> I bought a Kingston SNV3S/1000G NVMe drive. It supposedly can read and write at 6 GB/s and 5 GB/s respectively
That's a cheapo QLC drive making those numbers very conditional on several factors. Under sustained load it's not a surprise to see such drives drop to SATA speeds.
> So the limiting factor is the enclosure, that should still give me a generous 1000 MB/s.
That's capped by the USB controller, which means 512MB/s max.
I believe the USB controller still shares a bus with the NIC, which means maxing out the network connection will affect disk rw performance. You could confirm this by observing a concurrent iperf affecting fio results.
> The Raspberry Pi is notorious for not being excellent with encryption algorithms.
Well, that's specifically AES precompiles. Just like for kopia, you can benchmark and change LUKS algo for something more suitable for the platform like xchacha20: https://rr-developer.github.io/LUKS-on-Raspberry-Pi/
Author here, thanks for pointing out the mistake! The USB controller speed is indeed 4 Gbps and not 4 GB/s, I have updated the post accordingly.
> I believe the USB controller still shares a bus with the NIC, which means maxing out the network connection will affect disk rw performance. You could confirm this by observing a concurrent iperf affecting fio results.
Thanks for the pointers, I didn't know about iperf! That's something I'll likely try at some point for the sake of completeness, although I have other projects competing for my attention.
> Well, that's specifically AES precompiles. Just like for kopia, you can benchmark and change LUKS algo for something more suitable for the platform like xchacha20: https://rr-developer.github.io/LUKS-on-Raspberry-Pi/
Author here, that's a great question. I only host a handful of services on the VPS, including Nextcloud. I was mostly interested in storage, not so much in compute power.
Netcup offered VPS with 1 TB of storage for €30 / month back when I subscribed. The server has always been oversized for my needs, but that was the most financially interesting back then.
I recently switched to another ISP at home that gives me a fixed IP. I decided to move my services at home to shave off the VPS costs. Since I already had a Raspberry Pi 4, I only had to buy a SSD and an enclosure.
When you add up the component costs for RPi with NVMe, suitable case, cooling and power supply, it is not much cheaper than a mini PC which generally outperforms the RPi.
I use several RPis of various models. But there are times when even an old Intel/AMD laptop is a more suitable solution.
I’m not sure what electricity costs where you live, but my calculations tell me I’d have to run an intel n4000 for 5+ years before I break even compared to buying a CanaKit rpi 5.
Eh? I'm using an old MacBook as a home server; I kind of naively assumed that there was a whole load of protection circuitry and firmware controls in there that would turn it off before it expands or the magic smoke comes out.
This article has nothing to do with the speed of a Raspberry Pi as a server, unless the purpose of this server is to run "kopia".
I ran an 8 gig Raspberry Pi 4 with mirrored 8 TB spinning rust disks for quite a while. Performance is very good, once you stop playing with all the meta tasks and run, you know, server stuff. I ultimately switched to another system not because of CPU, but because of memory. SearXNG [1] likes a lot of memory.
That's a cheapo QLC drive making those numbers very conditional on several factors. Under sustained load it's not a surprise to see such drives drop to SATA speeds.
> So the limiting factor is the enclosure, that should still give me a generous 1000 MB/s.
That's capped by the USB controller, which means 512MB/s max.
I believe the USB controller still shares a bus with the NIC, which means maxing out the network connection will affect disk rw performance. You could confirm this by observing a concurrent iperf affecting fio results.
> The Raspberry Pi is notorious for not being excellent with encryption algorithms.
Well, that's specifically AES precompiles. Just like for kopia, you can benchmark and change LUKS algo for something more suitable for the platform like xchacha20: https://rr-developer.github.io/LUKS-on-Raspberry-Pi/
reply