Push vpnykuuvnrnx #82

Merged
coissac merged 11 commits from push-vpnykuuvnrnx into main 2026-11-19 06:17:44 +00:00
11 changed files with 497 additions and 2 deletions
Showing only changes of commit 03eaa4544a - Show all commits

No files matched your search

+61
View File
@@ -0,0 +1,61 @@
# Prompt — où optimiser l'étape `pack` du merge (diagnostic d'abord)
> À coller tel quel au début d'une nouvelle session Claude Code ouverte à la racine du dépôt obikmer.
> Cette session est un **diagnostic** : elle produit une analyse chiffrée et des pistes classées, pas du code. Toute modification du code attend mon feu vert explicite, piste par piste.
---
## Contexte
`obikmer merge` se termine par une étape `finalize`, dont l'essentiel est `pack` : les matrices de présence/comptage, écrites en colonnes pendant le merge (`MatrixBuilder`), sont transposées en un format compact par partition (sparse par défaut). Le point d'entrée est `KmerIndex::finalize_indexed` (`src/obikindex/src/index/builder.rs`, étape « pack » vers la ligne 109), qui appelle `KmerIndex::pack_matrices(sparse)` (`src/obikindex/src/index/kmer_index.rs`). Les primitives sont dans `src/obicompactvec/src/` : `pack_bit_matrix` (`bitmatrix/packed.rs`), `pack_compact_int_matrix` (`intmatrix.rs`), `pack_sparse_compact_int_matrix` (`sparse_intmatrix.rs`), et le côté sparse des bits dans `bitmatrix/`.
Lis d'abord `DevDocMD/implementation/merge_scaling.md` (état du pack sparse : passage par une matrice dense séquentielle puis sparse, un seul mapping par partition) et `DevDocMD/implementation/merge_mmap_audit.md`.
## Mesure de référence (production, 2026-11-18)
Merge de 186 sources (`Cedrus`), 256 partitions, binaire glibc compilé localement, **192 CPU non contraints**, 17,5 milliards de k-mers nouveaux au total. Mode **présence**, compaction par défaut, donc **sparse** (pas de `--dense`) :
| Étape | Durée | Efficacité CPU | RSS max |
|---|---|---|---|
| `merge_partitions` | 748 s | 141×/192 (73 %) | 354 Go |
| `pack` | 231 s | 65×/192 (34 %) | **503,7 Go** |
La RAM de la machine est de 754 Go. Le RSS max est celui du process : il inclut les pages de fichiers mappés, pas seulement la mémoire anonyme. **Je n'ai pas encore séparé les deux.**
Progression de `pack` (partitions terminées par tranche de 10 %, début à 04:08:09) :
`+53 s` (10 %) · `+35 s` (20 %) · `+13 s` (30 %) · `+48 s` (40 %) · `+24 s` (50 %) · `+9 s` · `+6 s` · `+6 s` · `+6 s` · `+11 s` (100 %), puis `+20 s` avant la fin de l'étape. Les partitions les plus grosses passent en premier (tri décroissant dans `merge.rs`), mais je ne sais pas si `pack` les traite dans le même ordre ni si l'ordre est le bon.
## Ce que je veux savoir
1. **Où part le temps** de `pack` ? Lecture des colonnes, transposition, écriture, `sync`/`fsync`, attente d'un verrou, déséquilibre entre partitions ? Efficacité de 34 % sur 192 workers : qu'est-ce qui les bloque ?
2. **Où part la mémoire** (504 Go) ? Combien est de la mémoire anonyme, combien des pages de fichiers mappés ? Est-elle proportionnelle au nombre de workers ? Quelle est la mémoire vivante par partition en cours de `pack` ?
3. **Quel est le parallélisme réel** : combien de partitions `pack` traite-t-il en même temps, et qu'est-ce qui le fixe (`PartitionRunner`, `--cpu-max`, NUMA) ?
4. Y a-t-il du **travail inutile** (relectures, copies intermédiaires, colonnes vides ou toutes-absentes ré-écrites, étape dense qui pourrait être évitée) ?
## Méthode (impérative)
- **Lire et mesurer avant de proposer.** Pas de récit plausible sans preuve : tout ce qui n'est pas vérifié par lecture du code ou par mesure est marqué « hypothèse ».
