V&V » RieCG: Irregular Sedov blast with load balancing

The Sedov problem

We computed the Sedov problem [3], widely used in shock hydrodynamics to test the ability of numerical methods to maintain symmetry. In this problem, a source of energy is defined to produce a shock in a single computational cell at the origin at t=0. The solution is a spherically spreading wave starting from a single point. We used a domain that is eighth of a sphere with the mesh consisting of 23,191,232 tetrahedra and 3,956,135 points. See also RieCG: Sedov blast.

A modified Sedov problem

To exercise Charm++'s built-in load balancing, we modified the Sedov problem by adding extra computational load to the function that computes the pressure based on density and internal energy: if fluid density > 2.0 then sleep(1ms). This increases the cost of the equation of state evaluation whose location propagates in space and time, which induces load imbalance across multiple mesh partitions in parallel.

The images below depict the Spatial distributions of the extra load, corresponding to the fluid density exceeding the value 2.0, during time evolution of the Sedov solution: shortly after the onset of load imbalance and at a later time. As the computational domain decomposed into multiple partitions, worked on by different compute cores and nodes, the parallel load becomes inhomogeneous and work units must all wait for the slowest one.

Code revision to reproduce

To reproduce the results below, use code revision fadd412, the following patch

diff --git a/src/Physics/EOS.hpp b/src/Physics/EOS.hpp
index e697caf..8403cb6 100644
--- a/src/Physics/EOS.hpp
+++ b/src/Physics/EOS.hpp
@@ -10,6 +10,8 @@
 // *****************************************************************************
 #pragma once

+#include <thread>
+
 #include "InciterConfig.hpp"

 namespace inciter {
@@ -29,6 +31,8 @@ using inciter::g_cfg;
 inline double
 pressure( double r, double e ) {
   auto g = g_cfg.get< tag::mat_spec_heat_ratio >();
+  using namespace std::chrono_literals;
+  if (r > 2.0) std::this_thread::sleep_for( 1ms );
   return r * e * (g-1.0);
 }

by saving the above patch to file "p" and applying it as

patch -p1 -i p

Then use the control file below.

Control file

-- vim: filetype=lua:

print "Sedov blast wave"

nstep = 500
ttyi = 1
cfl = 0.5

part = "rcb"

problem = {
  name = "sedov",
  -- p0 = 4.86e+3        -- sedov_coarse.exo
  -- p0 = 3.63e+4        -- sedov00.exo
  -- p0 = 2.32e+5        -- sedov01.exo
     p0 = 1.85e+6        -- sedov02.exo
  -- p0 = 14773333.33333 -- sedov02.exo+t0ref:u
}

mat = { spec_heat_ratio = 5/3 }

bc_sym = {
  sideset = { 1, 2, 3 }
}

--href = {
--  t0 = true,
--  init = { "uniform" }
--}

diag = {
  iter = 1000,
  format = "scientific"
}

fieldout = {
  iter = 1000
}

Run on 32 CPUs

# no extra load: run without applying the patch above
./charmrun +p32 Main/inciter -i sedov02.exo -c sedov.q -b -l 1000
# extra load, no lb: apply the patch above, and run with
./charmrun +p32 Main/inciter -i sedov02.exo -c sedov.q -b -l 1000
# extra load, lb: apply the patch above, and run with
./charmrun +p32 Main/inciter -i sedov02.exo -c sedov.q -b -u 0.995 -l 20 +balancer MetisLB +balancer GreedyRefineLB

In the last command line above, we apply overdecomposition with -u 0.995, yielding 6399 mesh partitions on 32 CPUs, and turn on load balancing in Charm++ to use the MetisLB strategy in the first load balancing step and GreedyRefine subsequently. The -l 20 command line argument tells Charm++ to perform load balancing every 20th time step. Consult the Charm++ manual for details on load balancing.

Timings on a single workstation

The figure and table below show the effect of load balancing on wall-clock time: in this particular case the extra load would make the simulation about 36x more expensive, which load balancing speeds up by 6.1x.

Measured wall-clock time of each time step during the Sedov calculation without extra load, as well as with extra load without and with load balancing. The area below each curve is proportional to the total computational cost.

Timings for the Sedov problem with and without load balancing on a single (shared-memory) workstation.

Run on multiple compute nodes

The following command line can be used to run the problem on a cluster of networked compute nodes each containing 32 CPU cores. Other than building Charm++ in SMP mode, no code changes are required.

# extra load, lb: apply the patch above, and run with
./charmrun +p $((31*$SLURM_NNODES)) Main/inciter -i sedov02.exo -c sedov.q +ppn 31 +pemap 1-31 +commap 0 -b -u 0.992 -l 20 +balancer DistributedLB

The above command line instructs the runtime system in SMP mode to execute on $SLURM_NNODES compute nodes. On each node it will assign 31 cores for computation and will designate 1 for communication only. Furthermore, Xyst will overdecompose the problem using a virtualization parameter of 0.992, which yields 30,962 mesh partitions assigned to the 248 compute cores. Such fine-grained work-units allow the runtime system to perform effective load balancing, which is invoked at every 20th time step.

Timings on a cluster

The following figure shows measured wall-clock times for each time step during computing the modified Sedov problem with and without extra load in SMP mode on 8 network-connected compute nodes. The area below each curve is proportional to the total computational cost of that simulation.

Timings for the Sedov problem with and without load balancing on 8 (network-connected, distributed-memory) compute nodes.

Summary

The above examples show that combining overdecomposition and Charm++'s automatic load balancing results in an order of magnitude reduction in wall-clock times for this particular irregular work-load. In this case the speed-up is about 6x on a single workstation and over 10x on multiple compute nodes.

References