SolidRun 1U 2 node Arm Server
servethehome.com
servethehome.com
No estimated price and no performance figures still.
Scaleway had their nicely priced C1 instances. Well, they still have and they run pretty solidly in my (limited) experience. But they have not updated the kernel for ages. It seems pretty clear that they will shut them down sooner or later without any successor.
Is there any other comparable offering out there?
(I intentionally mix owning and renting a server. Aren't we in the age of the cloud...)
C1 (ARM32 dedicated) was still available recently when I checked last time. Existing customers' instances are definitely still running, used one today.
I just wish they would let me run Alpine Linux on the DEV1-S. Its available for the more expensive DEV1 instances, but they obviously don't want me taking full advantage of the meager resources the DEV1-S offers.
Of course if you do it under time is money principle that's a trainwreck compared to using their image. But if you do it for the fun of it...
What does Brexit mean for EU customers? Do we slip paying VAT now? 24% in my case.
Well you can have an Intel Atom N2800 / 4GB DDR3 / 2Tb HDD dedicated server at OVH for 8 EUR / month.
And, like all their entry level servers, it's 100 Mbps max, which really hurts when you've got fiber at home.
But in some cases they can be convenient.
They have a lot of different offers at various prices (Kimsufi / So you start / OVH: all the same company basically).
Those VPS, while limited, actually come with what I would consider a fairly reasonable monthly quota (32T) for a low cost plan.
On the non-quantitive side I really prefer to run ARM over ugly Intel.
In case you didn't know, a Scaleway C1 is
* 2.99 per month, billed hourly
* 1.00 per month for the network (unlimited) and you can drop that if you have other machines in the same datacenter and do not need direct internet access
* ~ 1.00 per month for VAT (I guess that's 0,00 if you are outside of the EU)
Of course they are not for high performance computing, but there are enough use cases where the performance is prefectly good enough.
I use them for ssh jumpboxes and find their latency better than AWS EC2, which is much more expensive.
They were far too ahead of its time that didn't make any sense ( It isn't a pun against their Tag Line ). Once AWS Graviton 2 has the whole ecosystem ready ScaleWay could then enjoy the benefits. To kick start ARM on server space ScaleWay is biting off more than they can chew. The same answer to people ask why doesn't DO do Edge Worker / Container.
About twice as much as a C1, more if you have a lot of traffic, like e.g. tunneling video streams. Of course they are not fully comparable products. AWS maintenance might be better. But I don't have any high availability requirements for my toy projects.
Free trial until June (I guess you still pay for the traffic and maybe the storage. Haven't read the fineprint.)
It's mini-ITX, so there's a pile of cases you can use for it if you want.
In a well engineered 1U server setup you want the motherboard to be as low to the bottom of the chassis as possible, then the CPU in its socket, and a big passive heatsink (either aluminum or skived copper) occupying almost all the rest of the vertical room inside the case. The server should be like a front to back wind tunnel where the 40mm fans are moving air through the CPU heatsink(s) without the need for separate fans on top of the CPUs themselves.
The fan unit on top of that is really problematic from a above-the-fan vertical clearance and airflow perspective.
Additionally the SODIMM slots are blocking a typical airflow path from the front edge of the motherboard towards the rear. More normally a server motherboard for 1U would not use laptop size RAM, but would use normal size DIMMs oriented in such a way that they're parallel to the path of airflow.
(disclosure: I used to work for a server manufacturer and was responsible for procuring components from Taiwanese vendors, and designing new generations of 1U single and dual socket boxes to custom specs)
I'd love to have one of these to replace ODroid N2+ just for a rack mount solution, but not at that price.
I intend to cover that in depth when I get my hands on it.
Anything that's COM Express is produced in very small quantities and is very expensive. I'll be shocked if it isn't 3 to 4x the cost of a single socket 8-core Ryzen that runs circles around it in performance.
the SolidRun Article on their Systemsready work:
https://www.solid-run.com/news/how-honeycomb-lx2k-and-system...
