Implementation Guidelines

$ Production

# Tuned

* Engine

Same rules as baseline - standard framework configuration with no TCP tuning.

Same as baseline tuned rules. The MTU change is applied externally by the benchmark runner.

No specific rules.

Stress test with extreme TCP fragmentation - loopback MTU forced to 69 bytes, splitting every packet into ~29-byte payload segments. Uses 8 diverse request templates designed to maximize parser stress.

Connections: 512, 4,096, 16,384 Requests per connection: 2 (rotates through all templates)

How it works

Before the load generator starts, the benchmark script sets the loopback interface MTU to 69:

sudo ip link set lo mtu 69

This is the lowest practical MTU (IP minimum is 68). With 20-byte IP headers and 20-byte TCP headers, each TCP segment carries only 29 bytes of payload. The MTU is restored to 65,536 after the test.

Each connection sends 2 requests then reconnects, rotating to the next template. This ensures all 8 templates are exercised evenly across all connections.

Request templates

8 templates designed to stress different aspects of HTTP parsing under fragmentation:

Template	Size	Segments	Purpose
GET (×2)	355B	13	Browser-like headers (User-Agent, Accept, etc.)
POST	360B	13	Content-Length body across segments
Chunked	379B	14	`\r\n\r\n` header boundary splits at segment edge
Noise	843B	30	500-character junk path (expects 404)
Cookie bomb	2,097B	73	2KB cookie header parsed across 73 segments
30 Headers	380B	14	30 tiny headers, boundaries everywhere
Body29	146B	6	Body fills exactly one segment (29 bytes)

What it measures

TCP segment reassembly efficiency under extreme fragmentation
HTTP parser robustness when headers, bodies, and boundaries split across reads
Syscall overhead when the kernel processes many small packets
Cookie and multi-header parsing across fragmented buffers
Chunked transfer-encoding with split \r\n\r\n boundaries
Framework resilience to junk routes under fragmentation

Expected behavior

Frameworks typically see 60-90% throughput reduction compared to baseline (MTU 65,536). Some frameworks collapse at high connection counts (4,096+) as the kernel TCP stack gets overwhelmed by the volume of tiny segments. The stacked bar in the results shows which template types each framework handles well vs which cause errors.

Parameters

Parameter	Value
Endpoint	`/baseline11` (GET + POST)
Loopback MTU	69 bytes (29 bytes payload per segment)
Templates	8 (GET×2, POST, Chunked, Noise, Cookie, 30 Headers, Body29)
Connections	512, 4,096, 16,384
Requests/connection	2 (rotate on reconnect)
Pipeline	1
Duration	5s
Runs	3 (best taken)
Load generator	gcannon

Validation