JaxPM/notebooks/04-MultiGPU_PM_Solvers.ipynb

Multi-GPU Particle Mesh Simulation with Advanced Solvers¶

In [ ]:

import os
os.environ["EQX_ON_ERROR"] = "nan"
import jax
import jax.numpy as jnp
import jax_cosmo as jc

from jaxpm.kernels import interpolate_power_spectrum
from jaxpm.painting import cic_paint_dx
from jaxpm.pm import linear_field, lpt, make_diffrax_ode
from functools import partial
from diffrax import ConstantStepSize, LeapfrogMidpoint,Dopri5 , PIDController , ODETerm, SaveAt, diffeqsolve

Note: This notebook requires 8 devices (GPU or TPU).
If you're running on CPU or don't have access to 8 devices,
you can simulate multiple devices by adding the following code at the start BEFORE IMPORTING JAX:

import os
os.environ["JAX_PLATFORM_NAME"] = "cpu"
os.environ["XLA_FLAGS"] = "--xla_force_host_platform_device_count=8"

Recommended only for debugging. If used, you must probably lower the resolution of the mesh.

In [2]:

assert jax.device_count() >= 8, "This notebook requires a TPU or GPU runtime with 8 devices"

Setting Up Device Mesh and Sharding for Multi-GPU Simulation¶

This cell configures a 2x4 device mesh across 8 devices and sets up named sharding to distribute data efficiently.

• Device Mesh: pdims = (2, 4) arranges devices in a 2x4 grid. create_device_mesh(pdims) initializes this layout across available GPUs.
• Sharding with Mesh: Mesh(devices, axis_names=('x', 'y')) assigns the mesh grid axes, which allows flexible mapping of array data across devices.
• PartitionSpec and NamedSharding: PartitionSpec defines data partitioning across mesh axes ('x', 'y'), and NamedSharding(mesh, P('x', 'y')) specifies this sharding scheme for arrays in the simulation.

More info on Sharding in general in Distributed arrays and automatic parallelization

In [3]:

from jax.experimental.mesh_utils import create_device_mesh
from jax.experimental.multihost_utils import process_allgather
from jax.sharding import Mesh, NamedSharding
from jax.sharding import PartitionSpec as P

all_gather = partial(process_allgather, tiled=False)

pdims = (2, 4)
devices = create_device_mesh(pdims)
mesh = Mesh(devices, axis_names=('x', 'y'))
sharding = NamedSharding(mesh, P('x', 'y'))

In [ ]:

@partial(jax.jit , static_argnums=(2,3,4,5))
def run_simulation_with_fields(omega_c, sigma8,mesh_shape,box_size,halo_size , snapshots):
    mesh_shape = (mesh_shape,) * 3
    box_size = (box_size,) * 3
    # Create a small function to generate the matter power spectrum
    k = jnp.logspace(-4, 1, 128)
    pk = jc.power.linear_matter_power(
        jc.Planck15(Omega_c=omega_c, sigma8=sigma8), k)
    pk_fn = lambda x: interpolate_power_spectrum(x, k, pk, sharding)

    # Create initial conditions
    initial_conditions = linear_field(mesh_shape,
                                      box_size,
                                      pk_fn,
                                      seed=jax.random.PRNGKey(0),
                                      sharding=sharding)


    cosmo = jc.Planck15(Omega_c=omega_c, sigma8=sigma8)

    # Initial displacement
    dx, p, f = lpt(cosmo,
                   initial_conditions,
                   a=0.1,
                   order=2,
                   halo_size=halo_size,
                   sharding=sharding)

    # Evolve the simulation forward
    ode_fn = ODETerm(
        make_diffrax_ode(cosmo, mesh_shape, paint_absolute_pos=False))
    solver = LeapfrogMidpoint()

    stepsize_controller = ConstantStepSize()
    res = diffeqsolve(ode_fn,
                      solver,
                      t0=0.1,
                      t1=1.,
                      dt0=0.01,
                      y0=jnp.stack([dx, p], axis=0),
                      args=cosmo,
                      saveat=SaveAt(ts=snapshots),
                      stepsize_controller=stepsize_controller)
    ode_fields = [cic_paint_dx(sol[0], halo_size=halo_size, sharding=sharding) for sol in res.ys]
    lpt_field = cic_paint_dx(dx , halo_size=halo_size, sharding=sharding)
    return initial_conditions, lpt_field, ode_fields, res.stats

Large-Scale Simulation Across Multiple Devices¶

In this cell, we run a large simulation that would not be feasible on a single device. By distributing data across multiple devices, we achieve a higher resolution (mesh_shape = 1024 and box_size = 1000.) with effective boundary handling using a halo_size of 128.

We gather initial conditions and computed fields from all devices for visualization.

In [ ]:

from jaxpm.plotting import plot_fields_single_projection

mesh_shape = 1024
box_size = 1000.
halo_size = 128
snapshots = (0.5 , 1.0)

initial_conditions , lpt_field , ode_fields , solver_stats = run_simulation_with_fields(0.25, 0.8 , mesh_shape, box_size , halo_size , snapshots)
ode_fields[-1].block_until_ready()
%timeit initial_conditions , lpt_field , ode_fields , solver_stats = run_simulation_with_fields(0.25, 0.8 , mesh_shape, box_size , halo_size , snapshots);ode_fields[-1].block_until_ready()

initial_conditions_g = all_gather(initial_conditions)
lpt_field_g = all_gather(lpt_field)
ode_fields_g = [all_gather(p) for p in ode_fields]

fields = {"Initial Conditions" : initial_conditions_g , "LPT Field" : lpt_field_g}
for i , field in enumerate(ode_fields_g):
    fields[f"field_{i}"] = field
plot_fields_single_projection(fields,project_axis=0)

45.6 s ± 11.4 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

This simulation runs in 45 seconds (less than half a second per step), which is impressive for a setup with over one billion particles (since ( 1024^3 \approx 1.07 ) billion). This performance demonstrates the efficiency of distributing data and computation across multiple devices.

