POST /app on AWS: slopio, hyper, h2o

Graviton4 (Neoverse-V2), eu-west-1, 28 October 2026. Each server reads the whole request body, goes over every byte of it, then sends the 2048 bytes of /. Medians over the repetitions; h2load's closed loop.

How the runs were set up

Queues, cores and threads as each run's env.txt and server logs record them.

CPU and network: nothing was throttled

  • No CPU burst. c8gn instances have fixed performance: no CPU credits to spend or run out of, as burstable (T) instances have, and no throttling once they are spent. Every vCPU runs at full speed for the whole run.
  • One vCPU is one physical core. Graviton4 has no SMT: no server shared its cores with a sibling thread.
  • The servers were limited by taskset, not by AWS, to the CPUs in the table. The machine's other CPUs stayed idle, and the load generator ran on an instance of its own.

Requests per second, one core

The server on CPU 0 of a c8gn.4xlarge, with all 8 of the NIC's queue interrupts on that same CPU. Each panel has its own scale.

Table

Request bodies received, one core

Requests per second × body size, in Gbit/s, at 16 × 16. The instance allows 50 Gbit/s: no run was held back by the NIC.

The server's CPU per request, one core, 16 × 16

µs of user and kernel time of the server process per request. Interrupts on the same CPU are not the process's time and are not counted. Each panel has its own scale.

Table

Four cores, 16 KiB bodies, 64 × 32

  • Server: CPUs 0-3 of a c8gn.8xlarge (32 vCPUs, 100 Gbit/s); its other 28 CPUs idle.
  • NIC: 4 queues, queue n's interrupts on CPU n, irqbalance stopped. Each worker's CPU takes its own queue's packets, and no other worker's.
  • Connections steered: SO_INCOMING_CPU on each worker's listener hands a connection to the worker on the CPU its packets arrive on (on by default in slopio, INCOMING_CPU=1 for hyper and h2o). slopio-hash is slopio with it off (STEER=0), connections spread by SO_REUSEPORT's hash: in the table only.
  • Workers: four per server, one pinned to each CPU, each with its own listener. slopio: one io_uring ring per worker. hyper: one current-thread tokio runtime per worker (reuseport mode, not work-stealing). h2o: one event loop per thread.
  • Load: h2load on another c8gn.8xlarge, 32 threads on its 32 CPUs, 64 connections of 32 streams, so about 16 connections and 512 requests in flight per worker.
Requests per second
The server's CPU per request, µs

Table

20,000 connections, one core, 3 KiB bodies and answers

Each connection sends one request at a time, at so many per second, or as fast as it can. At 5 req/s each no server reached the 100,000 asked: h2load sends every connection's request at the same instant, so there the CPU each server took, in the table, says more than its req/s.

Requests per second
Memory per connection, KiB

Table