SEP 40 default overhead: edge reads cost more than 2 % on short line runs and on native Cartesian

オープン
  • 何か問題がありますか？
#396 コメント 0 件 リアクション 0 件 担当者 0 名 GitHub で見る

メンテナーはふだん 1 日以内に返信

まだ誰も着手していません。

評価

難易度
5/5
見積もり時間
1週間以上
初心者へのやさしさ
8/100
issue の種類
バグ
明瞭さ
説明が足りない
活発さ
活発
技術スタック
python
領域
performance

調査の方向性

The overhead is measured in the native backend's march and its edge reads, the k⊥ ramp and the Cartesian filament case, and the payload names no file or test. Start by profiling a native Cartesian read on Linux, including the march's return and its state copy, and compare against the ifftn and einsum costs listed in the table. Done would mean the per-read host cost is explained, though the maintainer has already recorded this as a miss and moved to Final.

索引モデルが issue の本文から書いたものです。

説明

area: boundaries performance priority: medium

SEP 40 §7.7 sets a 2 % limit on the default's cost (the k⊥ safety ramp plus the reads of the transverse edge) on clean CPU runs. It was measured at the end of the SEP: Xeon Gold 6240, one socket, 18 threads, same process, variants interleaved, medians of six rounds.

backend case steps (no reads → reads) reads ramp edge both
native line 27 → 28 19 −0.1 % +14.4 % +14.0 %
native planar 40 → 40 24 −9.9 % +10.8 % 0.0 %
native radial_filament_256 127 → 127 19 +0.1 % +2.1 % +2.2 %
native cartesian_filament_128 (warned, half length) 40 → 41 20 +0.6 % +10.1 % +10.5 %
numba_cpu line 27 → 28 19 +0.2 % +5.9 % +6.4 %
numba_cpu planar 40 → 40 24 −1.5 % +0.9 % −1.1 %
numba_cpu radial_filament_256 127 → 127 19 0.0 % −0.2 % −0.1 %
numba_cpu cartesian_filament_128 (warned, half length) 40 → 41 20 +0.4 % +3.0 % +2.6 %

The ramp costs nothing. The remaining cost of reading has two parts:

  1. The read ceiling. It adds one step on the line case (27 → 28, +3.7 %) and on the Cartesian filament. Steps are capped at the read spacing, min(L/16, 4g/v_q), and that cap binds on some steps.
  2. Per-read host cost.
    • The line case is a 0.2 s native run, so reads at about 1 ms each are 10 % of it.
    • On native Cartesian, each read costs about 180 ms on the Xeon Gold 6240. The read's own pieces should cost about 70 ms: an 18-thread ifftn of 128² × 433 in 12 ms, einsum sums of 9 ms, and copies.
    • On an Apple M5 Max, the same read costs about 13 ms.
    • Neither OMP_WAIT_POLICY=passive nor one FFT worker changed it, so it is not thread contention. The remainder is probably the native march's return and re-entry around a read, or the host copy of the state; it has not been profiled on Linux.

Leads:

  • profile a native Cartesian read on Linux, including the march's return and its state copy;
  • let a native backend reduce the edge's faces where its state lives, as ADR-057 does for the temporal monitor;
  • consider whether L/16 should bind on cases whose light stays far from the edge.

From SEP 40's phase 3 report (#71). The maintainer's decision (2026-11-01): record a miss and proceed to Final.

主要言語
Python
スター
3
フォーク
3
平均マージ
6時間 51分
マージ済み PR（30日）
136

環境構築

はじめの一歩

  1. issue を最後まで読み、次にプロジェクトのコントリビューションガイドを読みます。
  2. 着手することを issue にコメントします — 二人が同じ作業をするのを防げます。
  3. リポジトリをフォークし、ブランチを切って変更します。
  4. issue 番号を参照したプルリクエストを送ります。

LupoLab/solprop のほかの issue

LupoLab/solprop の issue をすべて見る

似ている issue

Python の issue をもっと見る

新しい issue をメールで受け取る

初心者向けの GitHub issue を短くまとめたダイジェスト。