Benchmarks

How fast do phyelds programs run on each host, and what happens as the agents grow? The harness in benchmarks/ measures it. This page shows its baseline: one world of 100 to 10,000 agents, on one core of a laptop.

The unit is the agent-round per second: the rounds of all the agents in a second of wall time. A thousand agents that each run ten rounds a second make 10,000 agent-rounds per second.

---

How the hosts scale

Every agent runs distance_to. The area grows with the agents, so each agent keeps about eight neighbors, and a flat line would mean that an agent costs the same however many there are.

• ARGoS and Mesa keep their pace. From 100 to 10,000 agents, ARGoS loses a factor of 1.5 and Mesa a factor of 3.
• The simulator loses a factor of 50. Its radius_neighborhood compares every node with every other one, so a round costs O(N²). At 10,000 agents a neighbor search takes 36 s, and a round is mostly that search.

VMAS and Webots are measured with 100 agents only (see what is missing).

---

What each layer costs

Each host runs three times: on its own, with the phyelds adapter running an empty program, and with distance_to. The further left a dot moves, the more that layer costs.

• The program costs about the same on every host. distance_to adds about 20 to 55 µs to each round of each agent, and count_nodes, which also collects, 60 to 165 µs.
• The adapter is what tells the hosts apart. A Mesa agent alone takes a fraction of a microsecond, and the adapter adds 9 µs per agent at 100 agents, 80 µs at 10,000. In the simulator, the adapter's cost is the neighbor search, which grows with the agents: 50 µs per agent at 100 agents, 3.6 ms at 10,000.
• VMAS is bound by its own step. On the CPU, its flocking scenario takes 2.4 ms per agent and step, so phyelds does not show.
• ARGoS and Webots have no host-alone dot. Their robots need a controller to run, and Webots alone runs its rounds faster than it starts.

---

All the measurements

Agent-rounds per second of each stage, M for millions. Empty runs the adapter with a program that does nothing; Gradient, Gossip and Count run distance_to, gossip_max and count_nodes (benchmarks/programs.py). At constant density, each agent has about eight neighbors. At constant area, the agents share the area of 1,000, so the density grows with them.

Host	Agents	Layout	Host alone	Empty	Gradient	Gossip	Count
Simulator	100	density	1.4M	18,861	13,982	16,788	8,995
Simulator	100	area	2.0M	25,914	16,418	17,356	10,209
Simulator	1,000	density	1.7M	2,991	2,722	2,784	2,384
Simulator	1,000	area	1.7M	2,979	2,726	2,782	2,387
Simulator	10,000	density	1.3M	274	280	282	279
Simulator	10,000	area	1.4M	285	276	274	267
Mesa	100	density	14.1M	108,815	24,000	28,383	11,626
Mesa	100	area	12.5M	121,326	27,769	32,265	13,901
Mesa	1,000	density	134.3M	65,026	20,420	22,689	10,288
Mesa	1,000	area	126.0M	66,624	19,774	23,198	10,270
Mesa	10,000	density	964.3M	12,598	8,317	9,021	5,845
Mesa	10,000	area	954.9M	11,306	6,517	7,096	3,628
VMAS	100	density	411	406	397	398	389
VMAS	100	area	413	403	397	396	391
ARGoS	100	density	—	41,657	13,693	16,704	6,998
ARGoS	100	area	—	63,600	21,708	24,677	10,303
ARGoS	1,000	density	—	28,019	11,512	13,428	6,113
ARGoS	1,000	area	—	29,027	11,567	13,534	6,287
ARGoS	10,000	density	—	15,445	9,039	7,650	4,365
Webots	100	density	—	20,959	10,931	11,433	6,837
Webots	100	area	—	23,403	12,355	12,388	9,009

The neighbor search of all the agents, timed on its own, the messages delivered in a round, and the peak memory: of the measurement, or of its largest process for ARGoS and Webots.

Host	Agents	Layout	Neighbor search (s)	Messages per round	Peak memory (MB)
Simulator	100	density	0.00448	636	30
Simulator	100	area	0.00317	0	30
Simulator	1,000	density	0.321	7,132	66
Simulator	1,000	area	0.321	7,132	66
Simulator	10,000	density	36.7	72,768	114
Simulator	10,000	area	35.4	749,840	114
Mesa	100	density	0.000594	636	120
Mesa	100	area	0.000536	0	121
Mesa	1,000	density	0.0102	7,132	129
Mesa	1,000	area	0.0103	7,132	129
Mesa	10,000	density	0.741	72,768	210
Mesa	10,000	area	0.783	749,840	210
VMAS	100	density	0.00328	636	230
VMAS	100	area	0.00333	0	230
ARGoS	100	density	—	—	51
ARGoS	100	area	—	—	50
ARGoS	1,000	density	—	—	113
ARGoS	1,000	area	—	—	114
ARGoS	10,000	density	—	—	752
Webots	100	density	—	—	300
Webots	100	area	—	—	300

• Machine: AMD Ryzen AI 9 HX PRO 370 with 30 GB, Linux 7.0, Python 3.13.2, phyelds 9.3.1. Mesa 3.5.1, VMAS 1.5.2 with PyTorch 2.5.1 on the CPU, ARGoS 3 and Webots R2025a.
• Cores: each measurement ran on one Zen 5 core (taskset -c 2). Webots, which runs a process per robot, ran on the eight Zen 5 CPUs (taskset -c 0-3,12-15).
• Rounds: 10 for the simulator, Mesa and VMAS (2 for the simulator at 10,000), 100 for ARGoS (60 at 10,000) and 500 for Webots. ARGoS and Webots run in processes of their own, so their time leaves the start-up out.
• Seed: 1, so every host gets its agents in the same places.

What is missing

• 100 agents at constant area: the agents are spread over the area of 1,000, out of each other's range, so they exchange no messages. At 1,000 the two layouts are the same.
• VMAS beyond 100 agents: on its own, its flocking scenario takes tens of seconds a step with 1,000 agents.
• ARGoS at 10,000, constant area: it takes over half an hour to start 10,000 foot-bots packed in the area of 1,000. At constant density it takes about 5 minutes.
• Webots beyond 100 robots: each robot is a Python process, which caps Webots at a few hundred robots on one machine.

These are measurement points, not promises: they change with the machine, the programs and the release.

---

Measuring again

poetry run python benchmarks/run.py --hosts mesa --nodes 100 1000 > results.jsonl
poetry run python docs/examples/benchmarks.py   # redraw the charts

Pin the measurements to one core, e.g. with taskset -c 2 on Linux: on a processor with cores of different speeds, a measurement that is not pinned can land on either.

The benchmarks README lists the hosts, stages and metrics, and how to measure each host. The measurements of this page are in benchmarks/baseline.jsonl, one JSON line each, with their environment.