"Jash

Weak scaling grows the problem with the number of ranks, holding the work per rank fixed. It answers the question distributed solvers actually exist for: can I solve a proportionally bigger problem in the same wall-clock time by adding hardware? Strong scaling (the sibling document) fixes the problem and asks for speedup; weak scaling is usually the more honest test for memory-bound sparse solves, because per-rank working set stays constant by construction.

We solve 2-D finite-difference Laplacians with PETSc's CG under GAMG via LinearSolve.jl's PETScAlgorithm, with the target unknown count proportional to the rank count. Two properties are on trial:

• Algorithmic scalability: GAMG's iteration count should stay near-constant as N grows. The iteration column makes this auditable on every run.
• Parallel weak efficiencyT_1 / T_P: with perfect weak scaling every rank count solves its (proportionally larger) system in the baseline's time.

using MPI            # provides mpiexec()
using Plots
using Printf

const WORKER = joinpath(@__DIR__, "run_solve.jl")
const PROJECT = Base.active_project()

# MPICH_jll 5.x hydra can fail PMI bootstrap on a single node; `-launcher fork`
# spawns ranks with fork() instead and is harmless where hydra is healthy.
const MPIEXEC_ARGS = `-launcher fork`

# Run run_solve.jl under `mpiexec -n P`; parse the CSV line rank 0 prints:
#   ranks,N,nnz,solver,pc,time_s,residual,iters,retcode
# One thread per rank so parallelism comes only from the rank count.
function run_ranks(P; N, solver = "cg", pc = "gamg")
    cmd = `$(mpiexec()) $(MPIEXEC_ARGS) -n $P $(Base.julia_cmd()) --project=$(PROJECT) $(WORKER) $N $solver $pc`
    out = read(addenv(cmd, "OMP_NUM_THREADS" => "1"), String)
    line = strip(last(filter(!isempty, split(out, '\n'))))
    f = split(line, ',')
    return (ranks = parse(Int, f[1]), N = parse(Int, f[2]), nnz = parse(Int, f[3]),
        time = parse(Float64, f[6]), residual = parse(Float64, f[7]),
        iters = parse(Int, f[8]), retcode = f[9])
end

run_ranks (generic function with 1 method)

Run the sweep

Work per rank is sized for the benchmark runner (the pipeline was proven at smaller sizes first); extending N_PER_RANK and RANKS further charts a larger envelope at proportional run cost. Note the replicated-SparseMatrixCSC input path stores a full matrix copy per rank, so total memory grows with P × N here even though the solver's per-rank work is constant; the closing section discusses this bound.

const N_PER_RANK = 100_000
const RANKS = [1, 2, 4, 8]

results = [run_ranks(P; N = N_PER_RANK * P) for P in RANKS]

4-element Vector{@NamedTuple{ranks::Int64, N::Int64, nnz::Int64, time::Floa
t64, residual::Float64, iters::Int64, retcode::SubString{String}}}:
 (ranks = 1, N = 99856, nnz = 498016, time = 0.357702307, residual = 2.6078
93230536068e-9, iters = 14, retcode = "Success")
 (ranks = 2, N = 199809, nnz = 997257, time = 0.572087908, residual = 8.168
35835539473e-9, iters = 15, retcode = "Success")
 (ranks = 4, N = 399424, nnz = 1994592, time = 0.582770386, residual = 3.43
5836440055526e-9, iters = 16, retcode = "Success")
 (ranks = 8, N = 799236, nnz = 3992604, time = 0.722665721, residual = 7.20
9407772457451e-9, iters = 16, retcode = "Success")

Weak efficiency

t1 = results[1].time
weak_eff = [t1 / r.time for r in results]

println("ranks |    N    |  time (s)  | iters | weak eff. | residual  | retcode")
println("------+---------+------------+-------+-----------+-----------+--------")
for (r, e) in zip(results, weak_eff)
    @printf("%5d | %7d | %10.4g | %5d | %8.1f%% | %9.2e | %s\n",
        r.ranks, r.N, r.time, r.iters, 100 * e, r.residual, r.retcode)
end

