReductionOptimizer()
# AB->AB
process = parse_process(process_string, ABCModel())
gen_time = @elapsed graph = parse_dag("input/$process_string.txt", ABCModel())
bench_process(process, "$process not optimized tape", graph, gen_time, 0.0)
opt_time = @elapsed optimize_to_fixpoint!(optimizer, graph)
bench_process(process, "$process reduced tape", graph, gen_time, opt_time)
return nothing
end
# sadly cannot put these in functions because the world age must increase after the function is created which happens only in the global scope
## -- WARMUP TO COMPILE FUNCTIONS first
optimizer = ReductionOptimizer()
process = parse_process("ke->kke", QEDModel())
gen_time = @elapsed graph = gen_graph(process)
opt_time = @elapsed optimize_to_fixpoint!(optimizer, graph)
bench_process(process, "warmup", graph, gen_time, opt_time)
# AB->AB^3
process = parse_process("AB->ABBB", ABCModel())
gen_time = @elapsed graph = parse_dag("input/AB->ABBB.txt", ABCModel())
opt_time = @elapsed optimize_to_fixpoint!(optimizer, graph)
bench_process(process, "warmup", graph, gen_time, opt_time)
## -- WARMUP END
# compton
bench_qed("ke->ke")
bench_qed("ke->kke")
bench_qed("ke->kkke")
bench_qed("ke->kkkke")
bench_qed("ke->kkkkke")
bench_qed("ke->kkkkkke")
bench_qed("ke->kkkkkkke", true)
bench_abc("AB->AB")
bench_abc("AB->ABBB")
bench_abc("AB->ABBBBB")
CSV.write(results_filename, df)