- Pour mesurer sans la production : `scripts/docker_debug/` fournit un conteneur Linux (build glibc ou musl) qui lance un merge de N sources synthétiques ou dérivées de génomes réels (`run.sh`, variables `GEN`, `REAL`, `GENOME_LEN`, `TARGET`, `TOOL=heaptrack` pour glibc) avec un moniteur CSV de mappings et de RSS. Les génomes de 2 Mb donnent des partitions assez grosses pour que `pack` ait quelque chose à faire. `heaptrack` fonctionne sur le build glibc (pas sur le build musl statique). `gdb` ne fonctionne pas sous Rosetta.
- Pour séparer mémoire anonyme et fichiers mappés, utilise `/proc/PID/status` (`RssAnon`, `RssFile`, `RssShmem`), **jamais `smaps` en entier** sur un process à beaucoup de mappings (il verrouille `mmap_lock` et ralentit le process).
- Les estimations de coût (octets, secondes) sont marquées **estimées** ; les chiffres observés sont marqués **mesurés**.
## Règles du projet (non négociables)
- Avant tout travail de code : `mcp__serena__initial_instructions`, puis Serena (`find_symbol`, `get_symbols_overview`, `find_referencing_symbols`) pour naviguer. `rust-analyzer-mcp` et `qmd` quand ils servent. Consulte `context7` avant d'utiliser une API de bibliothèque externe.
- **Aucune commande VCS mutante** (git/jj : commit, add, mv, rm, describe, push…). Lecture seule autorisée.
- **Aucune nouvelle dépendance, aucun changement de version** sans mon accord explicite. Si un problème de bibliothèque se pose, expose-le et propose des alternatives.
- **Une question appelle une réponse, pas une action.** Ne modifie aucun fichier de `src/` sans demande explicite. Cette session est en lecture seule sur le code.
- Le code existant est une hypothèse, pas une vérité : signale tout ce qui est devenu injustifié plutôt que de le garder par inertie.
- Si tu proposes d'ajouter une opération à une structure partagée, mets-la comme méthode générique sur cette structure (crate `obicompactvec`), pas comme fonction libre dans le consommateur.
- Ne jamais matérialiser en mémoire une structure volumineuse qui peut être streamée.
- Échanges en **français**, documentation en **anglais**.
## Livrable
Un rapport (dans la conversation, pas dans un fichier, sauf si je le demande) avec :
1. Le chemin de code de `pack`, de `finalize_indexed` jusqu'aux écritures disque, résumé en quelques lignes par étape.
2. Pour chaque question ci-dessus : ce qui est **mesuré**, ce qui est **hypothèse**, et comment vérifier l'hypothèse.
3. Une liste de **pistes d'optimisation classées** par gain attendu/coût/risque. Pour chaque piste : ce qu'elle changerait, ce qu'elle risque, et comment on verrait qu'elle marche. Aucune implémentation.
4. Les vérifications de non-régression à prévoir : le résultat sur disque doit rester **identique octet pour octet** (voir les oracles `scripts/smoke_test_merge.sh` et `scripts/bench_merge_oracle.sh`).
Termine par la question : « Quelle piste veux-tu que j'implémente ? » et attends ma réponse.
+16
View File
@@ -0,0 +1,16 @@
# Linux debug environment for the merge mapping/memory-growth hunt.
# Same toolchain as .gitea/workflows/release.yml for the musl target
# (cargo-zigbuild + zig cc), plus a glibc toolchain and diagnostic tools.
FROM rust:1-bookworm
RUN apt-get update && apt-get install -y --no-install-recommends \
build-essential cmake pkg-config autoconf automake libtool \
heaptrack valgrind gdb strace procps python3 python3-pip \
&& rm -rf /var/lib/apt/lists/*
RUN pip install ziglang --quiet --break-system-packages \
&& cargo install cargo-zigbuild \
&& rustup target add x86_64-unknown-linux-musl
RUN ZIG=$(python3 -c "import ziglang, os; print(os.path.join(os.path.dirname(ziglang.__file__), 'zig'))") \
&& printf '#!/bin/sh\nexec "%s" cc -target x86_64-linux-musl "$@"\n' "$ZIG" > /usr/local/bin/x86_64-linux-musl-gcc \
&& printf '#!/bin/sh\nexec "%s" c++ -target x86_64-linux-musl "$@"\n' "$ZIG" > /usr/local/bin/x86_64-linux-musl-g++ \
&& chmod +x /usr/local/bin/x86_64-linux-musl-gcc /usr/local/bin/x86_64-linux-musl-g++
WORKDIR /work
+6
View File
@@ -0,0 +1,6 @@
#!/usr/bin/env bash
# Builds the debug image. PLATFORM defaults to linux/amd64 (Rosetta on Apple Silicon).
set -euo pipefail
cd "$(dirname "${BASH_SOURCE[0]}")"
PLATFORM="${PLATFORM:-linux/amd64}"
docker build --platform "$PLATFORM" -t "obikmer-debug:${PLATFORM//\//-}" .
