Performance & Tracing Update

High level summary

• Benchmarking: Node 11.1.0 release benchmarks.
• Development: beacon benchmark: new metric capturing disk access time for UTxO lookup; replaced Cairo rendering with gnuplot to reduce bloat and increase portability.
• Tracing: Alert manager feature for cardano-tracer now in implementation phase.
• Leios: On-disk LedgerDB tx validation times benchmark delivered; distributable, self-contained benchmark package in the works.
• Organizational: Decentralization of delivery: The entire Performance & Tracing team successfully moved from IOG to ICAN Group.

Low level overview

Benchmarking

The P&T team has performed and analyzed benchmarks, as well as pre-release bencharks, for the Node 11.1.0 release. While we did measure an increased resident set size, we also saw a clear reduction in process CPU usage. The memory increase has already prompted an investigation and changes, which will make their way into a 11.1.1 patch release.

Development

beacon's comparison and variance reports now break totalTime outliers down into ledger-tick and table-read cost, and commit to a single totalTime metric instead of quietly double-counting table-read time into it — table-read time being what actually captures UTxO lookup cost against the on-disk backing store (beacon PR#16, second of three parts split off from #13, still under review).

We've also dropped beacon's Cairo-based plotting backend entirely in favour of a small gnuplot wrapper (beacon PR#21). The old Chart/Chart-cairo dependency pulled in a full X11 client stack for a tool that only ever renders to PNG files; the switch shrinks the binary's shared-library footprint from 25 down to 5, and let us finally add macOS to CI.

Tracing

Following on from the alarm-system design work reported last time, we've moved into implementation (cardano-node PR#6664). The dispatch payload sent to alarm consumers now carries the source, rule ID, severity and a summary, instead of just the name of the consumer being notified; on the timeseries side, alarm rules are now backed by a small, pure, edge-triggered state machine — cycling through Inactive/Pending/Active states with configurable for/repeatEvery timing and a deterministic series key — while the evaluator thread and value decoder driving it are left for a follow-up PR.

Native support for nested spans in Hermod remains ongoing work, with no PR created yet.

Leios

The on-disk LedgerDB tx validation benchmark work mentioned in our last two posts has been delivered: beacon run can now force the LSM backend to genuinely exercise disk I/O via a new --lsm-no-cache flag, bypassing the OS page cache through O_DIRECT, with --mem-limit/--heap-limit cgroup caps as a fallback for older db-analyser builds; every run also now reports peak RSS, block I/O counts, and a true wall-clock totalTime alongside the previous mutator-only timing (beacon PR#12). Concrete measurements from this on-disk backend are being published under Leios issue#553, the original ask behind this benchmarking effort.

In parallel, we're packaging this benchmark into a self-contained, distributable executable that SPOs can run on their own hardware without nix or network access (beacon PR#22, based on #16). glue bundles glibc and its own curl/unzip, patches its own ELF interpreter at extraction time, and drives measurement, hardware profiling and report assembly through auditable shell scripts rather than Haskell. It's currently under review.

Organizational

As of this month, the entire Performance & Tracing team has moved from IOG to ICAN Group. The team has been warmly welcomed, and the move comes with a genuine sense of a fresh start — new structure, same people, same focus. Delivery on the Cardano Maintenance & Support proposal has continued without disruption throughout the transition; everything reported above was produced across the move.