done

ethereum/reth/ethereum-mainnet-reth-minimal-trace--nimbus.yml

@@ -9,6 +9,17 @@ x-logging-defaults: &logging-defaults
 # this is why we have a minimal profile that doesn't actually exist in reth.
 # This node is built from source with architecture-specific optimizations
 # Build command: docker compose build --build-arg ARCH_TARGET=${ARCH_TARGET:-native} ethereum-mainnet-reth-minimal
+#
+# IMPORTANT: Cache optimization considerations
+
+# If running multiple nodes on the same machine, be aware that:
+# - L3 cache is shared across all cores, causing cache contention
+# - Multiple nodes compete for cache space, reducing optimization effectiveness
+
+# - Consider CPU pinning to minimize cache conflicts:
+#   docker run --cpuset-cpus="0-7" ethereum-mainnet-reth-minimal  # Pin to specific cores
+# - For AMD X3D CPUs, CCD0 (cores 0-7) has the 3D V-Cache
+# - For multi-node setups, generic builds may perform better than cache-optimized ones
 
 # Usage:
 #

op/reth/base-mainnet-op-reth-minimal-trace--benchmark.yml

@@ -30,7 +30,17 @@ x-logging-defaults: &logging-defaults
 
 services:
   base-mainnet-op-reth-minimal:
-    image: ${BASE_RETH_IMAGE:-ghcr.io/paradigmxyz/op-reth}:${BASE_MAINNET_RETH_VERSION:-v1.4.3}
+    build:
+      context: ./
+      dockerfile: reth.Dockerfile
+      args:
+        LLVM_IMAGE: ${LLVM_IMAGE:-snowstep/llvm}
+        LLVM_VERSION: ${LLVM_VERSION:-20250514100911}
+        RETH_VERSION: ${BASE_MAINNET_RETH_VERSION:-v1.4.3}
+        RETH_REPO: ${BASE_MAINNET_RETH_REPO:-https://github.com/paradigmxyz/reth}
+        ARCH_TARGET: ${ARCH_TARGET:-native}
+        PROFILE: ${RETH_BUILD_PROFILE:-maxperf}
+        BUILD_OP_RETH: true
     sysctls:
       # TCP Performance
       net.ipv4.tcp_slow_start_after_idle: 0       # Disable slow start after idle

op/reth/base-mainnet-op-reth-pruned-trace--benchmark.yml

@@ -30,7 +30,17 @@ x-logging-defaults: &logging-defaults
 
 services:
   base-mainnet-op-reth-pruned:
-    image: ${BASE_RETH_IMAGE:-ghcr.io/paradigmxyz/op-reth}:${BASE_MAINNET_RETH_VERSION:-v1.4.3}
+    build:
+      context: ./
+      dockerfile: reth.Dockerfile
+      args:
+        LLVM_IMAGE: ${LLVM_IMAGE:-snowstep/llvm}
+        LLVM_VERSION: ${LLVM_VERSION:-20250514100911}
+        RETH_VERSION: ${BASE_MAINNET_RETH_VERSION:-v1.4.3}
+        RETH_REPO: ${BASE_MAINNET_RETH_REPO:-https://github.com/paradigmxyz/reth}
+        ARCH_TARGET: ${ARCH_TARGET:-native}
+        PROFILE: ${RETH_BUILD_PROFILE:-maxperf}
+        BUILD_OP_RETH: true
     sysctls:
       # TCP Performance
       net.ipv4.tcp_slow_start_after_idle: 0       # Disable slow start after idle

reth.Dockerfile

@@ -83,9 +83,9 @@ ENV SCCACHE_DIR=/tmp/sccache
 ENV RUSTC_WRAPPER=/usr/local/bin/sccache
 
 # Set C/C++ flags for dependencies
-ENV CFLAGS_BASE="-O3 -flto=thin -fomit-frame-pointer -fno-semantic-interposition -funroll-loops -ffast-math"
+ENV CFLAGS_BASE="-O3 -flto -fomit-frame-pointer -fno-semantic-interposition -funroll-loops -ffast-math"
-ENV CXXFLAGS_BASE="-O3 -flto=thin -fomit-frame-pointer -fno-semantic-interposition -funroll-loops -ffast-math"
+ENV CXXFLAGS_BASE="-O3 -flto -fomit-frame-pointer -fno-semantic-interposition -funroll-loops -ffast-math"
-ENV LDFLAGS="-Wl,-O3 -Wl,--as-needed -Wl,--gc-sections -flto=thin"
+ENV LDFLAGS="-Wl,-O3 -Wl,--as-needed -Wl,--gc-sections -flto"
 
