Skip to content
Profiling catalogue
Projects

Vlasov-Ampere, CPU node vs GPU node: GPU node

all 6all levels

gpu-node.h52026-10-17 12:00 UTCsynthetic 4-GPU node4 ranks6 regions0.0895414 s profiled

← Vlasov-Ampere, CPU node vs GPU node · ‹ CPU node

Summary

Section titled “Summary”
Wall time89.5 mssetup to finalize
Ranks4
Regions6
In regions95%of the session
Top hotspotsolve80.3% of the session
Load imbalance+4%solve, slowest on rank 2
  • solve (in simulation) is the largest hotspot: 278 ms over 40 calls, 80.3% of the session.
  • The three largest hotspots account for 92% of the session.

Hotspots

Section titled “Hotspots”

The leaves of the call tree with the most time: regions without nested regions, and the own time parents spend outside theirs. Shares are of the session. Click one to find it in the table.

Region statistics

Section titled “Region statistics”
region% sessiontotal [s]gpu total [s]gpu avg [s]trend
└─ (own) 4.63% 0.016000
│ └─ (own) 0.00% 0.000000

Durations are in seconds. (own) rows show a region's time excluding its children. % session uses wall-clock coverage; overlapping recursive calls count once. Click a row for its call site and per-rank breakdown.

Load balance

Section titled “Load balance”

Each region's own time per rank, excluding nested regions, so a parent does not repeat its children's imbalance. Excess is how much longer the slowest rank spends than the average rank: the most balancing that region could save. Regions are ordered by it.

regionmeanminmaxslowest rankimbalanceexcessmin · mean · max
solve69.4 ms67.2 ms72.5 ms2+4%3.12 ms
setup4 ms4 ms4 ms0+0%0 s
copy_to_device3 ms3 ms3 ms0+0%0 s
exchange6 ms6 ms6 ms0+0%0 s

Rank comparison

Section titled “Rank comparison”

Each cell compares one rank's time in a region with the region's mean over the ranks. Rank 2, in red in the header, spends the most time in regions overall. Hover a cell for its time.

faster than the meanslower

regionmean0123
solve69.4 ms−2.6%+1.2%+4.5%−3.2%
exchange6 ms0%0%0%0%
setup4 ms0%0%0%0%
copy_to_device3 ms0%0%0%0%
all regions82.4 ms−2.2%+1.0%+3.8%−2.7%

Charts

Section titled “Charts”
Region durations

Each bar shows a region's total recorded duration, summed over the selected ranks, longest first.

Timeline: GPU node (rank 0)

Each bar is one recorded region call on rank 0. Its position and width show when the call started and how long it ran; colors identify regions.

Rank imbalance

Each line compares a region's total duration by rank; its dashed line marks the mean across ranks. Points far from the mean identify stragglers.

Rank heatmap

This heatmap uses exclusive timings. Exclusive duration is the time spent in a region itself, excluding time spent in nested child regions; this prevents the enclosing session region from dominating the heatmap.

Figures

Section titled “Figures”

Files

Section titled “Files”

Download run.py (715 B)

"""Landau damping of a Langmuir wave with a Vlasov-Ampere solver: the script both runs use."""


import sys


import yaml
from scope_profiler import ProfileManager




def main(parameter_file):
    with open(parameter_file) as handle:
        parameters = yaml.safe_load(handle)
    with ProfileManager.session(label=parameters["label"]):
        with ProfileManager.region("setup"):
            state = initialise(parameters)
        for _ in range(parameters["steps"]):
            with ProfileManager.region("solve"):
                state = advance(state, parameters["dt"])
            with ProfileManager.region("exchange"):
                exchange_halos(state)




if __name__ == "__main__":
    main(sys.argv[1])

Metadata

Section titled “Metadata”
commit0000000
systemsynthetic
chip_informationsynthetic 4-GPU node
mpi_size4
num_nodes1
omp_num_threads1
platformsynthetic
scope_profiler_versionsample
timestamp2026-10-17T12:00:00Z
total_cores4