Tor Relay Power Consumption Optimization in 2026

A Tor relay's CPU is consumed by: AES-256-CTR cell encryption/decryption (the dominant cost for high-bandwidth relays), TLS handshakes for relay-to-relay connections, circuit setup (public key operations: X25519 for ntor handshake), and Linux kernel network stack processing. AES is hardware-accelerated on all modern x86-64 CPUs via AES-NI and on ARM CPUs (Cortex-A53 and later) via ARM Cryptography Extensions. Verify AES-NI is enabled: grep -m1 aes /proc/cpuinfo. Without AES-NI, AES encryption is 5-10x slower in software. For high-bandwidth relays (100+ Mbit/s), a CPU without AES-NI will be saturated before the network interface. Choose VPS or bare metal hardware with AES-NI support - all modern Intel and AMD CPUs since 2010 support AES-NI.

Linux OS settings that affect relay power consumption: CPU frequency scaling (cpufreq). For relays with variable load, performance governor wastes power at idle; conservative or ondemand governor scales frequency based on load. Set: echo conservative > /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor. For relays with consistently high load, performance governor is appropriate. C-state configuration: deep C-states (C6, C7) reduce power significantly during idle periods but add latency when transitioning back to active. For low-latency requirements, disable deep C-states; for relays prioritizing energy efficiency at the cost of slightly higher latency: install cpupower and configure with cpupower idle-set -E. Network interrupt coalescing: ethtool -C rx-usecs 200 reduces CPU wakeups from frequent small interrupts, slightly reducing power at the cost of minimal latency increase.

Tor's configuration options that affect CPU utilization: NumCPUs - defaults to auto-detect, but setting this to the actual number of physical cores (not hyperthreads) can reduce context switching overhead. RelayBandwidthRate and RelayBandwidthBurst - limiting bandwidth reduces CPU usage proportionally. For operators who want to reduce power consumption, lower bandwidth limits reduce CPU load at the cost of reduced network contribution. DisableOOMKill 1 - prevents out-of-memory killer from terminating Tor during memory pressure, which would require restart overhead. CircuitBuildTimeout and LearnCircuitBuildTimeout - circuit building involves CPU-intensive public key operations; adjusting these affects how aggressively Tor builds backup circuits. MaxCircuitDirtiness can reduce circuit churn, reducing the rate of new circuit establishment (the most CPU-intensive operation per circuit).

For bare metal relay operators, hardware selection determines baseline energy efficiency. Most energy-efficient options for 2026: ARM servers (Ampere Altra, AWS Graviton-based servers) provide excellent performance-per-watt compared to x86 - a 32-core Ampere Altra at ~60W idle can handle 10 Gbit/s Tor relay traffic. AMD EPYC processors provide better performance-per-watt than comparable Intel Xeon in recent generations. Intel Xeon Scalable (Sapphire Rapids) includes AI accelerators that can offload some cryptographic operations. For VPS deployments, power consumption is indirect - but choosing shared VPS over dedicated hardware means power costs are pooled and not directly attributable. The energy-efficiency consideration for VPS operators is mainly about choosing a provider that uses renewable energy (Iceland has near-100% renewable electricity, making Iceland VPS hosting among the lowest-carbon options).

For bare metal servers, use IPMI power monitoring: ipmitool sdr type Power shows current power draw in watts. IPMI power logging provides historical consumption data. For sustained relay operation, compare power draw at different traffic loads to understand the relationship between bandwidth and power consumption. For VPS relays where direct power monitoring is unavailable, proxy metrics: CPU utilization percentage (correlates with power draw) and total CPU time consumed are available via /proc/stat and tool like htop. For environmental impact reporting, calculate kWh consumed: average watts * hours / 1000. Iceland's electricity is 99%+ renewable (geothermal + hydropower) - a relay in Iceland has near-zero net carbon impact regardless of power consumption.