NuttX First Impressions: A POSIX RTOS With Real Trade-offs

NuttX is the only embedded RTOS I know that takes POSIX seriously. That fidelity has a cost in footprint and build complexity — we evaluated NuttX 12.3 on i.MX RT1064 for a gateway SKU that needed open(), poll(), and real sockets without shipping Linux.

Evaluation setup

Hardware: MIMXRT1064-EVK, 1 MB flash, 512 KB OCRAM, Ethernet + SD. Workload: Modbus TCP gateway, local SQLite cache (yes, on MCU — bounded), HTTPS POST batch upload.

Compared against: Zephyr 3.4 (same silicon), FreeRTOS + lwIP (incumbent on smaller nodes).

Success criteria: POSIX code from desktop prototype runs with <20% rewrite; OTA via MCUboot slot layout; cold boot <3 s.

First impressions (good)

POSIX isn't marketing

int fd = open("/mnt/sd/log.bin", O_RDONLY);
struct pollfd pfd = { .fd = fd, .events = POLLIN };
poll(&pfd, 1, 1000);

This ran. Sockets API matched our Linux tools. Onboarding time for backend engineers dropped — they didn't learn k_mutex first.

NuttShell on serial

nsh for field debug beats re-flashing semihosting builds. dmesg, ifconfig, mount SD — ops team friendly.

Deterministic build with nuttx config system

menuconfig familiar to kernel people. Device drivers under drivers/ follow Linux naming enough to grep productively.

Trade-offs (real)

Footprint

Minimal config still ~350 KB flash vs. ~180 KB Zephyr same features vs. ~120 KB FreeRTOS+lwIP bare.

OCRAM pressure from SQLite page cache — we capped at 32 KB and accepted slower writes.

Build complexity

Out-of-tree app linking requires understanding Makefile + Application.mk + board defconfig. First clean build: 45 minutes CI. Zephyr west: 12 minutes same machine.

Error messages when Kconfig wrong: cryptic until you learn to search NuttX mailing list archives.

Scheduler vs. Zephyr/FreeRTOS

NuttX priority scheduling with POSIX threads — closer to Linux than FreeRTOS. Migration notes from Zephyr vs FreeRTOS apply doubly: read NuttX priority scheme (higher number = higher priority in some APIs — check docs per version).

We lost an afternoon to inverted priority assumption copied from FreeRTOS comments.

Network stack maturity

NuttX bundled network stack worked for Modbus TCP + TLS via mbedTLS port. Throughput 42 Mb/s on EVK — sufficient. UDP packet loss under load test worse than Linux baseline — tuned CONFIG_NET_NACTIVESOCKETS.

Not worse than lwIP on same board; different failure shapes.

OTA angle

Integrated MCUboot swap with NuttX boardctl reset hook. Dual slot on external QSPI flash — see OTA bootloader design for policy; NuttX-specific wrinkle: unmount filesystems before swap or first boot corrupts SD cache journal.

Our sequence:

1. Download to staging partition (HTTPS worker thread).
2. boardctl(BOARDIOC_RESET, ...) after MCUboot mark pending.
3. Boot new image; self-test; confirm or revert counter.

When we'd pick NuttX again

• Gateway class MCU with existing POSIX codebase
• Team with Linux background, no appetite for Zephyr Kconfig religion
• Product needs multi-process isolation (fork not on MCU — but separate tasks with MPU regions workable on RT1064)

When we wouldn't

• nRF52-class constrained nodes — Zephyr or bare metal
• Hard real-time kHz control loops — verify worst-case latency; NuttX OK on RT1064 at our rates, not proven for 100 kHz servo
• Tiny team without POSIX maintenance depth — build system tax hurts

Ship/no-ship decision

Gateway SKU shipped NuttX 12.3. Sensor leaf nodes stayed Zephyr. Hybrid fleet ops cost accepted — two update pipelines, one ops runbook.

Six-month field: 2 bricked units (SD journal + abrupt OTA — fixed in 12.4 bump); network uptime 99.7% vs. 99.6% prior FreeRTOS gateway (noise).

What I'd do next

Prototype same gateway on Zephyr native POSIX layer (CONFIG_POSIX_API) — compare rewrite percentage honestly. NuttX won our eval in 2024; Zephyr POSIX improved since.

Document NuttX defconfig diff from EVK default in repo — we spent 3 days re-deriving Ethernet PHY tweaks.

If you're POSIX-curious: budget 2 weeks for build system fluency before judging scheduler. The RTOS is fine; the Makefile is the onboarding boss fight.

Footprint breakdown (EVK config)

Rough size output from our gateway build:

Component	flash KB
NuttX kernel + scheduler	95
Network stack	78
mbedTLS (minimal ciphers)	62
SQLite (amalgamation trimmed)	48
App + Modbus	67

Zephyr equivalent measured 312 KB total with same features toggled — not apples-to-apples on crypto profile, but directionally NuttX cost is real.

POSIX surface we actually used

Not all of POSIX — fork absent, mmap limited, signals subset. We used: open/read/write, poll, socket, pthread, mqueue. Anything outside that list — verify in NuttX docs before porting Linux library wholesale.

Developer ergonomics vs Zephyr

Task	NuttX	Zephyr
Add I2C sensor	drivers/ pattern	devicetree + driver model
Shell debug	nsh built-in	CONFIG_SHELL
Unit test on host	limited	native_sim strong

Hybrid fleet ops: two bug databases, two CI pipelines — cost accepted for gateway POSIX win.

What I'd do next