mirror of
https://github.com/paboyle/Grid.git
synced 2026-08-29 13:59:36 +01:00
Debug page fault
This commit is contained in:
@@ -0,0 +1,108 @@
|
||||
#!/bin/bash -l
|
||||
#SBATCH --job-name=nrhs-fault
|
||||
#SBATCH --nodes=1
|
||||
#SBATCH --ntasks-per-node=8
|
||||
#SBATCH --cpus-per-task=7
|
||||
#SBATCH --gpus-per-node=8
|
||||
#SBATCH --time=0:45:00
|
||||
#SBATCH --account=phy157_dwf
|
||||
#SBATCH --gpu-bind=none
|
||||
#SBATCH --exclusive
|
||||
#SBATCH --mem=0
|
||||
#SBATCH -S 0
|
||||
#SBATCH -q debug
|
||||
|
||||
##############################################################################
|
||||
# GPU "Memory access fault ... Reason: Unknown" in the 3-level example,
|
||||
# 2026-08-27/28: NRHS=12 and NRHS=6 (adaptive gcr/gcr) deep in the solve; once
|
||||
# at Nrhs=1 after an Nrhs=4 solve (replay cell S6). NRHS=4 has never faulted.
|
||||
# --debug-signals catches nothing: the fault is raised by the ROCm runtime,
|
||||
# asynchronously, possibly long after the offending kernel/copy was queued.
|
||||
#
|
||||
# Plan: a 1-node reproducer (debug queue, minutes), then make the fault
|
||||
# ATTRIBUTABLE. AMD_SERIALIZE_KERNEL=3 / AMD_SERIALIZE_COPY=3 make every
|
||||
# kernel launch and copy synchronous, so the fault surfaces at the call that
|
||||
# caused it; AMD_LOG_LEVEL=3 prints each kernel's name as it launches, so the
|
||||
# last name before the fault is the culprit. ~5-10x slower: fine at this size.
|
||||
#
|
||||
# Cells: F0 NRHS=4 control | F1 NRHS=6 (reproduce?) | F2 NRHS=6 serialised
|
||||
# F3 NRHS=6 DENSE_CC=0 (no dense bottom) | F4 NRHS=6 --device-mem 48000 (no LRU eviction)
|
||||
# Read: grep -n -B8 "Memory access fault" log.fault.F* ; for F2 the last
|
||||
# "ShaderName"/kernel lines before the fault; F3/F4 say whether the
|
||||
# dense bottom or MemoryManager eviction is involved.
|
||||
##############################################################################
|
||||
|
||||
cat << EOF > select_gpu
|
||||
#!/bin/bash
|
||||
export GPU_MAP=(0 1 2 3 7 6 5 4)
|
||||
export NUMA_MAP=(3 3 1 1 2 2 0 0)
|
||||
export GPU=\${GPU_MAP[\$SLURM_LOCALID]}
|
||||
export NUMA=\${NUMA_MAP[\$SLURM_LOCALID]}
|
||||
export HIP_VISIBLE_DEVICES=\$GPU
|
||||
unset ROCR_VISIBLE_DEVICES
|
||||
if [ \$SLURM_PROCID = "0" ]; then echo \$*; fi
|
||||
exec numactl -m \$NUMA -N \$NUMA \$*
|
||||
EOF
|
||||
chmod +x ./select_gpu
|
||||
|
||||
root=$HOME/ParallelIO/systems/Frontier
|
||||
source $root/sourceme-rocm7.2.sh
|
||||
|
||||
export OMP_NUM_THREADS=7
|
||||
export MPICH_GPU_SUPPORT_ENABLED=1
|
||||
export MPICH_SMP_SINGLE_COPY_MODE=CMA
|
||||
export MPICH_OFI_NIC_POLICY=GPU
|
||||
module load libfabric
|
||||
|
||||
BIN=$root/examples/Example_pvdagm_v2_3level_DenseCoarseMatrix
|
||||
OPTS1="--accelerator-threads 8 --shm 4096 --shm-mpi 1"
