Local transport latency on naamah, 2026-08-23

This directory records two complete runs of the RESP and SBE local-transport latency matrix added in commit d122399a3c78be3a4941f7d4767387b7c070bff2.

Each case used one depth-one client, 10,000 warm-up requests, and 100,000 measured requests for each of PING, SET, and GET. SET and GET used the key goblin:local-transport-latency and a 16-byte value. Times are end-to-end client round trips, including protocol encoding and reply decoding.

Two-run mean

Median round-trip latency, in microseconds:

p99 round-trip latency, in microseconds:

The median ordering was ring, Aeron IPC, UDS, then Aeron UDP loopback. Aeron IPC added roughly 0.15--0.23 microseconds over the corresponding ring median. SBE had a lower median than RESP in every transport/operation pair.

Both runs reproduced an Aeron UDP SET p99 around 153--158 microseconds for both protocols. The benchmark establishes that this tail is repeatable on this configuration. Subsequent ordered-trace experiments found that the 16-byte SET value periodically grows and compacts its spilled string tail in the store arena. The Media Driver's default exponential-backoff agents amplified those brief pauses into contiguous UDP latency clusters.

Aeron polling policy

These archived runs used DEDICATED Media Driver threads but did not override Aeron 1.51's default backoff idle strategy. DEDICATED assigned conductor, sender, and receiver agents to separate pinned threads; it did not make those agents poll continuously. The recorded 153--158 microsecond SET p99 therefore describes the default-backoff policy.

The current benchmark launcher instead defaults to Aeron's low-latency C Media Driver profile: conductor spin, sender and receiver noop, and two messages per network send. Goblin's C++ Aeron client conductor also uses spin. In the controlled follow-up on this host, that policy reduced the same 16-byte Aeron UDP SET p99 to 18.4 microseconds for RESP and 19.2 microseconds for SBE; it continuously occupies the pinned C++ client-conductor and Media Driver agent cores. Checkout the recorded source commit above to reproduce these archived backoff results exactly.

Files

• run-1.csv and run-2.csv are the complete per-run statistics emitted by the benchmark.
• aggregate.csv places each run's p50 and p99 beside their arithmetic mean.
• metadata.txt records source, dependency, binary, host, and affinity provenance.

Isolation and topology

Every protocol/transport pair ran against a freshly started server. The server main loop was pinned to CPU 2 and the measuring client thread to CPU 3. Aeron client conductors used CPUs 10 and 11. Driver conductor, sender, and receiver agents used CPUs 4--6 for the server driver and CPUs 7--9 for the client driver. These are distinct physical cores on the single NUMA node; SMT siblings start at CPU 64.

Aeron IPC used one local Media Driver. Aeron UDP used separate server and client Media Drivers connected through 127.0.0.1, so its data path crossed the kernel loopback UDP stack. UDS used a Unix-domain stream socket. The ring and Goblin Aeron subscription paths were busy-polled, while the Media Driver agents used the historical backoff policy described above. UDS used the server's socket event loop.

The machine had no configured isolated CPUs. Its scaling governor reported performance, while frequency boost remained enabled. These are latency microbenchmarks, not throughput results.

Reproduction

The two runs used the committed launcher with only build/result locations and sample counts overridden:

env AERON_PREFIX=/tmp/packrat-aeron-d122399.JoXfbh/aeron-prefix \
    BUILD_DIR=/tmp/packrat-aeron-d122399.JoXfbh/build \
    OUT_DIR=/tmp/packrat-aeron-d122399.JoXfbh/results-full \
    SKIP_BUILD=1 SAMPLES=100000 WARMUP=10000 \
    benchmarks/local_transport_latency.sh

For the second run, OUT_DIR was changed to results-full-repeat.