+33
View File
@@ -0,0 +1,33 @@
#!/usr/bin/env python3
"""gen_fasta.py N OUTDIR [GENOME_LEN] — N deterministic synthetic genomes.
Each genome = unique random sequence + a random subset of a shared fragment
pool (some fragments lightly mutated), so a merge both hits k-mers already
in the destination and creates new ones, like related genomes would.
"""
import random, sys
n, out = int(sys.argv[1]), sys.argv[2]
glen = int(sys.argv[3]) if len(sys.argv) > 3 else 150_000
seed = int(sys.argv[4]) if len(sys.argv) > 4 else 20261118
rng = random.Random(seed)
def rnd(m):
return "".join(rng.choices("ACGT", k=m))
def mutate(s, rate=0.01):
s = list(s)
for i in range(len(s)):
if rng.random() < rate:
s[i] = rng.choice("ACGT")
return "".join(s)
pool = [rnd(2000) for _ in range(300)] # 600 kb shared pool
for g in range(n):
parts = [rnd(glen)]
for f in rng.sample(range(len(pool)), 60):
parts.append(mutate(pool[f]) if rng.random() < 0.3 else pool[f])
rng.shuffle(parts)
with open(f"{out}/G{g:03d}.fasta", "w") as fh:
for j, p in enumerate(parts):
fh.write(f">G{g:03d}_{j}\n{p}\n")
+39
View File
@@ -0,0 +1,39 @@
#!/usr/bin/env python3
"""gen_fasta_real.py N OUTDIR GENOMES_DIR [WINDOW] [SEED] [MUT]
N derived genomes built from *real* genomes (GENOMES_DIR/*.fna.gz): each one is
a random window of WINDOW bp of a randomly chosen real genome, point-mutated at
rate MUT (default 3 %, which alters ~47 % of 21-mers), so a merge both finds
k-mers already present in the destination and creates new ones, with realistic
sequence statistics (no dedicated degenerate structure).
"""
import glob, gzip, random, sys
n, out, gdir = int(sys.argv[1]), sys.argv[2], sys.argv[3]
window = int(sys.argv[4]) if len(sys.argv) > 4 else 300_000
seed = int(sys.argv[5]) if len(sys.argv) > 5 else 20261118
mut = float(sys.argv[6]) if len(sys.argv) > 6 else 0.03
rng = random.Random(seed)
def read_fna(path):
seq = []
with gzip.open(path, "rt") as fh:
for line in fh:
if not line.startswith(">"):
seq.append(line.strip().upper())
return "".join(seq)
genomes = [read_fna(p) for p in sorted(glob.glob(f"{gdir}/*.fna.gz"))]
genomes = [g for g in genomes if len(g) > window]
for i in range(n):
g = rng.choice(genomes)
start = rng.randrange(0, len(g) - window)
s = list(g[start:start + window])
for j in range(len(s)):
if s[j] in "ACGT" and rng.random() < mut:
s[j] = rng.choice("ACGT")
seq = "".join(c if c in "ACGT" else "A" for c in s)
with open(f"{out}/R{i:03d}.fasta", "w") as fh:
fh.write(f">R{i:03d}\n")
for k in range(0, len(seq), 80):
fh.write(seq[k:k + 80] + "\n")
+11
View File
@@ -0,0 +1,11 @@
#!/usr/bin/env bash
# monitor.sh PID OUT [INTERVAL_S] — samples mappings / VM / commit figures until PID exits.
PID=$1; OUT=$2; DT=${3:-5}
printf 'time,maps,vmsize_kb,vmdata_kb,vmrss_kb,threads,committed_as_kb,commit_limit_kb\n' >"$OUT"
while kill -0 "$PID" 2>/dev/null; do
maps=$(wc -l <"/proc/$PID/maps" 2>/dev/null) || break
st=$(grep -E '^(VmSize|VmData|VmRSS|Threads):' "/proc/$PID/status" | awk '{printf "%s,", $2}')
mi=$(grep -E '^(Committed_AS|CommitLimit):' /proc/meminfo | awk '{printf "%s,", $2}')
echo "$(date +%T),$maps,${st}${mi%,}" >>"$OUT"
sleep "$DT"
