Skip to content

ARM box as home server replacement: 6 month report

VPS Hosting by HieuTce 24 replies 2.3K views
#11

HieuTce said:
Measured with a cheapo P3 Kill-A-Watt clone from Lazada, so give or take 0.2W maybe.
Not the OP but I tested this on Hetzner's ARM64 cloud in Falkenstein. Their virtio-net does checksum offloading but no TSO/GSO on the smallest plans. Had to disable it for WireGuard or throughput halved. OP's on Contabo so different hypervisor probably.

HieuTce said:
Measured with a cheapo P3 Kill-A-Watt clone from Lazada, so give or take 0.2W maybe.
Xhci_hcd timeout, port reset, device reconnect. Classic. Seen it on three different Altra boards, always the same port if you push USB2 bandwidth hard.

#12

green15 said:
Their ARM lineup is VPS ARM 4, VPS ARM 6, and VPS ARM 8
Yeah it's the 4. And you're right about the USB resets, that's exactly my "one weird USB snag." It only hits when I'm pushing the RTL-SDR for ADS-B, which is USB2 bulk all day. Moved the SDR to a Pi 4 on the same switch and feed it over TCP now, problem solved. The ARM box just does the decoding.

#13

3.2W idle is making me cry into my 650W PSU. My "home server" is a 5950X with a 3090 that I tell myself I need for "AI inference" but mostly I use it to host a Minecraft server for three people. My denki bill last month was ¥34,000. I am the problem.

8-char NTLM found in 4min 32sec
6 #14

Hana5 said:
My denki bill last month was ¥34,000
Oof. I thought my £140 quarterly was bad running a Dell T30 in the cupboard. ARM's tempting but I need the PCIe slots for HBAs, can't fit that in these little boxes.

Honey badger don't care... about downtime
#15

HieuTce said:
Moved the SDR to a Pi 4 on the same switch and feed it over TCP
This is the way. I run rtl_tcp on a Pi Zero 2 W, draws 0.8W, feeds dump1090 on the ARM box. Zero USB drama. The Pi's entire job is "be a USB adapter with better drivers."

1ms or I don't want it
#16

LichunLars said:
Rtl_tcp on a Pi Zero 2 W
Salam kenal Lars! I do same but with Airspy Mini, Pi 3B+ because Zero cannot keep up with 10MHz sample. 1.4W for the Pi, still nothing. My total stack: Pi for SDR, ARM box for processing, old NUC sold to friend. Very clean.

traffic worse than my packet loss
6 #17

HieuTce said:
Wall. Measured with a cheapo P3 Kill-A-Watt clone
Those are surprisingly accurate actually, within 1-2% if it's the BRENENSTUHL clone. The BMC power reading is almost always DC rail only, misses fan and NVMe 3.3V rail. Your wall number is the real one.

I've been logging mine to InfluxDB for a year. Contabo ARM 6 in Singapore, idles 4.1W, peaks 11W during the nightly ZFS scrub.

#18

sslman762 said:
Contabo ARM 6 in Singapore
How's the peering from Singapore to Vultr's Singapore region? I got burned by Contabo's routing before, their German IPs sometimes trombone to Frankfurt before heading back to London.

#19

srelet said:
How's the peering from Singapore to Vultr's Singapore
Same building basically. 0.8ms. Contabo SIN and Vultr SIN are both at Digital Realty SIN10, different floors. The routing is fine when both ends are there, it's the cross-region that gets weird. You can check the peering yourself on https://bgp.he.net.

#20

I'm on Contabo's London ARM 4 for €5.99 and I can't complain. Runs WireGuard, Pi-hole, and a small Matrix server for friends. 3.8W at wall. The only annoyance is the €1.20 IPv4 surcharge, but that's Contabo being Contabo.

Post a reply

You need an account to reply. Log in or register to join the conversation.

Post reply Preview Save draft