# Auditability annotations, mirroring the strong-scaling document: iteration
# growth means the preconditioner is not algorithmically scaling (the time
# column then conflates solver work with parallel overhead), and efficiency
# above ~110% at these sizes usually reflects cache effects on a generic
# (JLL) PETSc build rather than real scaling.
itset = [r.iters for r in results]
if maximum(itset) - minimum(itset) > 0.25 * minimum(itset)
    @warn "GAMG iteration count grows >25% across the sweep — algorithmic " *
          "scalability is not holding at these sizes." iters = itset
end
if maximum(weak_eff) > 1.10
    @warn "Weak efficiency exceeds 110% — treat as a cache/working-set artifact " *
          "at small N, not a real result." weak_eff
end

ranks |    N    |  time (s)  | iters | weak eff. | residual  | retcode
------+---------+------------+-------+-----------+-----------+--------
    1 |   99856 |     0.3577 |    14 |    100.0% |  2.61e-09 | Success
    2 |  199809 |     0.5721 |    15 |     62.5% |  8.17e-09 | Success
    4 |  399424 |     0.5828 |    16 |     61.4% |  3.44e-09 | Success
    8 |  799236 |     0.7227 |    16 |     49.5% |  7.21e-09 | Success

Plot

p = plot(RANKS, 100 .* weak_eff;
    marker = :square, label = "PETSc CG + GAMG",
    xlabel = "MPI ranks (N = $(N_PER_RANK) × ranks)",
    ylabel = "weak efficiency (%)",
    title = "Weak scaling: constant work per rank",
    ylims = (0, 130), legend = :bottomleft)
hline!(p, [100]; linestyle = :dash, color = :gray, label = "ideal (100%)")
p

Reading the result

Flat iteration counts with efficiency near 100% is the headline weak-scaling claim: the solver does proportionally more work in the same time as ranks are added. Efficiency decay isolates parallel overhead (communication, setup), since constant per-rank work removes the serial-fraction confound that complicates strong scaling.

Two caveats bound what this document can show. First, the replicated-input path makes every rank assemble the full matrix, so the constant-work-per-rank premise holds for the solve phase but not for assembly: assembly cost grows with total N on every rank, which drags measured weak efficiency down even when the solver itself scales. Second, the same replication caps how far RANKS can extend before total memory becomes the binding constraint. A partitioned-input variant removes the second and third limits and is the natural follow-up for large-rank weak scaling.

Appendix

Appendix

These benchmarks are a part of the SciMLBenchmarks.jl repository, found at: https://github.com/SciML/SciMLBenchmarks.jl. For more information on high-performance scientific machine learning, check out the SciML Open Source Software Organization https://sciml.ai.

To locally run this benchmark, do the following commands:

using SciMLBenchmarks
SciMLBenchmarks.weave_file("benchmarks/LinearSolveDistributed","WeakScaling.jmd")

Computer Information:

Julia Version 1.12.7
Commit 6d172b025e4 (2026-08-15 08:05 UTC)
Build Info:
  Official https://julialang.org release
Platform Info:
  OS: Linux (x86_64-linux-gnu)
  CPU: 128 × AMD EPYC 7502 32-Core Processor
  WORD_SIZE: 64
  LLVM: libLLVM-18.1.7 (ORCJIT, znver2)
  GC: Built with stock GC
Threads: 128 default, 1 interactive, 128 GC (on 128 virtual cores)
Environment:
  JULIA_NUM_THREADS = auto

Package Information:

Status `~/github-runners/amdci3-1/_work/SciMLBenchmarks.jl/SciMLBenchmarks.jl/benchmarks/LinearSolveDistributed/Project.toml`
  [6e4b80f9] BenchmarkTools v1.8.0
⌃ [7ed4a6bd] LinearSolve v5.17.3
  [da04e1cc] MPI v0.20.27
  [3da0fdf6] MPIPreferences v0.1.12
⌃ [ace2c81b] PETSc v0.4.10
  [91a5bcdd] Plots v1.41.7
  [0bca4576] SciMLBase v3.54.0
  [31c91b34] SciMLBenchmarks v0.2.1 [loaded: `/home/crackauc/github-runners/amdci3-1/_work/SciMLBenchmarks.jl/SciMLBenchmarks.jl/src/SciMLBenchmarks.jl` (v0.2.1) expected `/home/crackauc/.julia/packages/SciMLBenchmarks/ceJyd/src/SciMLBenchmarks.jl` (v0.2.1)]
  [a0a7dd2c] SparseMatricesCSR v0.6.12
  [37e2e46d] LinearAlgebra v1.12.0
  [de0858da] Printf v1.11.0
  [2f01184e] SparseArrays v1.12.0
Info Packages marked with ⌃ have new versions available and may be upgradable.