done
+92
View File
@@ -0,0 +1,92 @@
#!/usr/bin/env bash
# run.sh — builds obikmer (release, debug symbols) and merges the benchmark presence
# indexes in the container while sampling mappings/RSS/commit in parallel.
#
# Env:
# PLATFORM linux/amd64 (default, Rosetta) | linux/arm64
# GEN if set: build GEN synthetic indexes (gen_fasta.py) instead of using
# the benchmark ones; GENOME_LEN (default 150000), K (21), M (9), PART (256)
# REAL if set (with GEN): derive genomes from benchmark/run/genomes (gen_fasta_real.py)
# SRC_DIR alternative source tree (default: <repo>/src); VARIANT: volume suffix;
# LOCKED=0 to allow Cargo.lock changes (needed when a dep is added)
# NSRC number of source indexes to merge (default: all 20)
# TARGET gnu (default, glibc) | musl (static, same chain as the release CI)
# TOOL none (default) | heaptrack (heaptrack output goes to OUT/)
# MAXMAP if set, vm.max_map_count is lowered to this inside the VM for the
# run and restored afterwards (needs --privileged; affects the whole
# OrbStack VM while it runs)
# MALLOC_ARENA_MAX, MALLOC_MMAP_THRESHOLD_, MALLOC_TRIM_THRESHOLD_ passed through
# OUT host output directory (default: ./out/<timestamp>)
set -euo pipefail
HERE="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
REPO="$(cd "$HERE/../.." && pwd)"
PLATFORM="${PLATFORM:-linux/amd64}"
NSRC="${NSRC:-20}"
TOOL="${TOOL:-none}"
TARGET="${TARGET:-gnu}"
OUT="${OUT:-$HERE/out/$(date +%Y%m%d-%H%M%S)}"
mkdir -p "$OUT"
IMG="obikmer-debug:${PLATFORM//\//-}"
VOL="obikmer-debug-target-${PLATFORM//\//-}-${TARGET}${VARIANT:+-$VARIANT}"
docker run --rm --privileged --platform "$PLATFORM" \
-v "${SRC_DIR:-$REPO/src}:/work/src:$([ -n "${SRC_DIR:-}" ] && echo rw || echo ro)" \
-v "$REPO/benchmark/run/specimen_index_presence:/idx_ro:ro" \
-v "$VOL:/target" \
-v "$OUT:/out" \
-v "$HERE/monitor.sh:/monitor.sh:ro" \
-v "$HERE/gen_fasta.py:/gen_fasta.py:ro" \
-v "$HERE/gen_fasta_real.py:/gen_fasta_real.py:ro" \
-v "$REPO/benchmark/run/genomes:/genomes_ro:ro" \
-e CARGO_TARGET_DIR=/target -e CARGO_PROFILE_RELEASE_DEBUG=1 \
-e GEN="${GEN:-}" -e LOCKED_FLAG="$([ "${LOCKED:-1}" = 0 ] || echo --locked)" -e SEED="${SEED:-20261118}" -e REAL="${REAL:-}" -e GENOME_LEN="${GENOME_LEN:-150000}" -e K="${K:-21}" -e M="${M:-9}" -e PART="${PART:-256}" -e RUST_LOG="${RUST_LOG:-}" -e NSRC="$NSRC" -e TARGET="$TARGET" -e TOOL="$TOOL" -e MAXMAP="${MAXMAP:-}" \
-e MALLOC_ARENA_MAX="${MALLOC_ARENA_MAX:-}" \
-e CXXFLAGS="${CXXFLAGS:-}" \
-e MALLOC_MMAP_THRESHOLD_="${MALLOC_MMAP_THRESHOLD_:-}" \
-e MALLOC_TRIM_THRESHOLD_="${MALLOC_TRIM_THRESHOLD_:-}" \
"$IMG" bash -c '
set -euo pipefail
for v in MALLOC_ARENA_MAX MALLOC_MMAP_THRESHOLD_ MALLOC_TRIM_THRESHOLD_ RUST_LOG; do
[ -n "${!v}" ] || unset $v
done
if [ "$TARGET" = musl ]; then
PKG_CONFIG_ALLOW_CROSS=1 cargo zigbuild --release $LOCKED_FLAG --target x86_64-unknown-linux-musl \
--manifest-path /work/src/Cargo.toml --bin obikmer 2>&1 | tail -3
BIN=/target/x86_64-unknown-linux-musl/release/obikmer
else
cargo build --release $LOCKED_FLAG --manifest-path /work/src/Cargo.toml --bin obikmer 2>&1 | tail -3
BIN=/target/release/obikmer
fi
file "$BIN" 2>/dev/null | cut -c1-150 || true
# merge writes into its sources (lock files…): work on a copy, never on benchmark/
rm -rf /tmp/idx && mkdir -p /tmp/idx
