Reference Summaries: Measurement, Benchmarks & Jitter Detection

← Back to the reference list · Each entry below summarizes one reference from the collection, so you can decide whether it is worth your reading time.

STAC Benchmarks (STAC-T0, STAC-N1, STAC-TS)

Source: docs.stacresearch.com · current record: AMD240422 · intermediate · ongoing (STAC acquired by ISG, 2020)

The industry’s audited benchmark authority for trading technology: STAC-T0 measures tick-to-trade network-stack round trip (current record 13.9 ns, Exegy/AMD 2024), STAC-N1 measures host network stacks under market-data workloads, and STAC-TS covers timestamping/clock-sync compliance tooling for MiFID II. Reports disclose full hardware/software configurations, which makes them valuable tuning references even if you never run the benchmarks. Free registration required; the old stacresearch.com/SFC170206 link this site carried now lives at docs.stacresearch.com/SFC170206.

cns-sysjitter — Measure OS-induced Jitter

Source: GitHub (Xilinx-CNS) · advanced · tool ~2010s, repo current under AMD/Xilinx

The open-source home of Solarflare’s classic sysjitter tool (v1.4, by David Riddoch): it spins a thread on each isolated core and records every interval where the thread was knocked off-CPU, producing per-core interruption counts and histograms. This is the ground-truth check after applying isolcpus/nohz_full tuning — did the ticks, IRQs and SMIs actually go away? Successor URL to the dead openonload.org download page this site previously linked.

hiccups — Erez Strauss’s OS Hiccup Meter

Source: GitHub · advanced · dormant

A small tool measuring OS “hiccups” — scheduling interruptions experienced by a busy-spinning thread — in the same spirit as sysjitter. Still online but unmaintained; kept because it’s simple enough to read in one sitting and adapt. For maintained alternatives use cns-sysjitter above, or cyclictest/rtla from the kernel’s rt-tests ecosystem.

cyclictest & rt-tests

Source: man page · rt-tests at kernel.org · intermediate · actively maintained

The standard real-time latency measurement suite: cyclictest measures timer-wakeup latency distributions (the canonical RT kernel benchmark), hwlatdetect isolates firmware/SMI-induced latency, and rteval wraps them for whole-system evaluation under load. On modern kernels (5.17+), the in-kernel rtla tool with its timerlat/osnoise tracers is the recommended companion — it tells you not just that jitter happened but which kernel path caused it.

sockperf — Socket-level Latency Benchmarking

Source: GitHub (Mellanox) · beginner · v3.10, 2022

The go-to network latency microbenchmark from Mellanox/NVIDIA: ping-pong, under-load and throughput modes over TCP/UDP with full percentile histograms (not just averages — in latency work the tail is the story). Works with kernel sockets and kernel-bypass stacks (VMA/XLIO, Onload), which makes it the standard tool for before/after comparisons of NIC and stack tuning. Pair with iperf3/netperf for throughput and MoonGen for hardware-rate packet generation.

Corvil (Pico) — Production Latency Analytics

Source: pico.net · beginner · current (2025 platform)

The dominant commercial platform for passive, packet-level latency measurement in trading: nanosecond timestamping, 200 Gbps sustained capture with real-time analytics (Corvil 12000), 600+ trading-protocol decoders, and MiFID II/RegNMS/CAT compliance reporting with traceable time. Shows what production-grade latency monitoring looks like beyond DIY tools — no agents, no application changes, every order measured from market-data tick to order response. Corvil was acquired by Pico in 2019.

c2clat & core-to-core-latency

Source: rigtorp/c2clat · nviennot/core-to-core-latency · intermediate · 2020/2022, still the reference

Two open-source tools that measure cache-coherence latency between every pair of CPU cores and render the result as a matrix/heatmap. They expose the topology effects that decide thread-placement strategy: ~8–25 ns within a CCX or between SMT siblings, 85–180 ns across chiplet boundaries, several-fold penalties across sockets. Running one of these is now a standard characterization step for any new trading server — vendors use them in QA too. (Also summarized in the hardware section.)

Google Benchmark

Source: GitHub · beginner · v1.9.5, January 2026

The standard C++ microbenchmarking harness — statistically robust timing, automatic iteration counts, DoNotOptimize barriers and perf-counter integration. Listed here as the code-level complement to the system-level tools above; see the software section for the full summary.