mimalloc 2.0.9: 5-Way Allocator Test and 200-Relay Deployment
We tested 5 memory allocators with Debian 13.2 on 100 Tor relays to find the best option for long-running Guard relays. The result: mimalloc 2.0.9 used 7.4× less memory than mimalloc 3.0.1. We then migrated all 200 relays on the server to the winner.
Experiment Design
| Group | Allocator | Relays | Test Period | Duration |
|---|---|---|---|---|
| A | mimalloc 3.0.1 | 20 | Dec 31 – Jan 10 | 10 days |
| B | mimalloc 2.0.9 | 20 | Jan 2 – Jan 10 | 8 days |
| C | mimalloc 2.1.7 | 20 | Jan 2 – Jan 10 | 8 days |
| D | jemalloc 5.3.0 | 20 | Jan 4 – Jan 10 | 6 days |
| E | glibc 2.41 | 20 | Jan 2 – Jan 10 | 8 days |
All 5 groups ran simultaneously from Jan 4–10 (6 days). Each group had 20 relays with identical hardware and Tor configuration.
Results (Jan 10, 2026)
| Allocator | Avg Memory | vs mimalloc 3.0.1 |
|---|---|---|
| mimalloc 2.0.9 | 1.41 GB | 7.4× less |
| mimalloc 2.1.7 | 1.83 GB | 5.7× less |
| jemalloc 5.3.0 | 2.74 GB | 3.8× less |
| glibc 2.41 | 4.15 GB | 2.5× less |
| mimalloc 3.0.1 | 10.44 GB | — |
mimalloc 3.0.1 Regression
mimalloc 3.0.1 showed unbounded memory growth over 10 days:
| Day | Memory |
|---|---|
| 1 | ~1.9 GB |
| 5 | ~5.3 GB |
| 10 | 10.44 GB |
Memory continued growing at ~1 GB/day with no sign of stabilizing. In contrast, mimalloc 2.0.9 stabilized at ~1.4 GB by day 3 and remained flat.
Observation
We don't know why mimalloc 3.0.1 behaves this way. The data shows memory growth; we haven't investigated the root cause in mimalloc's code.
Production Migration
Based on these results, we migrated all 200 relays to mimalloc 2.0.9:
# Build mimalloc 2.0.9 from source (Debian 13 ships 3.0.1)
wget https://github.com/microsoft/mimalloc/archive/refs/tags/v2.0.9.tar.gz
tar xzf v2.0.9.tar.gz && cd mimalloc-2.0.9
mkdir build && cd build && cmake .. && make
sudo mkdir -p /usr/local/lib/mimalloc
sudo cp libmimalloc.so.2.0 /usr/local/lib/mimalloc/libmimalloc-2.0.9.so
# Configure systemd override for each relay
for relay in $(ls /var/lib/tor-instances/); do
sudo mkdir -p /etc/systemd/system/tor@${relay}.service.d/
cat <<EOF | sudo tee /etc/systemd/system/tor@${relay}.service.d/allocator.conf
[Service]
Environment="LD_PRELOAD=/usr/local/lib/mimalloc/libmimalloc-2.0.9.so"
EOF
done
sudo systemctl daemon-reloadlibmimalloc2.0 package ships mimalloc 3.0.1—the version with the regression. Build 2.0.9 from source instead.
Summary
| Recommendation | Allocator | Memory |
|---|---|---|
| Best | mimalloc 2.0.9 | 1.41 GB |
| Good | mimalloc 2.1.7 | 1.83 GB |
| Acceptable | jemalloc 5.3.0 | 2.74 GB |
| Avoid | mimalloc 3.0.1 | 10.44 GB |