if [ -n "$GEN" ]; then
mkdir -p /tmp/fa
if [ -n "$REAL" ]; then python3 /gen_fasta_real.py "$GEN" /tmp/fa /genomes_ro "$GENOME_LEN" "$SEED"
else python3 /gen_fasta.py "$GEN" /tmp/fa "$GENOME_LEN" "$SEED"; fi
NSRC=$GEN
echo "indexing $GEN synthetic genomes (k=$K m=$M p=$PART)…"
ls /tmp/fa/*.fasta | xargs -P 8 -I{} sh -c \
"b=\$(basename {} .fasta); $BIN index -k $K -m $M --theta 0 -p $PART --label \$b -o /tmp/idx/\$b.idx {} >/dev/null 2>&1 || echo index failed: \$b"
else
cp -r /idx_ro/. /tmp/idx/
fi
mapfile -t SRC < <(find /tmp/idx -mindepth 1 -maxdepth 1 -type d | sort | head -n "$NSRC")
echo "sources: ${#SRC[@]} target: $TARGET tool: $TOOL max_map_count: $(cat /proc/sys/vm/max_map_count)"
if [ -n "$MAXMAP" ]; then
OLD=$(cat /proc/sys/vm/max_map_count); sysctl -w vm.max_map_count="$MAXMAP" >/dev/null
trap "sysctl -w vm.max_map_count=$OLD >/dev/null" EXIT
fi
rm -rf /tmp/merged
CMD=("$BIN" merge --output /tmp/merged --force --force-presence "${SRC[@]}")
if [ "$TOOL" = heaptrack ]; then CMD=(heaptrack -o /out/heaptrack "${CMD[@]}"); fi
"${CMD[@]}" 2>/out/merge.stderr &
PID=$!
/monitor.sh "$PID" /out/monitor.csv 1 &
MON=$!
set +e; wait "$PID"; RC=$?; set -e
wait "$MON" || true
echo "merge exit code: $RC" | tee /out/exit_code.txt
tail -5 /out/merge.stderr
'
echo "outputs: $OUT"
+102
View File
@@ -0,0 +1,102 @@
#!/usr/bin/env bash
# monitor_pack.sh — 1 Hz sampler for I/O- and memory-bound stages (pack, merge_partitions).
#
# Usage:
# scripts/monitor_pack.sh --pid PID OUT_DIR [OUTPUT_INDEX_DIR] # attach to a running process
# scripts/monitor_pack.sh --cmd OUT_DIR [OUTPUT_INDEX_DIR] -- CMD... # start CMD and sample it
# env: DT=1 (seconds between samples)
#
# Writes OUT_DIR/samples.csv (one row per sample) and OUT_DIR/env.txt (context
# needed to read it). Linux only. Never reads smaps (it takes mmap_lock and
# slows a process with many mappings); only /proc/PID/{status,io,task/*/stat}
# and /proc/meminfo.
#
# Columns of samples.csv
# t epoch seconds
# rss_anon_kb, rss_file_kb, rss_shmem_kb from /proc/PID/status
# threads thread count
# n_R n_D n_S threads running / in uninterruptible sleep (I/O wait) / sleeping
# (futex, channel, sleep: NOT disk). n_D is the direct count of
# workers blocked on I/O.
# maps number of mappings (wc -l /proc/PID/maps)
# read_bytes write_bytes /proc/PID/io (cumulative, bytes submitted to the
# block layer; buffered writes counted at page dirtying)
# rchar wchar /proc/PID/io (cumulative, bytes through read()/write() syscalls)
# dirty_kb writeback_kb cached_kb mem_avail_kb /proc/meminfo
#
# Pair it with the process's own log (RUST_LOG=info, plus obisys=debug for the
# "activated N worker(s)" lines and obikindex=debug for per-partition timings)
# to align stages on wall-clock time: the log timestamps are UTC ISO-8601, `t`
# is epoch seconds.
set -euo pipefail
MODE=${1:-}; shift || true
case "$MODE" in
--pid) PID=${1:?PID}; shift ;;
--cmd) ;;
*) sed -n 2,12p "$0"; exit 2 ;;
esac
OUT=${1:?OUT_DIR}; shift
IDX=""
if [ "${1:-}" != "" ] && [ "${1:-}" != "--" ]; then IDX=$1; shift; fi
mkdir -p "$OUT"
if [ "$MODE" = "--cmd" ]; then
[ "${1:-}" = "--" ] || { echo "--cmd needs: -- CMD..." >&2; exit 2; }
shift
"$@" 2>"$OUT/cmd.stderr" &
PID=$!
