portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2026-08-29 13:59:36 +01:00
Code Releases Activity

Ring allgather too. Let's nail the Dense CoarseCoarse

Browse Source
This commit is contained in:
Peter Boyle
2026-08-26 21:05:36 -04:00
parent 8e4a304626
commit dc1ae3185a
4 changed files with 144 additions and 16 deletions
Download Patch File Download Diff File Expand all files Collapse all files
systems/Frontier/smoother_modes.job
+15 -12
View File
@@ -4,7 +4,7 @@
#SBATCH --ntasks-per-node=8
#SBATCH --cpus-per-task=7
#SBATCH --gpus-per-node=8
#SBATCH --time=1:00:00
#SBATCH --time=1:45:00
#SBATCH --account=phy157_dwf
#SBATCH --gpu-bind=none
#SBATCH --exclusive
@@ -89,7 +89,7 @@ export DENSE_CC=1
export DENSE_APPLY_PROFILE=1
unset DENSE_CC_CHECK
export DENSE_SPLITK=128
export DENSE_DEVICE_SUM=2 # cartesian P2P ring: no 8 MB device-allreduce cliff (NRHS=12 abort)
export DENSE_DEVICE_SUM=4 # cartesian P2P ring ALLGATHER: ~8x fewer bytes than the padded allreduce (=2); no collectives, no size cliff
export GRID_ALLOC_NCACHE_LARGE=64
export NRHS=4
export PowerIterations=0
@@ -129,16 +129,19 @@ run_cell () {
grep -h "SCHUR fp64 distributed invert took\|GB/s/rank" $fname | sed 's/^Grid : Message : [0-9.]* s : //' | cut -c1-120 | head -2
}
# name Fso Fss Csn fine coarse
run_cell L0_overshoot 12 1.0 6 replay gcr # the converged overshoot (M3), now with last-call selection + refresh 5
run_cell L1_csn2 12 1.0 2 replay gcr # coarse smoother back to the banked 2 steps
run_cell L2_fso8 8 1.0 2 replay gcr
run_cell L3_fss05 8 0.5 2 replay gcr
run_cell L4_banked 6 0.5 2 replay gcr # nearest to the banked adaptive point
# Coarse replay: same lesson as the fine level -- record a DECENT polynomial
# first (overshoot: 6-step coarse smoother, shift 2.0) and back off from there.
run_cell L5_coarse6 8 0.5 6 replay replay # coarse frozen at a 6-step polynomial (overshoot)
run_cell L6_coarse4 8 0.5 4 replay replay # back off
# The (order x shift) table at Csn=2, one axis at a time from the overshoot,
# so a failing cell identifies WHICH knob it needed. Reference: 28.57 s.
# name Fso Fss Csn fine coarse
run_cell L0_overshoot 12 1.0 6 replay gcr # M3's point with last-call selection + record 8..16 + refresh 5
run_cell L1_12_10 12 1.0 2 replay gcr # coarse smoother back to 2 steps; the row/column anchor
run_cell L2_08_10 8 1.0 2 replay gcr # order axis
run_cell L3_06_10 6 1.0 2 replay gcr
run_cell L4_12_05 12 0.5 2 replay gcr # shift axis
run_cell L5_08_05 8 0.5 2 replay gcr
run_cell L6_06_05 6 0.5 2 replay gcr # nearest to the banked adaptive point (Fso6/Fss0.1)
# Coarse replay: record a DECENT polynomial first (6-step, shift 2.0) and back off.
run_cell L7_coarse6 8 1.0 6 replay replay
run_cell L8_coarse4 8 1.0 4 replay replay
echo "========================================================="
echo "summary"