FLoDeck

FLoDeck is a discrete-event queueing simulation framework for modeling the flow of computational load through constrained computing systems — it models the flow of work through queues and compute resources. In the context of HPC, it simulates job arrivals, queueing, scheduling and execution to generate realistic job traces and analyze resource utilization under configurable workload and scheduling policies.

The name is inspired by Flow and Load.

Features

• Queueing-theory core — arrival processes are workload flows (Poisson, tiered-random or replayed from trace files), the service process is a pool of identical nodes, and the queueing discipline is FIFO or Priority with aging.
• Admission control — backlog limits (total and/or per workload flow); over-limit jobs are either dropped and counted, or kept in a holding buffer and promoted back when a slot frees up.
• Backfill scheduling — an optional scheduler plans job starts over per-node timelines (walltime based), letting small jobs start in gaps ahead of wide blocked ones.
• Pluggable site policies — a SitePolicy maps node counts to priority tiers (aging boosts, walltime caps); the bundled FRONTIER policy follows the OLCF Frontier user guide ("job priority by node count" bins and the batch-partition limit of four eligible jobs per user).
• Traces and statistics — per-job records and system-state traces written to files, plus time-averaged job count, mean flow time (delay) and node utilization.

Package layout

Installation

Install flodeck from PyPI:

pip install flodeck

Development Install

To install in editable development mode (for contributing or local development):

# Clone the repository
git clone https://github.com/swing-lab/flodeck.git
cd flodeck

# Install in development mode with test and formatting extras
pip install -e ".[dev]"

Note: flodeck uses setuptools-scm to derive its version from git tags; on a clone without a release tag, the fallback version from pyproject.toml is used.

With the dev extras installed, run pytest for the test suite (tests/) and ruff check . for linting.

Quick start

from flodeck import FloDeck, poisson_flow

simulator = FloDeck(num_nodes=100)
simulator.run(flows=[poisson_flow(arrival_rate=22. / 72,
                                  execution_rate=1. / 3,
                                  time_limit=1000.)])
simulator.report()

or from the command line:

flodeck --nodes 100 --arrival-rate 0.3 --execution-rate 0.33 \
        --time-limit 1000
flodeck --policy frontier --backfill --hold-overflow \
        --arrival-rate 0.0017 --time-limit 86400

Simulator options

FloDeck(num_nodes, ...) accepts:

• backlog_limit — total limit of the waiting backlog
• backlog_rules — BacklogRules (ordering, per-flow limits, admission hook)
• hold_overflow — buffer rejected jobs instead of dropping them
• backfill — enable the backfill scheduler
• time_limit — timestamp when the processing must stop (otherwise the run lasts while the flows produce jobs)
• output_path — per-job records: arrived_at, started_at, finished_at, span, flow, label
• trace_path — system-state trace: at, event, running, queued, held (also printed with run(..., verbose=True))

Workload flows

A workload flow is any generator that yields Job objects in arrival order; predefined ones cover the common cases.

Rates are parameters of exponential distributions, expressed in events per unit of simulated time: on average, jobs arrive every 1 / arrival_rate and run for 1 / execution_rate time units. Time itself is unit-free — the unit chosen for the rates is also the unit of time_limit, walltimes and the reported statistics (the bundled FRONTIER policy uses seconds).

from flodeck import file_flow, poisson_flow

flows = [poisson_flow(arrival_rate=11. / 36, execution_rate=1. / 3,
                      span=100, flow='main', time_limit=1000.),
         file_flow(path='flodeck_input.txt', flow='external',
                   time_limit=1000.)]

Either num_jobs or time_limit must be set for generated flows.

Modeling a specific machine

from flodeck import FRONTIER, FloDeck, tiered_flow

simulator = FloDeck(num_nodes=FRONTIER.node_count,
                    backlog_rules=FRONTIER.backlog_rules(),
                    hold_overflow=True,
                    backfill=True,
                    time_limit=86400.)
simulator.run(flows=[tiered_flow(arrival_rate=1. / 600,
                                 tiers=FRONTIER.tiers.values(),
                                 time_limit=86400.)])
simulator.report()

After a run: simulator.completed holds the finished jobs, simulator.trace the state snapshots, and mean_job_count(), mean_flow_time(), utilization() and dropped_count provide the summary statistics.

Examples

See examples/:

• general.py — two merged flows on an M/M/c-like system
• theory.py — analytic Erlang-C reference values for validation
• streamed.py — a custom job generator with a holding buffer
• frontier.py — Frontier policy with priority aging and backfill

Acknowledgments

FLoDeck is an independent implementation and design effort inspired by the ATLAS-Titan/allocation-modeling project. The original software was developed by M. Titov, with the conceptualization and methodological framework designed in collaboration with A. Poyda and S. Jha.

License

This project is licensed under the Apache License, Version 2.0.