fi
{
echo "date_utc: $(date -u +%FT%TZ)"
echo "pid: $PID"
echo "nproc: $(nproc)"
echo "kernel: $(uname -r)"
grep -E 'MemTotal|SwapTotal' /proc/meminfo
for k in vm.dirty_ratio vm.dirty_background_ratio vm.dirty_bytes vm.dirty_background_bytes \
vm.dirty_expire_centisecs vm.dirty_writeback_centisecs vm.max_map_count; do
echo "$k: $(sysctl -n $k 2>/dev/null || echo '?')"
done
echo "ulimit -n: $(ulimit -n)"
echo "cmdline: $(tr '\0' ' ' </proc/$PID/cmdline 2>/dev/null)"
if [ -n "$IDX" ]; then
echo "output_dir: $IDX"
echo "fs_type: $(stat -f -c %T "$IDX" 2>/dev/null || echo '?')"
echo "mount: $(df -P "$IDX" 2>/dev/null | tail -1)"
command -v beegfs-ctl >/dev/null && echo "beegfs: $(beegfs-ctl --getentryinfo "$IDX" 2>&1 | head -8 | tr '\n' ' ')"
fi
env | grep -E '^(MALLOC_|RUST_LOG|RAYON_)' || true
} >"$OUT/env.txt"
DT=${DT:-1}
printf 't,rss_anon_kb,rss_file_kb,rss_shmem_kb,threads,n_R,n_D,n_S,maps,read_bytes,write_bytes,rchar,wchar,dirty_kb,writeback_kb,cached_kb,mem_avail_kb\n' >"$OUT/samples.csv"
while [ -d "/proc/$PID" ]; do
t=$(date +%s)
st=$(awk '/^RssAnon:/{a=$2} /^RssFile:/{f=$2} /^RssShmem:/{s=$2} /^Threads:/{th=$2} END{printf "%s,%s,%s,%s",a,f,s,th}' "/proc/$PID/status" 2>/dev/null) || break
# thread states: field 3 of /proc/PID/task/TID/stat, after the ")" closing comm
states=$(cat /proc/$PID/task/*/stat 2>/dev/null | sed -E 's/^.*\) (.).*/\1/' | sort | uniq -c \
| awk '{c[$2]=$1} END{printf "%d,%d,%d", c["R"]+0, c["D"]+0, c["S"]+0}')
maps=$(wc -l <"/proc/$PID/maps" 2>/dev/null || echo 0)
io=$(awk '/^rchar:/{rc=$2} /^wchar:/{wc=$2} /^read_bytes:/{r=$2} /^write_bytes:/{w=$2} END{printf "%s,%s,%s,%s",r,w,rc,wc}' "/proc/$PID/io" 2>/dev/null || echo ",,,")
mi=$(awk '/^Dirty:/{d=$2} /^Writeback:/{w=$2} /^Cached:/{c=$2} /^MemAvailable:/{m=$2} END{printf "%s,%s,%s,%s",d,w,c,m}' /proc/meminfo)
echo "$t,$st,$states,$maps,$io,$mi" >>"$OUT/samples.csv"
sleep "$DT"
done
if [ "$MODE" = "--cmd" ]; then wait "$PID" || echo "command exit code: $?" >>"$OUT/env.txt"; fi
# ── summary ──────────────────────────────────────────────────────────────────
awk -F, 'NR==1{next}
{ n++; if(NR==2){t0=$1; r0=$10; w0=$11}
if($2>ma)ma=$2; if($3>mf)mf=$3; if($4>ms)ms=$4; if($5>mt)mt=$5
if($6>mR)mR=$6; if($7>mD)mD=$7; if($8>mS)mS=$8; sR+=$6; sD+=$7; sS+=$8
if($14>md)md=$14; if($15>mw)mw=$15; if($9>mm)mm=$9
t1=$1; r1=$10; w1=$11 }
END{ if(!n)exit
printf "samples: %d over %ds\n", n, t1-t0
printf "peak RssAnon %.1f GB | RssFile %.1f GB | RssShmem %.1f GB | threads %d | maps %d\n", ma/1e6, mf/1e6, ms/1e6, mt, mm
printf "threads by state (mean / max): R %.1f / %d | D(I/O wait) %.1f / %d | S %.1f / %d\n", sR/n, mR, sD/n, mD, sS/n, mS
printf "peak Dirty %.1f GB | Writeback %.1f GB\n", md/1e6, mw/1e6
printf "block I/O over the run: read %.1f GB, write %.1f GB\n", (r1-r0)/1e9, (w1-w0)/1e9 }' "$OUT/samples.csv" | tee "$OUT/summary.txt"
