RackNerd NVMe shared: 90 days, not transformative

Web Hosting by LichunLars 1mo ago 3 replies 178 views

LichunLars

Member OP

Joined:: Jun 2024
Posts:: 190
From:: Stockholm, SE

1mo ago · edited

Migrated from Hetzner SSD (2019) to RackNerd NVMe shared (2026-04). Expected I/O revolution. Got incremental improvement at best.

fio --name=randread --ioengine=libaio --iodepth=32 --rw=randread --bs=4k --direct=1 --size=4G --numjobs=4

Hetzner SSD:  IOPS=4,200  latency=7.6ms
RackNerd NVMe:  IOPS=5,800  latency=5.4ms

Sequential reads better — 550MB/s vs 180MB/s — but who runs databases sequentially?

ping -c 100 RackNerd-node-07.driftnet.local
avg: 12.3ms  max: 89.4ms  stddev: 14.2ms

Route via their internal mesh adds 8-15ms vs advertised. Contention spikes during backup windows.

Not bad. Not transformative. Marketing oversold. Numbers attached, judge yourselves.

1ms or I don't want it

marcus_qc

Member

Joined:: Jun 2024
Posts:: 301
From:: Montreal, CA

1mo ago · edited

We ran NVMe shared at InterServer for like 6 months and killed it. Super stoked on the hardware, not stoked on the reality. Controller saturation hit at like 40% capacity. Customers mad. We moved everyone to dedicated slices and ate the cost.

Our bad. Fixed fast tho.

42U and still growing

srelet

Member

Joined:: Jul 2024
Posts:: 135
From:: San Francisco, USA

1mo ago

LichunLars said:

Contention spikes during backup windows

This is your error budget burning in real-time. No transparency from RackNerd on noisy neighbor policies?

We publish our contention ratios. We publish our backup windows. Blameless postmortem culture means we also publish when we get it wrong. Transparency isn't a feature, it's the baseline.

sofialund

Member

homelab heatstroke

Joined:: Jul 2024
Posts:: 140
From:: Buenos Aires, Argentina

1mo ago

¡the numbers they are honest, thanks you for this! In my work the engineer of storage he always says «the NVMe fast» but the controller the shared is bottleneck red. ¡the IOPS they do not scale linear no!

Post a reply