From 960bc1587f56d12583f07f172592aff6ca0a059e Mon Sep 17 00:00:00 2001
From: Jonathan Hao <jonathan@aztec-labs.com>
Date: Fri, 13 Feb 2026 13:55:29 +0000
Subject: [PATCH] fix: heap-allocate thread-local random buffer to reduce
 thread creation overhead
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

The 1 MiB `thread_local` random buffer in `engine.cpp` was stored inline
in the TLS segment (.tbss), causing every `pthread_create` to allocate and
zero-initialize 1 MiB per thread. With 31 worker threads this added ~18 ms
of overhead to thread pool creation — making UltraHonk verification slower
at 32 cores than at 8 cores.

Moving the buffer to heap allocation on first use reduces the TLS footprint
from ~1 MiB to 16 bytes per thread, dropping thread pool creation from
18 ms to ~1.6 ms for 31 threads.

Benchmark results (parity_base circuit, noir-rollup verifier target):

  Cores | Before | After
  8     | 63 ms  | 59 ms
  32    | 72 ms  | 51 ms
  64    | 87 ms  | 53 ms

Resolves AztecProtocol/barretenberg#1624
---
 .../cpp/src/barretenberg/numeric/random/engine.cpp  | 13 +++++++++----
 1 file changed, 9 insertions(+), 4 deletions(-)

diff --git a/barretenberg/cpp/src/barretenberg/numeric/random/engine.cpp b/barretenberg/cpp/src/barretenberg/numeric/random/engine.cpp
index 6d30a12a75e6..a1eb6a6d259b 100644
--- a/barretenberg/cpp/src/barretenberg/numeric/random/engine.cpp
+++ b/barretenberg/cpp/src/barretenberg/numeric/random/engine.cpp
@@ -9,6 +9,7 @@
 #include <array>
 #include <cstring>
 #include <functional>
+#include <memory>
 #include <random>
 #if defined(__APPLE__)
 #include <TargetConditionals.h>
@@ -44,8 +45,9 @@ constexpr size_t RANDOM_BUFFER_SIZE = 1UL << 20;
 
 #endif
 struct RandomBufferWrapper {
-    // Buffer with randomness sampled from a CSPRNG
-    uint8_t buffer[RANDOM_BUFFER_SIZE];
+    // Buffer with randomness sampled from a CSPRNG (heap-allocated on first use to avoid
+    // bloating TLS — a 1 MiB inline array adds ~0.6 ms per thread creation)
+    std::unique_ptr<uint8_t[]> buffer;
     // Offset into the unused part of the buffer
     ssize_t offset = -1;
 };
@@ -67,8 +69,11 @@ template <size_t size_in_unsigned_ints> std::array<unsigned int, size_in_unsigne
     // We could preserve the leftover bytes, but it's a bit messy
     if (random_buffer_wrapper.offset == -1 ||
         (static_cast<size_t>(random_buffer_wrapper.offset) + random_data_buffer_size) > RANDOM_BUFFER_SIZE) {
+        if (!random_buffer_wrapper.buffer) {
+            random_buffer_wrapper.buffer = std::make_unique<uint8_t[]>(RANDOM_BUFFER_SIZE);
+        }
         size_t bytes_left = RANDOM_BUFFER_SIZE;
-        uint8_t* current_offset = random_buffer_wrapper.buffer;
+        uint8_t* current_offset = random_buffer_wrapper.buffer.get();
         // Sample until we fill the buffer
         while (bytes_left != 0) {
 #if defined(__wasm__) || defined(__APPLE__)
@@ -97,7 +102,7 @@ template <size_t size_in_unsigned_ints> std::array<unsigned int, size_in_unsigne
         random_buffer_wrapper.offset = 0;
     }
 
-    memcpy(&random_data, random_buffer_wrapper.buffer + random_buffer_wrapper.offset, random_data_buffer_size);
+    memcpy(&random_data, random_buffer_wrapper.buffer.get() + random_buffer_wrapper.offset, random_data_buffer_size);
     random_buffer_wrapper.offset += static_cast<ssize_t>(random_data_buffer_size);
     return random_data;
 }