 # Configure architecture-specific flags and build
 RUN --mount=type=cache,target=/usr/local/cargo/registry \
@@ -159,6 +159,26 @@ RUN --mount=type=cache,target=/usr/local/cargo/registry \
         RUSTFLAGS="-C target-cpu=native -C link-arg=-fuse-ld=/usr/local/bin/mold -C opt-level=3" \
         CFLAGS="$CFLAGS_BASE -march=native" \
         CXXFLAGS="$CXXFLAGS_BASE -march=native"; \
+    elif [ "$ARCH_TARGET" = "multinode-zen4" ] || [ "$ARCH_TARGET" = "multinode-7950x" ]; then \
+        # Optimized for multiple nodes on same machine - reduced cache assumptions
+        RUSTFLAGS="-C target-cpu=znver4 -C link-arg=-fuse-ld=/usr/local/bin/mold -C opt-level=3 -C llvm-args=-enable-machine-outliner" \
+        CFLAGS="$CFLAGS_BASE -march=znver4 -mtune=znver4 --param l1-cache-line-size=64 --param l1-cache-size=32 --param l2-cache-size=512" \
+        CXXFLAGS="$CXXFLAGS_BASE -march=znver4 -mtune=znver4 --param l1-cache-line-size=64 --param l1-cache-size=32 --param l2-cache-size=512"; \
+    elif [ "$ARCH_TARGET" = "multinode-zen5" ] || [ "$ARCH_TARGET" = "multinode-9950x" ]; then \
+        # Optimized for multiple nodes on same machine - reduced cache assumptions
+        RUSTFLAGS="-C target-cpu=znver5 -C link-arg=-fuse-ld=/usr/local/bin/mold -C opt-level=3 -C llvm-args=-enable-machine-outliner" \
+        CFLAGS="$CFLAGS_BASE -march=znver5 -mtune=znver5 --param l1-cache-line-size=64 --param l1-cache-size=48 --param l2-cache-size=512" \
+        CXXFLAGS="$CXXFLAGS_BASE -march=znver5 -mtune=znver5 --param l1-cache-line-size=64 --param l1-cache-size=48 --param l2-cache-size=512"; \
+    elif [ "$ARCH_TARGET" = "multinode-epyc" ]; then \
+        # Optimized for multiple nodes on EPYC systems - assume shared resources
+        RUSTFLAGS="-C target-cpu=znver4 -C link-arg=-fuse-ld=/usr/local/bin/mold -C opt-level=3" \
+        CFLAGS="$CFLAGS_BASE -march=znver4 -mtune=znver4 --param l1-cache-line-size=64 --param l1-cache-size=32 --param l2-cache-size=256" \
+        CXXFLAGS="$CXXFLAGS_BASE -march=znver4 -mtune=znver4 --param l1-cache-line-size=64 --param l1-cache-size=32 --param l2-cache-size=256"; \
+    elif [ "$ARCH_TARGET" = "multinode-generic" ]; then \
+        # Generic multinode optimization - conservative cache assumptions
+        RUSTFLAGS="-C target-cpu=x86-64-v3 -C link-arg=-fuse-ld=/usr/local/bin/mold -C opt-level=3" \
+        CFLAGS="$CFLAGS_BASE -march=x86-64-v3 --param l1-cache-line-size=64 --param l1-cache-size=32 --param l2-cache-size=256" \
+        CXXFLAGS="$CXXFLAGS_BASE -march=x86-64-v3 --param l1-cache-line-size=64 --param l1-cache-size=32 --param l2-cache-size=256"; \
     else \
         RUSTFLAGS="-C target-cpu=$ARCH_TARGET -C link-arg=-fuse-ld=/usr/local/bin/mold -C opt-level=3" \
         CFLAGS="$CFLAGS_BASE -march=$ARCH_TARGET" \