Comparing ODE Solvers: Leapfrog vs. Dopri5¶

Next, we compare the Leapfrog solver with Dopri5 (an adaptive Runge-Kutta method) to observe differences in accuracy and performance for particle evolution.

In [11]:

mesh_shape = 512
box_size = 512.
halo_size = 64
snapshots = (0.5 , 1.0)

initial_conditions , lpt_field , ode_fields , solver_stats = run_simulation_with_fields(0.25, 0.8 , mesh_shape, box_size , halo_size , snapshots)
ode_fields[-1].block_until_ready()
%timeit initial_conditions , lpt_field , ode_fields , solver_stats = run_simulation_with_fields(0.25, 0.8 , mesh_shape, box_size , halo_size , snapshots);ode_fields[-1].block_until_ready()

5.04 s ± 9.5 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

In [ ]:

mesh_shape = 512
box_size = 512.
halo_size = 64
snapshots = (0.5, 1.0)

@partial(jax.jit , static_argnums=(2,3,4,5))
def run_simulation_with_dopri(omega_c, sigma8,mesh_shape,box_size,halo_size , snapshots):
    mesh_shape = (mesh_shape,) * 3
    box_size = (box_size,) * 3
    # Create a small function to generate the matter power spectrum
    k = jnp.logspace(-4, 1, 128)
    pk = jc.power.linear_matter_power(
        jc.Planck15(Omega_c=omega_c, sigma8=sigma8), k)
    pk_fn = lambda x: interpolate_power_spectrum(x, k, pk, sharding)

    # Create initial conditions
    initial_conditions = linear_field(mesh_shape,
                                      box_size,
                                      pk_fn,
                                      seed=jax.random.PRNGKey(0),
                                      sharding=sharding)


    cosmo = jc.Planck15(Omega_c=omega_c, sigma8=sigma8)

    # Initial displacement
    dx, p, f = lpt(cosmo,
                   initial_conditions,
                   a=0.1,
                   order=2,
                   halo_size=halo_size,
                   sharding=sharding)

    # Evolve the simulation forward
    ode_fn = ODETerm(
        make_diffrax_ode(cosmo, mesh_shape, paint_absolute_pos=False))
    solver = Dopri5()

    stepsize_controller = PIDController(rtol=1e-5,atol=1e-5)
    res = diffeqsolve(ode_fn,
                      solver,
                      t0=0.1,
                      t1=1.,
                      dt0=0.01,
                      y0=jnp.stack([dx, p], axis=0),
                      args=cosmo,
                      saveat=SaveAt(ts=snapshots),
                      stepsize_controller=stepsize_controller)
    ode_fields = [cic_paint_dx(sol[0], halo_size=halo_size, sharding=sharding) for sol in res.ys]
    lpt_field = cic_paint_dx(dx , halo_size=halo_size, sharding=sharding)
    return initial_conditions, lpt_field, ode_fields, res.stats

initial_conditions , lpt_field , ode_fields , solver_stats = run_simulation_with_dopri(0.25, 0.8 , mesh_shape, box_size , halo_size , snapshots)
ode_fields[-1].block_until_ready()
%timeit initial_conditions , lpt_field , ode_fields , solver_stats = run_simulation_with_dopri(0.25, 0.8 , mesh_shape, box_size , halo_size , snapshots);ode_fields[-1].block_until_ready()

print(f"Solver Stats : {solver_stats}")

4.44 s ± 8.12 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)
Solver Stats : {'max_steps': Array(4096, dtype=int32, weak_type=True), 'num_accepted_steps': Array(12, dtype=int32, weak_type=True), 'num_rejected_steps': Array(0, dtype=int32, weak_type=True), 'num_steps': Array(12, dtype=int32, weak_type=True)}

In [10]:

initial_conditions_g = all_gather(initial_conditions)
lpt_field_g = all_gather(lpt_field)
ode_fields_g = [all_gather(p) for p in ode_fields]

fields = {"Initial Conditions" : initial_conditions_g , "LPT Field" : lpt_field_g}
for i , field in enumerate(ode_fields_g):
    fields[f"field_{i}"] = field
plot_fields_single_projection(fields,project_axis=0)

We can see how easily we can switch solvers here. Although Dopri5 offers adaptive stepping, it didn’t yield a significant performance boost over Leapfrog in this case.

Note: Dopri5 uses a PIDController for adaptive stepping, which might face challenges in distributed setups. In my experience, it works well without triggering all-gathers, but make sure to set:

os.environ["EQX_ON_ERROR"] = "nan"

before importing diffrax to handle any errors gracefully.

However, Dopri5 requires more memory than Leapfrog, making a $1024^3$ mesh simulation unfeasible on eight A100 GPUs with 80GB memory each!!. For larger setups, we’ll need more compute resources—this is covered in the final notebook, 05-MultiHost_PM.ipynb.