+53
View File
@@ -0,0 +1,53 @@
#!/usr/bin/env bash
# watch_merge.sh PID OUT.csv — samples a running process and kills it past a threshold.
#
# Every INTERVAL seconds appends one CSV row: number of memory mappings
# (lines of /proc/PID/maps), VmSize, VmData, VmRSS, threads, and the
# system-wide Committed_AS / CommitLimit. If the mapping count exceeds
# MAPS_MAX, VmRSS exceeds RSS_MAX_GB, or VmSize exceeds VMSIZE_MAX_GB (stay
# below `ulimit -v`: past it, allocations fail instead of the guard firing), the process receives SIGTERM (then
# SIGKILL after 30 s) and the reason is written to OUT.csv.reason — a guard
# so a runaway merge ends with a clean message instead of a failed
# allocation (or the OOM killer on a shared node).
#
# Reading /proc/PID/maps is itself slow when the process has hundreds of
# thousands of mappings (it held the kernel's mmap lock in earlier
# measurements): keep INTERVAL >= 5. It never reads /proc/PID/smaps.
#
# Usage: scripts/watch_merge.sh PID OUT.csv
# env: INTERVAL=5 MAPS_MAX=500000 RSS_MAX_GB=400 VMSIZE_MAX_GB=440
#
# Typical use (Linux only):
# nohup scripts/watch_merge.sh "$(pgrep -f 'obikmer merge' | head -1)" /tmp/watch.csv &
set -u
PID=${1:?usage: watch_merge.sh PID OUT.csv}
OUT=${2:?usage: watch_merge.sh PID OUT.csv}
INTERVAL=${INTERVAL:-5}
MAPS_MAX=${MAPS_MAX:-500000}
RSS_MAX_KB=$(( ${RSS_MAX_GB:-400} * 1024 * 1024 ))
VMSIZE_MAX_KB=$(( ${VMSIZE_MAX_GB:-440} * 1024 * 1024 ))
[ -r "/proc/$PID/maps" ] || { echo "cannot read /proc/$PID/maps" >&2; exit 1; }
printf 'time,maps,vmsize_kb,vmdata_kb,vmrss_kb,threads,committed_as_kb,commit_limit_kb\n' >"$OUT"
while kill -0 "$PID" 2>/dev/null; do
maps=$(wc -l <"/proc/$PID/maps" 2>/dev/null) || break
st=$(grep -E '^(VmSize|VmData|VmRSS|Threads):' "/proc/$PID/status" 2>/dev/null | awk '{printf "%s,", $2}')
rss=$(grep '^VmRSS:' "/proc/$PID/status" 2>/dev/null | awk '{print $2}')
vsz=$(grep '^VmSize:' "/proc/$PID/status" 2>/dev/null | awk '{print $2}')
mi=$(grep -E '^(Committed_AS|CommitLimit):' /proc/meminfo | awk '{printf "%s,", $2}')
echo "$(date +%T),$maps,${st}${mi%,}" >>"$OUT"
reason=""
[ "$maps" -gt "$MAPS_MAX" ] && reason="mappings $maps > MAPS_MAX $MAPS_MAX"
[ -n "${rss:-}" ] && [ "$rss" -gt "$RSS_MAX_KB" ] && reason="VmRSS ${rss} kB > RSS_MAX_GB"
[ -n "${vsz:-}" ] && [ "$vsz" -gt "$VMSIZE_MAX_KB" ] && reason="VmSize ${vsz} kB > VMSIZE_MAX_GB"
if [ -n "$reason" ]; then
echo "$(date +%T) GUARD: $reason — sending SIGTERM to $PID" | tee "$OUT.reason" >&2
kill -TERM "$PID" 2>/dev/null
for _ in $(seq 30); do kill -0 "$PID" 2>/dev/null || exit 0; sleep 1; done
kill -KILL "$PID" 2>/dev/null
exit 0
fi
sleep "$INTERVAL"
done
+76 -2
View File
@@ -296,7 +296,14 @@ impl KmerIndex {
};
let n = self.n_partitions();
let order: Vec<usize> = (0..n).collect();
// Largest partitions first (LFD), so the long jobs do not end up as
// stragglers at the tail. Sizes are `stat()` only; the sort is stable,
// ties keep partition-number order. Partitions are independent, so
// the order has no effect on what is written.
