chore: adjusting resources for e2e test + adding strict_scheduling

[mrzeszutko]

Summary

• Increase CPU/memory allocation for E2E tests that spin up multiple nodes, prover nodes, or validators
• Enable strict (CPU-aware) scheduling for E2E test execution to prevent oversubscription

Problem

All E2E tests currently run with the default CPUS=2, MEM=8g, regardless of how many processes they start. A standard test (1 Anvil + 1 Aztec Node + 1 PXE = 3 processes) fits comfortably in 2 CPUs. But P2P tests spin up 6-10+ full validator nodes, epoch tests run multiple validators plus prover nodes, and fee tests always start a prover node — all squeezed into the same 2 CPUs.

Meanwhile, GNU parallel runs up to 64 jobs concurrently (num_cpus / 2) without awareness of per-test CPU needs. It happily runs 64 tests each claiming 2 CPUs even when some actually need 6-8. This causes CPU starvation, missed sequencer timing windows, and flaky timeout failures.

Changes

1. Per-test resource overrides (get_test_resources)

Added a function that maps test file paths to appropriate CPUS/MEM based on the number of processes each test actually runs. The values were chosen by analyzing the source code of each test to count processes at peak load:

Category	Tests	CPUS	MEM	Processes at peak	Examples
Standard (1 node)	~70	2 (default)	8g	AN + AZ + PXE = 3	e2e_token_contract/*, e2e_deploy_contract/*
Single node + prover	~15	3	12g	AN + AZ + PN + PXE = 4	e2e_fees/*, e2e_simple, e2e_epochs/epochs_multiple
Multi-validator (3-4 nodes)	~4	4	16g	AN + 3-4 AZ + PXE = 5-6	e2e_epochs/epochs_simple_block_building, epochs_multi_proof
P2P medium (4 validators, no prover)	~10	4	16g	AN + BS + 4 AZ = 6	e2e_p2p/duplicate_proposal_slash, rediscovery
Multi-validator + prover	~12	6	24g	AN + 4-6 AZ + PN = 7-9	e2e_p2p/gossip_network, epochs_mbps*, reqresp/*
Extremely heavy	2	8	32g	10-13 processes	e2e_p2p/preferred_gossip_network, add_rollup
Prover full fake	1	3	12g	AN + AZ + PN + PXE = 4	e2e_prover/full (non-CI_FULL)
Prover full real	1	16	96g	(unchanged)	e2e_prover/full (CI_FULL)

Process abbreviations: AN = Anvil, AZ = Aztec Node (sequencer + archiver + world-state + P2P + validator), PN = Prover Node, PXE = Private eXecution Environment, BS = Bootstrap Node.

Memory formula: MEM = CPUS * 4g (same ratio as the existing default of CPUS=2/MEM=8g), which provides headroom for each process's heap, native code, and world-state trees.

2. Strict scheduling for standalone E2E test runs

Changed the test (and test_and_collect_avm_inputs) function in yarn-project/end-to-end/bootstrap.sh to use STRICT_SCHEDULING=1. This only affects standalone execution of E2E tests — i.e., when running ./bootstrap.sh test directly (e.g., grind runs, local dev, or any workflow that invokes the test function). It does not affect normal CI runs, where the Makefile calls test_cmds and feeds commands to the test engine, which uses plain parallelize without strict scheduling.

The strict scheduler:

• Tracks available CPU cores with a semaphore
• Only starts a test when enough cores are free to satisfy its CPUS requirement
• Pins each test to specific cores via CPU_LIST / taskset

This prevents oversubscription in standalone runs: with 128 cores and tests requesting 2-8 CPUs each, the scheduler naturally limits concurrency to ~30-40 tests instead of the previous 64, with each test getting the cores it actually needs.

This is the same scheduler already used by benchmarks (bench function).