Hacker News Reader

Comment by KozmoNau7 | Hacker News Reader

I built a DIY NAS in a small cube case (Fractal Node 304), which can house a mini-ITX motherboard and six 3.5" drives, and has pretty good (and quiet!) cooling for a case that size (two 92mm fans in front and a 140mm in back). The drive bays are not hot swappable, but are very easy to replace when the top is taken off the case.

For the OS, I put openSUSE on it, set up a btrfs pool and configured SMB, NFS, FTP and SSH/SCP manually. It also functions as my HTPC, with Kodi as the frontend.

The Celeron J4105 is passively cooled and rarely sees 50C even under full load, plus it has hardware video decoding, so it can play back 4K video all day with no strain.

It's not a turnkey solution like a Synology or QNAP box, but it's a lot more flexible.

View on HN

Interesting. What is roughly the total price, the average power consumption, and what is the main advantage compared to ARM systems? Am I correct to assume the Intel chips still draw a lot more power than the single-digit Watt numbers of ARM systems, which would become a substantial accumulated cost over the years?

The price for the hardware was DKK 2,500 or $400. Synology doesn't have a 3.5" 6-bay NAS, but their cheapest 4-bay is the DS420j at DKK 2300, or around $370. So while that is slightly less expensive and probably uses less power, it also has less room for disks and is significantly less flexible. If you have modest demands and just need something to store your files, it'll work.

Or you could go for the DS1612+ to have room for 6 disks like in my build, which will set you back DKK 6,900 or $1,100. And that uses an AMD Ryzen CPU, not ARM. On the upside, it also has transcoding and media playback functionality like my build, and some features my build doesn't have, like an upgrade to 10gbit ethernet.

I'm not counting disks in these prices, since I used disks I already had, which I would also have done if I had bought a prebuilt NAS.

The TDP for the Celeron J4105 in my build is 10W, representing "the average power, in watts, the processor dissipates when operating at Base Frequency with all cores active under an Intel-defined, high-complexity workload". It very rarely runs at anything approaching a high-complexity load, even when transferring files over gigabit ethernet. It idles at around 30-35C passively cooled aside from the case fans, which run inaudibly at 500rpm.

Obviously the chipset and RAM also draw some power, as do the disks and fans and so on, but it compares very favorably to the quoted 51W under load/25W idle for the DS1612+. That is why I specifically chose a motherboard with that class of CPU.

Let's say the 51W figure is the same for my build and assume a worst-case load, running at full power 24/7. With the current electricity prices here, that's DKK 773 or $125/year.

Compare to the DS420j, which is quoted at 22W, DKK 335 or $54/year, for a device that is significantly less capable.

And my build is not just a NAS. It also serves as my HTPC and couch gaming machine for retro and emulated games, among other things. Plus it's 100% under my control, I get to decide what's installed and which features are enabled/disabled. For me, the choice is obvious.

> It's not a turnkey solution like a Synology or QNAP box, but it's a lot more flexible.

That is not even the biggest advantage. The biggest advantage, in my opinion, is that you will always get updates and can run that for an unlimited time, without breaking your setup.

Building your own is a cool thing, and gives you many benefits. But updates and Linux are very prone to breakages - there’s too much hw and sw combinations to expect solid experience.

That is why I chose very commodity hardware and run openSUSE Leap 15.2 on it. I wanted a solid distro with a reasonable update policy, not something that pushes 100+ new updates every week. Debian Stable or perhaps Testing would also have been good choices, but I like YaST.

In the interest of full disclosure, I have had one issue with my build. The motherboard has four SATA connectors; Two from the Intel chipset, which work flawlessly, and two from an additional ASM1061 chipset, which don't. They seem to work initially, but when you start pushing larger amounts of data through them, errors start to pop up.

As far as I can tell, it's a either a driver issue or a BIOS issue, I can't remember whether I got those same errors on an ASM1061-based 2-port SATA PCIe expansion card or not. Either way, I replaced it with Marvell-based 4-port card instead, and just avoid using the ASM1061 ports for now.

Next time I add a disk to the system, I'll connect it to the potentially troublesome ports and run some tests, to see whether it was a driver issue that has been cleared up.