|
||||
vol=16.16.16.32
|
||||
MPI_GEOM=1.1.2.4
|
||||
|
||||
# Small-lattice setup: 16^3x32 on 8 GCDs, blocks 2^4 twice -> coarse 8.8.8.16, coarse-coarse
|
||||
# 4.4.4.8 (N = 512 x 60 = 30720, 8 blocks of 3840 on a 2x4 grid; SCHUR2D_LEAF_SPAN=2 -> W=7680 leaves).
|
||||
# The subspace is generated on first use and saved to SUBSPACE_FILE (a few minutes).
|
||||
export SUBSPACE_FILE=$PWD/subspace_16x32_nb64.scidac
|
||||
unset SLAB_FILE
|
||||
export HOT_START=1
|
||||
export MASS=0.00078
|
||||
export BLOCK=2.2.2.2
|
||||
export BLOCK2=2.2.2.2
|
||||
export DENSE_SCHUR=1
|
||||
export DENSE_SCHUR2D=1
|
||||
export DENSE_CC=1
|
||||
export DENSE_SPLITK=8
|
||||
export DENSE_DEVICE_SUM=4
|
||||
export SCHUR2D_LEAF_SPAN=2
|
||||
export SCHUR2D_LEAF_LU=1
|
||||
export SCHUR2D_PROBE=0
|
||||
export GRID_ALLOC_NCACHE_LARGE=64
|
||||
# banked adaptive solver settings
|
||||
export FineSmootherShift=0.1 FineSmootherOrder=6 FineSmootherMmax=4
|
||||
export CoarseSmootherShift=2.0 CoarseSmootherNstep=2 CoarseSmootherMmax=2
|
||||
export CoarseSolverTol=0.05 CoarseSolverOrder=200 CoarseSolverMmax=8
|
||||
export OuterTol=1e-8 OuterMmax=6 OuterNstep=12
|
||||
export PowerIterations=0 SmootherCoeffLog=0 PolyVerbose=0
|
||||
|
||||
run_cell () {
|
||||
name=$1; devmem=$2; shift 2
|
||||
GRID_LOG_EXTRA=""; for kv in "$@"; do case $kv in GRID_LOG_EXTRA=*) GRID_LOG_EXTRA=${kv#GRID_LOG_EXTRA=};; esac; done
|
||||
echo "----- $name : --device-mem $devmem $* -----"
|
||||
fname=log.fault.$name
|
||||
logopt=""; [ -n "$GRID_LOG_EXTRA" ] && logopt="--log Error,Warning,Message,$GRID_LOG_EXTRA"
|
||||
env "$@" srun -N1 -n8 --kill-on-bad-exit=1 ./select_gpu $BIN --mpi ${MPI_GEOM} --grid $vol $OPTS1 --device-mem $devmem --comms-overlap $logopt > $fname 2>&1
|
||||
echo " exit $?"; sleep 10
|
||||
grep -h "V2 3-level solve\|Memory access fault\|GRID_ASSERT\|Converged on iteration" $fname | sed 's/^Grid : Message : [0-9.]* s : //' | tail -4 | cut -c1-140
|
||||
grep -n -B3 "Memory access fault" $fname | tail -5 | cut -c1-160
|
||||
}
|
||||
|
||||
# name devmem environment
|
||||
run_cell F0_nrhs4 16000 NRHS=4 # control: never faulted at 288
|
||||
run_cell F1_nrhs6 16000 NRHS=6 # reproduce on 1 node?
|
||||
run_cell F2_nrhs6_ser 16000 NRHS=6 AMD_SERIALIZE_KERNEL=3 AMD_SERIALIZE_COPY=3 AMD_LOG_LEVEL=3 GRID_LOG_EXTRA=Memory # attributable fault
|
||||
run_cell F3_nrhs6_nodense 16000 NRHS=6 DENSE_CC=0 # dense bottom out of the loop
|
||||
run_cell F4_nrhs6_bigmem 48000 NRHS=6 # no MemoryManager eviction
|
||||
|
||||
echo "========================================================="
|
||||
for f in log.fault.F*; do echo "$f: $(grep -c 'Memory access fault' $f) faults; $(grep -h 'V2 3-level solve' $f | tail -1 | sed 's/.*V2/V2/' | cut -c1-60)"; done
|
||||
echo "========================================================="
|
||||
Reference in New Issue
Block a user