GET on AWS: slopio, hyper, h2o

Graviton4 (Neoverse-V2), eu-west-1, 26 to 29 October 2026. / answers 2048 bytes from static memory; /app?size_kb=3 3 KiB. Medians over the repetitions. Closed loop (h2load: a request leaves as soon as an answer lands) unless said open loop (bench/loadgen: requests leave at a set rate whether or not answers have come).

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 on /, by cores

As fast as h2load goes. slopio against hyper on 1, 4 and 8 cores; h2o was not yet in the bench for the 8-core run. Each panel has its own scale.

  • 1 core: CPU 0 of a c8gn.4xlarge, the NIC's 8 queues on that same CPU; h2load on another c8gn.4xlarge, 4 threads.
  • 4 cores: CPUs 0-3 of a c8gn.8xlarge, NIC queue n on CPU n, connections steered; h2load on another c8gn.8xlarge, 32 threads.
  • 8 cores: CPUs 0-7 of a c8gn.8xlarge (100 Gbit/s), the NIC's interrupts left to irqbalance; an older build (26 October), h2load as for 4 cores.
  • hyper in reuseport mode, one current-thread runtime per core, as slopio runs; its work-stealing multi mode is in the table.

Table

Latency under an open loop, against h2o

Requests sent at a set rate, Poisson arrivals, each timed from when it was due: the latency a client sees, which a closed loop hides. Four workers, NIC queues aligned with them.

p99 latency, ms, 600k req/s on 4 workers, by connections
p99 latency, ms, 300k req/s on 4 workers, by connections

Table

50,000 connections on 8 cores, by where the NIC's interrupts go

Each connection sends its next request as soon as it has the answer. CPUs 0-7 of a c8gn.8xlarge, 8 NIC queues; h2load on another, 32 threads. slopio against hyper (reuseport), 26-27 October build; h2o was not measured here.

Table

20,000 connections, one core, GET /app?size_kb=3

Each connection sends one request at a time, at so many per second, or as fast as it can. CPU 0 of a c8gn.4xlarge, the NIC's 8 queues on that CPU; h2load on another, 16 threads.

Requests per second
Memory per connection, KiB

Table