let sizes: Vec<u64> = (0..n).map(|i| self.pending_matrix_bytes(i)).collect();
let mut order: Vec<usize> = (0..n).collect();
order.sort_by_key(|&i| std::cmp::Reverse(sizes[i]));
let mut timings: Vec<(usize, u64, std::time::Duration)> = Vec::with_capacity(n);
let pb = progress_bar("pack", n as u64, "partitions");
obisys::PartitionRunner::new().run(
&order,
@@ -327,14 +334,50 @@ impl KmerIndex {
}
Ok(())
},
|_, _, _| {
|i, _, elapsed| {
tracing::debug!(
partition = i,
matrix_bytes = sizes[i],
secs = elapsed.as_secs_f64(),
"pack: partition done"
);
timings.push((i, sizes[i], elapsed));
pb.inc(1);
},
)?;
pb.finish_and_clear();
report_pack_timings(&timings);
Ok(())
}
/// Bytes of matrix data (`presence/` and `counts/` of every layer) still
/// present in partition `i`, from `stat()` only. Before packing this is
/// the columnar data, i.e. a proxy for the work `pack_matrices` has to do
/// on that partition. Missing partition or directories count as 0.
pub fn pending_matrix_bytes(&self, i: usize) -> u64 {
let Ok(n_layers) = self.n_layers(i) else {
return 0;
};
let mut total = 0u64;
for l in 0..n_layers {
let Ok(layer_dir) = self.layer_dir(i, l) else {
continue;
};
for sub in ["presence", "counts"] {
let Ok(entries) = fs::read_dir(layer_dir.join(sub)) else {
continue;
};
total += entries
.flatten()
.filter_map(|e| e.metadata().ok())
.filter(|m| m.is_file())
.map(|m| m.len())
.sum::<u64>();
}
}
total
}
// /// Write a `layer_meta.json` in any layer directory that is missing one.
// ///
// /// Old indexes were built before this file was required. The number of
@@ -394,3 +437,34 @@ impl KmerIndex {
// Ok(())
// }
}
/// One `info!` summary of `pack_matrices` per-partition timings: the sum of
/// partition times (divide by the stage wall time to get the mean number of
/// concurrently busy workers), the median and maximum, and the slowest
/// partitions with their pre-pack matrix size. `timings` holds
/// `(partition, matrix_bytes, elapsed)`.
fn report_pack_timings(timings: &[(usize, u64, std::time::Duration)]) {
if timings.is_empty() {
return;
}
let mut by_time: Vec<&(usize, u64, std::time::Duration)> = timings.iter().collect();
by_time.sort_by_key(|t| std::cmp::Reverse(t.2));
let sum: f64 = timings.iter().map(|t| t.2.as_secs_f64()).sum();
let bytes: u64 = timings.iter().map(|t| t.1).sum();
let median = by_time[by_time.len() / 2].2.as_secs_f64();
tracing::info!(
"pack: {} partition(s), sum of partition times {:.0}s, median {:.1}s, max {:.1}s, {:.1} GB of matrix data",
timings.len(),
sum,
median,
by_time[0].2.as_secs_f64(),
bytes as f64 / 1e9,
);
for (i, b, d) in by_time.iter().take(5).map(|t| (t.0, t.1, t.2)) {
tracing::info!(
"pack: slowest partition {i}: {:.1}s, {:.2} GB of matrix data",
d.as_secs_f64(),
b as f64 / 1e9,
);
}
}
+8
View File
@@ -49,6 +49,14 @@ impl Stage {
wall_secs = format_args!("{:.3}", stats.wall_secs),
rss = %fmt_bytes(stats.max_rss_bytes),
swaps = stats.swaps,
user_secs = format_args!("{:.1}", stats.user_secs),
sys_secs = format_args!("{:.1}", stats.sys_secs),
minor_faults = stats.minor_faults,
major_faults = stats.major_faults,
vol_ctx = stats.vol_ctx,
invol_ctx = stats.invol_ctx,
in_blocks = stats.in_blocks,
out_blocks = stats.out_blocks,
"done"
);
if stats.swaps > 0 {