And the full manifest:

Status `~/github-runners/amdci3-1/_work/SciMLBenchmarks.jl/SciMLBenchmarks.jl/benchmarks/LinearSolveDistributed/Manifest.toml`
  [47edcb42] ADTypes v1.24.0
  [14f7f29c] AMD v0.5.4
  [7d9f7c33] Accessors v0.1.45
  [79e6a3ab] Adapt v4.7.0
  [66dad0bd] AliasTables v1.1.3
⌃ [4fba245c] ArrayInterface v7.30.1
  [a9b6321e] Atomix v1.2.1
  [6e4b80f9] BenchmarkTools v1.8.0
  [62783981] BitTwiddlingConvenienceFunctions v0.1.6
  [2a0fbf3d] CPUSummary v0.2.7
  [fb6a15b2] CloseOpenIntervals v0.1.13
  [35d6a980] ColorSchemes v3.31.0
  [3da002f7] ColorTypes v0.12.1
  [c3611d14] ColorVectorSpace v0.11.0
  [5ae59095] Colors v0.13.1
  [38540f10] CommonSolve v0.2.14
  [bbf7d656] CommonSubexpressions v0.3.1
  [f70d9fcc] CommonWorldInvalidations v1.2.2
  [34da2185] Compat v4.18.1
  [a33af91c] CompositionsBase v0.1.2
  [2569d6c7] ConcreteStructs v0.2.8
  [187b0558] ConstructionBase v1.6.0
  [d38c429a] Contour v0.6.3
  [adafc99b] CpuId v0.3.1
  [a8cc5b0e] Crayons v4.2.0
  [9a962f9c] DataAPI v1.16.0
  [864edb3b] DataStructures v0.19.6
  [e2d170a0] DataValueInterfaces v1.0.0
  [8bb1440f] DelimitedFiles v1.9.1
  [163ba53b] DiffResults v1.1.0
  [b552c78f] DiffRules v1.16.0
  [ffbed154] DocStringExtensions v0.9.5
  [4e289a0a] EnumX v1.0.7
  [e2ba6199] ExprTools v0.1.11
  [c87230d0] FFMPEG v0.4.5
  [64ca27bc] FindFirstFunctions v3.2.1
⌅ [53c48c17] FixedPointNumbers v0.8.6
  [1fa38f19] Format v1.3.7
  [f6369f11] ForwardDiff v1.4.6
  [069b7b12] FunctionWrappers v1.1.3
  [77dc65aa] FunctionWrappersWrappers v1.13.0
  [46192b85] GPUArraysCore v0.2.0
  [28b8d3ca] GR v0.73.27
⌅ [eafb193a] Highlights v0.5.3
  [615f187c] IfElse v0.1.1
  [3587e190] InverseFunctions v0.1.17
  [92d709cd] IrrationalConstants v0.2.6
  [82899510] IteratorInterfaceExtensions v1.0.0
  [1019f520] JLFzf v0.1.11
  [692b3bcd] JLLWrappers v1.8.0
⌅ [682c06a0] JSON v0.21.4
  [ba0b0d4f] Krylov v0.10.10
  [2faa5264] LHLFactorization v2.2.2
  [b964fa9f] LaTeXStrings v1.4.1
  [23fbe1c1] Latexify v0.16.12
  [10f19ff3] LayoutPointers v0.1.17
⌃ [7ed4a6bd] LinearSolve v5.17.3
  [2ab3a3ac] LogExpFunctions v1.0.1
  [e6f89c97] LoggingExtras v1.2.0
  [da04e1cc] MPI v0.20.27
  [3da0fdf6] MPIPreferences v0.1.12
  [1914dd2f] MacroTools v0.5.16
  [d125e4d3] ManualMemory v0.1.8
  [299715c1] MarchingCubes v0.1.11
  [442fdcdd] Measures v0.3.3
  [e1d29d7a] Missings v1.2.0
  [ffc61752] Mustache v1.0.21
  [77ba4419] NaNMath v1.1.4
  [6fe1bfb0] OffsetArrays v1.17.0
