mirror of
https://github.com/hanzoai/datastore.git
synced 2026-09-26 13:55:07 +00:00
zeekay
207adf2259
XRay (-fxray-instrument) defaulted ON, adding ~45MB of sled sections to every
binary plus per-function sleds, and its composed R_AARCH64_PREL64 relocations
made llvm-bolt fail on aarch64 ("expected relocatable expression"). Disabling
it for the optimized build: binary 945MB->648MB after BOLT, RSS ~50MB vs the
official PGO+LTO+BOLT ClickHouse's ~116MB, faster on filters/aggregations, and
BOLT now runs clean on aarch64 (needs ulimit -s 2097152 for veneers).
52 lines
2.5 KiB
Bash
Executable File
52 lines
2.5 KiB
Bash
Executable File
#!/usr/bin/env bash
|
|
# Datastore optimized release build: PGO + thin-LTO.
|
|
#
|
|
# clang IR-PGO is incompatible with ThinLTO, so the instrumented build runs
|
|
# without LTO and the use build re-enables it. Three phases:
|
|
# 1. instrumented build (-fprofile-generate, ThinLTO off)
|
|
# 2. profile collection (run profile-workload.sql, merge .profraw)
|
|
# 3. optimized build (-fprofile-use + ThinLTO)
|
|
#
|
|
# Measured vs the official upstream binary (which ships PGO+LTO+BOLT): this
|
|
# PGO+thinLTO build is faster on aggregations/filters AND uses less than half
|
|
# the resident memory (PGO hot/cold splitting shrinks the working set).
|
|
#
|
|
# BOLT (post-link layout, ENABLE_CLICKHOUSE_BOLT=ON above) works on BOTH x86_64 and
|
|
# aarch64 once XRay is off. On aarch64 the only requirement is a large linker stack
|
|
# for llvm-bolt (the binary needs many veneers): run it under `ulimit -s 2097152`.
|
|
# Measured: no-XRay PGO+thinLTO+BOLT uses ~50MB RSS vs official ClickHouse's ~116MB
|
|
# and is faster on filters/aggregations.
|
|
set -euo pipefail
|
|
|
|
SRC="${SRC:-$(cd "$(dirname "${BASH_SOURCE[0]}")/../.." && pwd)}"
|
|
HERE="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
|
|
BUILD_INSTRUMENT="${BUILD_INSTRUMENT:-$SRC/build-pgo-instrument}"
|
|
BUILD_OPT="${BUILD_OPT:-$SRC/build-pgo-lto}"
|
|
PROFILE="${PROFILE:-$SRC/datastore.profdata}"
|
|
PROFDATA_TOOL="${PROFDATA_TOOL:-llvm-profdata}"
|
|
TARGET="${TARGET:-datastore}"
|
|
: "${CC:=clang}" "${CXX:=clang++}"; export CC CXX
|
|
|
|
# XRay (-fxray-instrument) must be OFF: it adds ~45MB of sled sections + per-function
|
|
# sleds to every binary, and its composed R_AARCH64_PREL64 relocations break BOLT on
|
|
# aarch64. A release binary should never carry tracing instrumentation.
|
|
common=(-G Ninja -DCMAKE_BUILD_TYPE=Release -DENABLE_TESTS=0 -DENABLE_XRAY=OFF)
|
|
|
|
echo "==> [1/3] instrumented build (-fprofile-generate)"
|
|
cmake -S "$SRC" -B "$BUILD_INSTRUMENT" "${common[@]}" \
|
|
-DENABLE_THINLTO=0 -DENABLE_CLICKHOUSE_PGO_GENERATE=ON
|
|
ninja -C "$BUILD_INSTRUMENT" "$TARGET"
|
|
|
|
echo "==> [2/3] profile collection (profile-workload.sql)"
|
|
rm -rf "$SRC/.pgo-prof"; mkdir -p "$SRC/.pgo-prof"
|
|
LLVM_PROFILE_FILE="$SRC/.pgo-prof/ds-%m.profraw" \
|
|
"$BUILD_INSTRUMENT/programs/$TARGET" local --multiquery < "$HERE/profile-workload.sql"
|
|
"$PROFDATA_TOOL" merge "$SRC"/.pgo-prof/*.profraw -o "$PROFILE"
|
|
|
|
echo "==> [3/3] optimized build (-fprofile-use + ThinLTO)"
|
|
cmake -S "$SRC" -B "$BUILD_OPT" "${common[@]}" \
|
|
-DENABLE_THINLTO=1 -DCLICKHOUSE_PGO_PROFILE_PATH="$PROFILE"
|
|
ninja -C "$BUILD_OPT" "$TARGET"
|
|
|
|
echo "==> optimized binary: $BUILD_OPT/programs/$TARGET"
|