Performance

Performance comparison between snmpkit (Rust core) and pyagentx3 (pure Python).

Benchmarks

Analysis

PDU Encode: 11.5x faster

The most critical operation for agent response times. snmpkit encodes PDUs in Rust with zero-copy byte handling, while pyagentx3 uses Python’s struct.pack with multiple allocations.

Value Create: 6.4x faster

Creating SNMP values (Integer, OctetString, Counter64, etc.) is significantly faster in Rust. This benefits agents with large table responses where thousands of values are created per request.

OID Parse: 1.5x faster

Parsing OID strings like “1.3.6.1.4.1.12345” into internal representation. snmpkit uses a Rust parser with pre-allocated buffers.

Header Decode: ~1x (equal)

Header decode is roughly equal between implementations. This 20-byte operation is so fast (~1M ops/sec) that the PyO3 FFI overhead of crossing the Python↔Rust boundary negates the Rust speedup. This is expected and acceptable since:

1. Header decode happens once per PDU (not a bottleneck)
2. The operations that matter (encode, value creation) are much faster
3. Real-world performance is dominated by network I/O

Test Environment

All benchmarks measure raw encoding/decoding performance without network I/O overhead. Each operation runs 10,000 iterations.

Real-World Impact

Response Latency

For a typical SNMP GET request with 10 OIDs:

snmpkit responds ~9.5x faster before network latency.

Throughput

At 1000 OIDs per response:

• pyagentx3: ~500 responses/sec (bottlenecked by encoding)
• snmpkit: ~5,000 responses/sec (bottlenecked by network)

Memory

snmpkit uses less memory per OID:

• pyagentx3: OIDs stored as Python strings with full object overhead
• snmpkit: OIDs stored as compact Rust tuples, exposed to Python on-demand

Optimization Tips

Enable Parallel Encoding

For agents serving 1000+ OIDs:

agent = Agent(parallel_encoding=True)

This uses Rust’s rayon for multi-threaded PDU encoding.

Tune Update Frequency

Don’t update faster than needed:

# Fast updates only where necessary
agent.register("1.3.6.1.4.1.12345.1", RealtimeMetrics(), freq=1)
 
# Slow updates for stable data
agent.register("1.3.6.1.4.1.12345.2", ConfigData(), freq=60)

Use Appropriate Types

Counter64 is larger to encode than Integer32. Use the smallest type that fits your data:

# Good: use Counter32 for values < 4 billion
self.set_COUNTER32("1.0", packet_count)
 
# Only use Counter64 when needed
self.set_COUNTER64("2.0", bytes_transferred)

Reproducing Benchmarks

Run the benchmarks yourself:

cd benchmarks
python run_benchmarks.py

Results are saved to benchmarks/results.json.