  [bac558e1] OrderedCollections v2.0.1
⌃ [ace2c81b] PETSc v0.4.10
⌅ [69de0a69] Parsers v2.8.8
  [eebad327] PkgVersion v0.3.3
  [ccf2f8ad] PlotThemes v3.3.0
  [995b91a9] PlotUtils v1.4.4
  [91a5bcdd] Plots v1.41.7
  [f517fe37] Polyester v0.7.19
  [1d0040c9] PolyesterWeave v0.2.2
  [d236fae5] PreallocationTools v1.7.1
  [aea7be01] PrecompileTools v1.3.4
  [21216c6a] Preferences v1.6.0
  [43287f4e] PtrArrays v1.4.0
  [0c0d3e7f] PureKLU v1.5.0
  [3cdcf5f2] RecipesBase v1.3.4
  [01d81517] RecipesPipeline v0.6.12
  [731186ca] RecursiveArrayTools v4.5.1
  [189a3867] Reexport v1.2.2
  [05181044] RelocatableFolders v1.0.1
  [ae029012] Requires v1.3.1
  [7e49a35a] RuntimeGeneratedFunctions v0.5.26
  [94e857df] SIMDTypes v0.1.0
  [0bca4576] SciMLBase v3.54.0
  [31c91b34] SciMLBenchmarks v0.2.1 [loaded: `/home/crackauc/github-runners/amdci3-1/_work/SciMLBenchmarks.jl/SciMLBenchmarks.jl/src/SciMLBenchmarks.jl` (v0.2.1) expected `/home/crackauc/.julia/packages/SciMLBenchmarks/ceJyd/src/SciMLBenchmarks.jl` (v0.2.1)]
  [a6db7da4] SciMLLogging v2.1.0
⌃ [c0aeaf25] SciMLOperators v1.30.0
  [431bcebd] SciMLPublic v1.3.0
  [53ae85a6] SciMLStructures v1.10.5
  [6c6a2e73] Scratch v1.3.0
  [efcf1570] Setfield v1.1.2
  [992d4aef] Showoff v1.1.1
  [a2af1166] SortingAlgorithms v1.2.3
  [a57abbd0] SparseColumnPivotedQR v2.1.8
  [a0a7dd2c] SparseMatricesCSR v0.6.12
  [276daf66] SpecialFunctions v2.9.0
  [860ef19b] StableRNGs v1.0.4
  [aedffcd0] Static v1.4.6
  [0d7ed370] StaticArrayInterface v1.10.0
  [90137ffa] StaticArrays v1.9.20
  [1e83bf80] StaticArraysCore v1.4.4
  [10745b16] Statistics v1.11.5
  [82ae8749] StatsAPI v1.8.0
  [2913bbd2] StatsBase v0.34.13
  [7792a7ef] StrideArraysCore v0.5.9
  [69024149] StringEncodings v0.3.7
  [2efcf032] SymbolicIndexingInterface v0.3.55
  [3783bdb8] TableTraits v1.0.1
  [bd369af6] Tables v1.14.0
  [62fd8b95] TensorCore v0.1.1
  [8290d209] ThreadingUtilities v0.5.6
  [1cfade01] UnicodeFun v0.4.1
  [b8865327] UnicodePlots v3.8.4
  [013be700] UnsafeAtomics v0.3.2
  [41fe7b60] Unzip v0.2.0
  [44d3d7a6] Weave v0.10.12
  [ddb6d928] YAML v0.4.16
  [6e34b625] Bzip2_jll v1.0.9+0
  [83423d85] Cairo_jll v1.18.7+0
  [ee1fde0b] Dbus_jll v1.16.2+0
  [2702e6a9] EpollShim_jll v0.0.20230411+1
  [2e619515] Expat_jll v2.8.4+0
⌅ [b22a6f82] FFMPEG_jll v8.1.2+0
  [a3f928ae] Fontconfig_jll v2.17.1+0
  [d7e528f0] FreeType2_jll v2.14.3+1
  [559328eb] FriBidi_jll v1.0.17+0
  [0656b61e] GLFW_jll v3.5.1+0
