Skip to content

Instantly share code, notes, and snippets.

CLIMA Overlap Timings

Runs

Unless noted otherwise all the runs were on an Amazon EC2 instance of type p3.16xlarge with 8 Titan V100s.

Bomex

Before Overlap

@lcw
lcw / README.md
Last active March 31, 2020 00:30

README

┌ Warning: Replacing docs for `CLIMA.DGmethods.knl_nodal_update_aux! :: Union{}` in module `CLIMA.DGmethods`
└ @ Base.Docs docs/Docs.jl:223
┌ Info: Configuration
│ ArrayType = CuArrays.CuArray
│ FT = Float64
│ NumericalFlux = Rusanov
└ dims = 2
┌ Info: Starting refinement level 1
│ numelems = (5, 5, 0)
│ dt = 3.5087719298245615e-06
signal (15): Terminated
in expression starting at /home/lucas/research/code/CLIMA/test/Arrays/mpi_comm.jl:180
clock_gettime at linux-vdso.so.1 (unknown line)
__clock_gettime at /lib64/libc.so.6 (unknown line)
unknown function (ip: 0x7f24a86e58fd)
unknown function (ip: 0x7f24a87896d6)
unknown function (ip: 0x7f24a87ab343)
unknown function (ip: 0x7f24a86bef7c)
unknown function (ip: 0x7f24a85cd08d)
┌ Info: Waiting on CUDAEvent
│ origin =
│ 14-element Array{Base.StackTraces.StackFrame,1}:
│ CudaEvent at cuda.jl:50 [inlined]
│ ⋮
│ current =
│ 12-element Array{Base.StackTraces.StackFrame,1}:
│ macro expansion at logging.jl:320 [inlined]
│ ⋮
└ token = Symbol("##546")
┌ Warning: Replacing docs for `CLIMA.DGmethods.knl_nodal_update_aux! :: Union{}` in module `CLIMA.DGmethods`
└ @ Base.Docs docs/Docs.jl:223
signal (15): Terminated
in expression starting at /home/lucas/research/code/CLIMA/test/Arrays/mpi_comm.jl:180
epoll_pwait at /build/glibc-t7JzpG/glibc-2.30/misc/../sysdeps/unix/sysv/linux/epoll_pwait.c:42
uv__io_poll at /workspace/srcdir/libuv/src/unix/linux-core.c:270
uv_run at /workspace/srcdir/libuv/src/unix/core.c:359
jl_process_events at /buildworker/worker/package_linux64/build/src/jl_uv.c:227
process_events at ./libuv.jl:104 [inlined]
┌ Warning: Replacing docs for `CLIMA.DGmethods.knl_nodal_update_aux! :: Union{}` in module `CLIMA.DGmethods`
└ @ Base.Docs docs/Docs.jl:223
signal (15): Terminated
in expression starting at /home/lucas/research/code/CLIMA/test/Arrays/mpi_comm.jl:180
epoll_pwait at /build/glibc-t7JzpG/glibc-2.30/misc/../sysdeps/unix/sysv/linux/epoll_pwait.c:42
uv__io_poll at /workspace/srcdir/libuv/src/unix/linux-core.c:270
uv_run at /workspace/srcdir/libuv/src/unix/core.c:359
jl_process_events at /buildworker/worker/package_linux64/build/src/jl_uv.c:227
process_events at ./libuv.jl:104 [inlined]
┌ Warning: Replacing docs for `CLIMA.DGmethods.knl_nodal_update_aux! :: Union{}` in module `CLIMA.DGmethods`
└ @ Base.Docs docs/Docs.jl:223
signal (15): Terminated
in expression starting at /home/lucas/research/code/CLIMA/test/Arrays/mpi_comm.jl:180
epoll_pwait at /build/glibc-t7JzpG/glibc-2.30/misc/../sysdeps/unix/sysv/linux/epoll_pwait.c:42
uv__io_poll at /workspace/srcdir/libuv/src/unix/linux-core.c:270
uv_run at /workspace/srcdir/libuv/src/unix/core.c:359
jl_process_events at /buildworker/worker/package_linux64/build/src/jl_uv.c:227
process_events at ./libuv.jl:104 [inlined]
┌ Warning: Package KernelAbstractions does not have MPI in its dependencies:
│ - If you have KernelAbstractions checked out for development and have
│ added MPI as a dependency but haven't updated your primary
│ environment's manifest file, try `Pkg.resolve()`.
│ - Otherwise you may need to report an issue with KernelAbstractions
└ Loading MPI into KernelAbstractions from project dependency, future warnings for KernelAbstractions are suppressed.
┌ Warning: Replacing docs for `CLIMA.DGmethods.knl_nodal_update_aux! :: Union{}` in module `CLIMA.DGmethods`
└ @ Base.Docs docs/Docs.jl:223
┌ Warning: The call to compilecache failed to create a usable precompiled cache file for CLIMA [777c4786-024f-11e9-21a3-85d5d4106250]
│ exception = Required dependency KernelAbstractions [63c18a36-062a-441e-b654-da1e3ab1ce7c] failed to load from a cache file.

Benchmarking Direct Columnwise Solvers

Here we are benchmarking the direct columnwise solvers using CLIMA version 4b8f986deb4f8c5d9db3db16083a0518bc971bba.

Global Simulations

The goal here is to test the cost of the solvers per time step. We leave it to other studies to see if the expected time step size can be obtained.