So yeah, not perfect, but good enough for me :-)

My home "NAS" running opensuse has been continually upgraded since something like OpenSuse 12. I have data drives and I have an OS ssd that's only 50GB. It's rock solid.

> But updates and Linux are very prone to breakages - there’s too much hw and sw combinations to expect solid experience.

That seems like FUD to me. I am running Linux since over 20 years. A NAS's hardware interfaces are mostly HDD and network drivers, and I have not seen them breaking in many years.

And especially for PC-platform hardware, or ARM SoC in which Linux is the dominant OS. I could understand if NAS vendors would say that in their advertisement, but what is the substance of claims like this? Especially since these things runs Linux as well.

There are way more things that can go wrong, not just drivers.

I've run a ThinkServe machine with CentOS and ZFS for several years. Kernel and ZFS updates would break it regularly, requiring tinkering to make it work again (my favorite: zfs#8885).

Nowadays I prefer when somebody else did the integration work.

Well, I see better what point of view this is coming from.

However, ZEF is not part of the standard kernel, IIRC there are still license issues and it needs to be built separately, which is not necessary with any normal file system. I enjoy some amount of tinkering, too, but I believe it is expecting too much that everything works out of the box with non-standard kernel modules.

In the same vein, I also do not understand entirely why people want to have something still experimental like btrfs precisely on their NAS, and apparently use that without an extra load of backups. btrfs has many qualities, but it is still not considered as robust as ext4 (the current standard Linux file system), and for a long time it had practically no recovery option if file system data became corrupt. For me, it is very clearly not the first choice for a NAS or server which should run completely hassle-free. IMO, everything one needs for that is on Debian stable, so why not use that? And why in heaven should one make oneself dependent on what a specific hardware vendor does?

The nice think about RHEL/CentOS is, that their kernels have stable ABIs (kABI). That allows for third parties to provide binary kernel modules, and the OpenZFS project actually provides the kABI-kmod tracking repo, so you can have an installation without any developer tools installed (which I prefer). So the problem wasn't really with kernel modules, the problem was the integration with the system, and nobody really cared - the above mentioned issue is still open, 2 years later.

Another nice thing about CentOS is, that it was supported for 10 years. Debian is supported only for 3. It tkaes 5 years for Synology to get a new release done, so they also do have kinda-long term support.

Btrfs is stable and has been widely used in production for years. Facebook uses it, SUSE CaaS uses it, the officially recommended platform for SAP is SLES, which has defaulted to btrfs for the root filesystem since 2014.

Btrfs brings snapshots, rollback, deduplication, storage pools, checksumming, send/receive and a whole host of other features. It's a completely different type of filesystem to ext4.

Calling btrfs "experimental" at this point is like calling electric cars "experimental".

I whish I could do the same with a CPU that supports ECC ram. Unfortunately, a fanless CPU with ECC support (the Atom x6000 for example) is almost a chimera, or really really expensive.

I built a custom home server using Mini-ITX form factor parts and I would have to agree with this conclusion. ECC ram in that form factor is very expensive. The biggest issue I had was that fanless ATX power supplies are not reliable. There was also a hardware bug that was bricking Atom C2xxx processors.

I think that if form factor is not an issue, one of the i3's that support ECC ram are low enough TDP to be cooled passively. Especially if it is not expected to run at full load for long periods. However I simply switched to using a big case with low RPM, good quality FDB fans. It hasn't been a problem since, and is much less cramped when I do want to tinker with it.

I would absolutely go for a micro-ATX or even full ATX board if that allows you better choices in hardware to get ECC or other features.

My goal was to build something not too much larger than a 4-6 disk NAS, that would also function as a HTPC, sitting under my TV. That meant small size and quiet operation were the main priorities. As none of the popular consumer NAS devices use ECC RAM, I reasoned that I would do without it in my build, too.

The fans are running on the lowest variable setting controlled by the BIOS, it's whisper-quiet even with your ear next to it, but I've got some even quieter fans from be quiet, that I'm planning to swap in. The PSU is also from be quiet!, and with the placement deep in the case, with the 120mm fan oriented downwards, it's completely inaudible.

Had I not been constrained by space and noise concerns in my small apartment, I would have build a larger, more powerful and ECC-equipped machine to be hidden away in an equipment rack in the basement.

Bingo, this is what I recommend. It's much more flexible.

Reply on news.ycombinator.com