  [d2c73de3] GR_jll v0.73.27+0
⌅ [b0724c58] GettextRuntime_jll v0.22.4+0
  [61579ee1] Ghostscript_jll v9.55.1+0
  [7746bdde] Glib_jll v2.88.3+0
  [3b182d85] Graphite2_jll v1.3.16+0
  [2e76f6c2] HarfBuzz_jll v100.14004.0+0
  [e33a78d0] Hwloc_jll v2.14.0+0
  [1d5cc7b8] IntelOpenMP_jll v2025.2.0+0
  [aacddb02] JpegTurbo_jll v3.2.0+1
  [c1c5ebd0] LAME_jll v3.100.3+0
  [88015f11] LERC_jll v4.2.0+0
  [1d63c593] LLVMOpenMP_jll v23.1.1+0
⌅ [e9f186c6] Libffi_jll v3.4.7+0
  [7e76a0d4] Libglvnd_jll v1.7.1+1
  [94ce4f54] Libiconv_jll v1.18.0+0
  [4b2f31a3] Libmount_jll v2.42.0+0
  [89763e89] Libtiff_jll v4.7.3+0
  [38a345b3] Libuuid_jll v2.42.0+0
  [856f044c] MKL_jll v2025.2.0+0
  [b5ada748] MPIABI_jll v1.0.0+0
  [7cb0a576] MPICH_jll v5.0.1+0
  [f1f71cc9] MPItrampoline_jll v5.5.6+0
  [9237b28f] MicrosoftMPI_jll v10.1.4+3
  [e7412a2a] Ogg_jll v1.3.6+0
  [656ef2d0] OpenBLAS32_jll v0.3.34+0
  [fe0851c0] OpenMPI_jll v5.0.11+0
  [efe28fd5] OpenSpecFun_jll v0.5.6+0
  [91d4177d] Opus_jll v1.6.1+0
  [8fa3689e] PETSc_jll v3.22.2+0
  [36c8627f] Pango_jll v1.58.2+0
  [30392449] Pixman_jll v0.46.4+0
  [c0090381] Qt6Base_jll v6.10.2+2
  [629bc702] Qt6Declarative_jll v6.10.2+2
  [ce943373] Qt6ShaderTools_jll v6.10.2+1
  [6de9746b] Qt6Svg_jll v6.10.2+0
  [e99dba38] Qt6Wayland_jll v6.10.2+1
  [aabda75e] SCALAPACK32_jll v2.2.302+0
  [a44049a8] Vulkan_Loader_jll v1.3.243+0
  [a2964d1f] Wayland_jll v1.24.0+0
⌅ [02c8fc9c] XML2_jll v2.13.9+0
  [ffd25f8a] XZ_jll v5.8.4+0
  [f67eecfb] Xorg_libICE_jll v1.1.2+0
  [c834827a] Xorg_libSM_jll v1.2.6+0
  [4f6342f7] Xorg_libX11_jll v1.8.13+0
  [0c0b7dd1] Xorg_libXau_jll v1.0.13+0
  [935fb764] Xorg_libXcursor_jll v1.2.4+0
  [a3789734] Xorg_libXdmcp_jll v1.1.6+0
  [1082639a] Xorg_libXext_jll v1.3.8+0
  [d091e8ba] Xorg_libXfixes_jll v6.0.2+0
  [a51aa0fd] Xorg_libXi_jll v1.8.4+0
  [d1454406] Xorg_libXinerama_jll v1.1.7+0
  [ec84b674] Xorg_libXrandr_jll v1.5.6+0
  [ea2f1a96] Xorg_libXrender_jll v0.9.12+0
  [a65dc6b1] Xorg_libpciaccess_jll v0.19.0+0
  [c7cfdc94] Xorg_libxcb_jll v1.17.1+0
  [cc61e674] Xorg_libxkbfile_jll v1.2.0+0
  [e920d4aa] Xorg_xcb_util_cursor_jll v0.1.6+0
  [12413925] Xorg_xcb_util_image_jll v0.4.1+0
  [2def613f] Xorg_xcb_util_jll v0.4.1+0
  [975044d2] Xorg_xcb_util_keysyms_jll v0.4.1+0
  [0d47668e] Xorg_xcb_util_renderutil_jll v0.3.10+0
  [c22f9ab0] Xorg_xcb_util_wm_jll v0.4.2+0
  [35661453] Xorg_xkbcomp_jll v1.4.7+0
  [33bec58e] Xorg_xkeyboard_config_jll v2.47.0+2
