intp

Interference profiling. Precise observability. For Linux.

---

This repository is a name reservation. A public marker for the place where the intp utility will land once it is ready to be packaged and — if all goes well — submitted as an official Linux package (PPA on Launchpad → Debian → Ubuntu → other distributions).

The code that will live here is not born in this directory. It is designed, compared, and hardened in intp-comparison, where seven parallel variants compete, side by side, to become the implementation that earns the intp name on apt install.

The vision in one sentence

To measure, precisely and cheaply, how one workload interferes with another when they share the same Linux hardware — CPU, last-level cache, memory bandwidth, block I/O, and network. In production. Without assembling a zoo of tools.

# the end goal
sudo apt install intp
intp -p $(pgrep -f my-application) -i 100ms

Seven metrics. One single utility.

intp is the direct heir of the original interference profiler published by Xavier & De Rose (SBAC-PAD 2022, PUCRS), rewritten to survive modern kernels and contemporary clouds. In short windows it collects seven signals that answer the question "who is stealing what from whom?":

Metric	What it measures	Source
netp	NIC physical utilization (TX + RX in bps)	driver counters
nets	Kernel network stack service time	softirq / NAPI
blk	Block device busy percentage	block layer
mbw	Memory bandwidth (LLC ↔ DRAM)	uncore IMC PMU
llcmr	LLC miss ratio	perf_event
llcocc	LLC occupancy in bytes	Intel RDT / resctrl
cpu	CPU utilization (user + sys)	scheduler / cgroup

Each metric is a window; the set is the fingerprint of interference. The technical deep-dive lives in METRICS-DEEP-DIVE.

Why this exists

Every cloud — public or private — sells the illusion that vCPUs are independent. They are not. Co-located workloads fight over cache, over memory bandwidth, over NIC queues, over I/O queues. The symptom is latency that appears with no obvious cause; the cause is the noisy neighbour nobody can see.

Traditional tools (top, sar, iostat, perf) show what each thing consumes. intp shows how that consumption translates into observable interference — the signal that interference-aware schedulers (such as IADA) need in order to decide.

intp is not a replacement for htop. It is the missing sensor between "this metric is high" and "this is hurting the neighbour".

Relationship with intp-comparison

The sibling repository intp-comparison is the test bench. Seven implementations coexist there, each with a different combination of instrumentation paradigm, cost, portability, and fidelity:

Variant	Paradigm	Kernel module	Min. kernel
v0	Classic SystemTap (original paper)	yes	≤ 6.6
v0.1	SystemTap, patched for 6.8+	yes	6.8+
v1	SystemTap native probes	yes	6.8+
v1.1	SystemTap + userspace helper	yes	6.8+
v2	Pure C: procfs / perf_event / resctrl	no	4.10+
v3	eBPF / CO-RE with libbpf	no	5.8+
v3.1	bpftrace + Python orchestrator	no	5.8+

Only the kernel-module-free variants (v2 and v3) survive the upstream packaging bar — official distributions do not accept out-of-tree modules without a long-term maintenance path. When the experimental comparison is done, the winning variant lands here and becomes intp.

The research behind that decision is the Master's dissertation of André Sacilotto Santos, advised by Prof. Cesar De Rose (PPGCC/PUCRS), with defense expected in March 2027.

Roadmap to the official package

The journey from "GitHub placeholder" to sudo apt install intp has well-defined stops. The full breakdown lives in docs/ROADMAP.md; at a glance:

1. Public name reservation — you are here.
2. Variant selection — intp-comparison converges on a single kernel-module-free implementation.
3. Debian packaging — debian/ directory, dh_make, manpages, optional systemd unit, reproducible install tests.
4. Personal PPA on Launchpad — first public channel for Ubuntu LTS (22.04, 24.04, 26.04).
5. Debian sponsorship — Intent To Package (ITP), mentors.debian.net review, sponsored upload by a Debian Developer.
6. Ubuntu universe — automatic sync from Debian.
7. Adjacent distributions — Fedora (COPR → official), Arch (AUR → official), Alpine, Nix.

Every step has its own technical gate (quality, tests, distro policy) and its own bureaucratic gate. None of them is trivial. That is why the "if all goes well" is not false modesty — it is honesty about a process that involves more than code.

How to follow along

Until the package is born, the place to watch the work happening is intp-comparison — home of the variants, the benchmarks, and the technical documents (METRICS-DEEP-DIVE, KERNEL-6.8-CHANGES, VARIANT-COMPARISON).

Local documentation:

• docs/VISION.md — Why intp, and why now.
• docs/ROADMAP.md — The steps to the official package.

Citation

Until the defense happens, please cite the source repository (intp-comparison) using the metadata in CITATION.cff. The dissertation reference will be added after the defense, expected in March 2027.

Original work:

Xavier, M. G. and De Rose, C. A. F. (2022). IntP: Quantifying Cross-Application Interference via System-Level Instrumentation. SBAC-PAD 2022, Bordeaux, France, pp. 221–230. IEEE. PUCRS. PDF · IEEE

License

MIT.

---

The name is reserved. The code is on its way.