Its a form-factor specifically designed for server rooms: where you put ear-plugs in before entering the room. The money spent on the racks / physical infrastructure is small compared to the ongoing costs of actually powering or cooling the room
Not just that but 1U servers also use a fair percentage of their wattage moving air with 40mm fans. The hotter the TDP of all the electronics in your server is, the more air you need to move. 40x28mm high RPM 12VDC fans are quite inefficient in terms of a ratio of cubic meters of air moved per hour vs their watt-hours consumed.
In a 2U chassis if you can find a way to pack in multiple motherboards and use 60mm height fans the possible efficiencies are much greater.
If you look at the 'wall of fans' in the center of something like a Dell 1U dual socket system, there's eight or ten fans each of which if running at full speed can be an 8W to 10W load. This is necessary because you might have two 120W TDP CPUs under passive heatsinks that need a LOT of air pushed past them.
Additionally a stack of 30 or 40 1U servers each with its own discrete 110-240VAC input, DC output power supply is quite inefficient in large numbers. A fair bit of wattage is wasted to spinning the 40mm fan in the power supply and in the densely packed AC to DC conversion circuitry. This is one reason why things like the FB Open Compute platform servers are sometimes 1.5RU high (so they can use 60mm fans), and use a single large AC-to-DC power supply that can take in 277VAC (or even 480VAC!), and output 12VDC to 48VDC to each motherboard in the same rack cabinet.
Even 1-node in 2U is definitely getting into 220V dedicated outlets and dedicated buildings. 1U or 1/2U per node certainly exists, but you need to start spending a good amount of thought on power and cooling to actually support that kind of infrastructure.
-----------
> If you look at the 'wall of fans' in something like a Dell 1U dual socket system, there's eight or ten fans each of which if running at full speed can be an 8W to 10W load. This is necessary because you might have two 120W TDP CPUs under passive heatsinks that need a LOT of air pushed past them.
That's possibly an advantage here. ARM's are known for lower power consumption, so maybe the fans can run slower (and therefore draw less power). I'd have to think about these specs more though...
A 2-node x 1U ARM server might work out if these ARM servers have very low power requirements in a very strong "horizontal scale-out" kind of setup.
But its still hard for me to think exactly what the use case would be. An I/O based server would probably be better in 4U: for 50 to 100 Hard Drives on a beefier 4U Xeon or 4U EPYC chassis.
So its one of those "what's the use" products. Under typical circumstances, a 2U or 4U beefy server split up into multiple VMs probably is a superior architecture.
But having more "tiny" physical machines that are mostly independent (maybe sharing a PSU, but otherwise a fully independent node) for "bare metal hosting" has some benefits over VMs. I mean... do you want a single dual-socket 128-core EPYC in 4U or do you want 8x16 ARMs in 1/2U each.
I dunno, I think the 128-core dual-socket 4U EPYC is gonna be better for cooling, power, and VM-flexibility. At least, that's my instinct. Unless you know that you absolutely want individual non-VM nodes.
------
EDIT: The I/O options discussed here are exceptional. If low CPU-power but high I/O is needed (more SPF+ ports or whatever), then that's probably the ARM's advantage over a single dual-socket EPYC system.
EDIT2: 16x A72 seems weak. But 10Gbps SFP+ x4? That's... actually really really good. Might be hell to actually write software that takes advantage of all that bandwidth though.
Ordinarily in a x86-64 intel or AMD chipset server if you wanted 4 x 10Gbps SFP+, in addition to whatever NICs are on the motherboard, you'll end up using one of the Intel chipset, low-profile PCI-Express interface cards in a slot. Using some 1U long depth Dells as an example you might get a motherboard that has 2 x 1000BaseT and 2 x 10GbE SFP+ in daughtercard plugged into the motherboard, and then you'd have two or three low-profile PCIE slots to add more NICs. Or some combination like one full-height size PCIE slot and one low profile slot.
That said, you are absolutely correct that rackmount equipment is expensive by it's form-factor - but given this is a mini-itx board clearly not designed for use in a rackmounted chassis (given clearance for CPU cooling and orientation of the memory modules) it's not a factor here.