  [c5fb5394] Xorg_xtrans_jll v1.6.0+0
  [3161d3a3] Zstd_jll v1.5.7+1
  [35ca27e7] eudev_jll v3.2.14+0
⌅ [214eeab7] fzf_jll v0.61.1+0
  [a4ae2306] libaom_jll v3.14.1+0
  [0ac62f75] libass_jll v0.17.5+0
  [1183f4f0] libdecor_jll v0.2.2+0
  [8e53e030] libdrm_jll v2.4.134+0
  [2db6ffa8] libevdev_jll v1.13.4+0
  [f638f0a6] libfdk_aac_jll v2.0.4+0
  [36db933b] libinput_jll v1.28.1+0
  [b53b4c65] libpng_jll v1.6.58+0
  [9a156e7d] libva_jll v2.23.0+0
  [f27f6e37] libvorbis_jll v1.3.8+0
  [9aeb927a] mpif_jll v1.0.0+0
  [009596ad] mtdev_jll v1.1.7+0
  [1317d2d5] oneTBB_jll v2022.3.0+0
⌅ [1270edf5] x264_jll v10164.0.1+0
  [dfaa095f] x265_jll v4.1.0+0
  [d8fb68d0] xkbcommon_jll v1.13.0+0
  [0dad84c5] ArgTools v1.1.2
  [56f22d72] Artifacts v1.11.0
  [2a0f44e3] Base64 v1.11.0
  [ade2ca70] Dates v1.11.0
  [8ba89e20] Distributed v1.11.0
  [f43a241f] Downloads v1.7.0
  [7b1f6079] FileWatching v1.11.0
  [9fa8497b] Future v1.11.0
  [b77e0a4c] InteractiveUtils v1.11.0
  [ac6e5ff7] JuliaSyntaxHighlighting v1.12.0
  [4af54fe1] LazyArtifacts v1.11.0
  [b27032c2] LibCURL v0.6.4
  [76f85450] LibGit2 v1.11.0
  [8f399da3] Libdl v1.11.0
  [37e2e46d] LinearAlgebra v1.12.0
  [56ddb016] Logging v1.11.0
  [d6f4376e] Markdown v1.11.0
  [a63ad114] Mmap v1.11.0
  [ca575930] NetworkOptions v1.3.0
  [44cfe95a] Pkg v1.12.1
  [de0858da] Printf v1.11.0
  [9abbd945] Profile v1.11.0
  [3fa0cd96] REPL v1.11.0
  [9a3f8284] Random v1.11.0
  [ea8e919c] SHA v0.7.0
  [9e88b42a] Serialization v1.11.0
  [6462fe0b] Sockets v1.11.0
  [2f01184e] SparseArrays v1.12.0
  [f489334b] StyledStrings v1.11.0
  [4607b0f0] SuiteSparse
  [fa267f1f] TOML v1.0.3
  [a4e569a6] Tar v1.10.0
  [8dfed614] Test v1.11.0
  [cf7118a7] UUIDs v1.11.0
  [4ec0a83e] Unicode v1.11.0
  [e66e0078] CompilerSupportLibraries_jll v1.3.1+2
  [deac9b47] LibCURL_jll v8.15.0+0
  [e37daf67] LibGit2_jll v1.9.0+0
  [29816b5a] LibSSH2_jll v1.11.3+1
  [14a3606d] MozillaCACerts_jll v2025.11.4
  [4536629a] OpenBLAS_jll v0.3.29+0
  [05823500] OpenLibm_jll v0.8.7+0
  [458c3c95] OpenSSL_jll v3.5.6+0
  [efcefdf7] PCRE2_jll v10.44.0+1
  [bea87d4a] SuiteSparse_jll v7.8.3+2
  [83775a58] Zlib_jll v1.3.1+2
  [8e850b90] libblastrampoline_jll v5.15.0+0
  [8e850ede] nghttp2_jll v1.64.0+1
  [3f19e933] p7zip_jll v17.7.0+0
Info Packages marked with ⌃ and ⌅ have new versions available. Those with ⌃ may be upgradable, but those with ⌅ are restricted by compatibility constraints from upgrading. To see why use `status --outdated -m`