Micro-server

Share here your ideas for new projects
fvolk
Posts: 571
Joined: Sun Jun 05, 2016 11:04 pm
languages_spoken: english
ODROIDs: C2, C4, H2
Has thanked: 0
Been thanked: 61 times
Contact:

Re: Micro-server

Post by fvolk »

powerful owl wrote: I misunderstood your numbers, ... However I realize now that each "job" is the same amount of work.
Sorry for the confusion :-/
Running on the DM a single job, it runs in 7m42sm, so one core at 4.5GHz max boost all the time.
powerful owl wrote: It seems you could pack more compute power into a tighter space, but there's no magic, you still have to get rid of the heat...
I think the DeskMinis are the reasonable trade-off in price/size/performance/cost when you need the performance.
I bought the H470 with 64GiB memory and paid almost 800 EUR (meanwhile memory got more expensive again :-/...)
Today you get double the cores and double the memory for the same money as about 3 years ago.
You get about one hyperthreaded i7 core with 8 GiB memory for about 100 EUR.
Or you go with less powerful and pay less, and also older DeskMinis already are available on the used market.

At the low end with a C4 you get per one core 1 GiB.
A larger cluster of C4s does not make sense to me, pure distributed compute nodes need to be leaner and cheaper for that, as exlibris also observed.
I think my mix of leftover C2s and C4s in a cluster is a good patchwork for learning and experimenting, but scale up to more nodes... no?

Between these high-end and low-end options is the H2.
Odroid chose very wisely the market segment: small size AND silent AND 2 nics AND low idle power.
Use as NAS or build a small cluster, when you need x86 for compatibility reasons or performance.

But having benchmarked various things over the last months, I'm seriously considering to tear down and sell my H2 cluster now.
Real compute jobs run faster on DeskMinis, and time is also money while maintaining the "middle ground" in performance with H2s.
I learned a lot from my H2 cluster, I'm grateful, but maintaining an extra rootfs and non-AVX builds is work,
and now that the H2+ is the standard the H2 lost some resale value -- sell them now or keep them for fun?

Alternatively, a cluster of N2+ derived compute modules as suggested by others?
I don't know where to position that in the market?

Just thinking aloud... ;-)
These users thanked the author fvolk for the post:
odroid (Fri Feb 19, 2021 9:19 am)

powerful owl
Posts: 201
Joined: Thu Mar 28, 2019 8:57 pm
languages_spoken: english
ODROIDs: 2 x HC1, 3 x H2, 1 x H2+
Has thanked: 58 times
Been thanked: 21 times
Contact:

Re: Micro-server

Post by powerful owl »

My H2s are happy running Proxmox with various services and also web servers for some development work. I'm also running MooseFS on them. Possibly a couple of ideas for repurposing yours?

With regard to the distributed computing, I think this is a similar question I have been wondering, somewhere between a million Pi zeroes (still works out to zero? ;) ) and massive enterprise CPUs there is an optimum node size. What is that? For exilibris use case density and massive parallelism is important, for hobbyist applications manageable and economical with off the shelf hardware is a factor.
fvolk wrote:
Fri Feb 19, 2021 6:21 am
I think the DeskMinis are the reasonable trade-off in price/size/performance/cost when you need the performance.
I bought the H470 with 64GiB memory and paid almost 800 EUR (meanwhile memory got more expensive again :-/...)
It seems I can only get an AMD version here, I will try and understand the AMD CPU choices over the weekend. The options I have considered so far is 3 x NUC8i5 which right now is a bit less than 800EUR equiv and I can steal memory sticks from my H2s so they are "free", or for about the same 6xH2+ but I would need to buy more drives and memory sticks.

BTW my motivation is I want to develop a distributed optimization program. Ray looks like a good candidate so I'd start with that. I am wondering if you could elaborate on your earlier comment "multiple VMs and multiple physical nodes are not equal"? [Update] My last question is not very clear - I am wondering if there's a disadvantage to running in containers that have use of all cores of a node, other than some amount of overhead.

fvolk
Posts: 571
Joined: Sun Jun 05, 2016 11:04 pm
languages_spoken: english
ODROIDs: C2, C4, H2
Has thanked: 0
Been thanked: 61 times
Contact:

Re: Micro-server

Post by fvolk »

powerful owl wrote: It seems I can only get an AMD version here, I will try and understand the AMD CPU choices over the weekend.
The AMD version of the DeskMini got popular with HTPC users because they do HDMI 4K@60, while Intel only does HDMI 4K@30. Soon to be released 11.Gen Intel CPUs should change that in about a month - we'll see.
Note that AMD CPUs are typically 65W and some users complain in forums that AMD APUs at full load run against the limit of the DeskMini PSU - and the heat slowly cooking the system to early death? The 4000 APUs are unbuyable everywhere and the X300 seems to be a repainted A300 with just a new UEFI version. Google also for "deskmini a300 boot problem", "deskmini a300 hyperv problem" etc. How the situation with AMD will change when in summer the 5000 APUs are released, nobody knows. However, I will consider checking out a Cezanne based AMD APU DeskMini "for science"...
I wouldn't buy the AMD version right now and wait until summer.
powerful owl wrote: I am wondering if you could elaborate on your earlier comment "multiple VMs and multiple physical nodes are not equal"?
Multiple physical nodes teach the pain of replicating and synchronizing installs over "slow" 1Gbit network links.
Or fighting with a real PXE+diskless boot setup and the network load of rootfs provided via NFS (but then you don't need a local disk on node, only USB stick for boot)
Real network has latency, packet drops, or you could simply pull the network cable and see how the node and rest of the cluster behaves - and hopefully recovers - or not.
etc.


In order for this post to be more on topic for this forum and preventing Odroid punishing me for talking about other's products,
enjoy a board size comparison of a H2 with a DeskMini. This is why a H2 successor can't move "up" - that market is already taken?
h2dm.jpg
h2dm.jpg (99.42 KiB) Viewed 59 times

powerful owl
Posts: 201
Joined: Thu Mar 28, 2019 8:57 pm
languages_spoken: english
ODROIDs: 2 x HC1, 3 x H2, 1 x H2+
Has thanked: 58 times
Been thanked: 21 times
Contact:

Re: Micro-server

Post by powerful owl »

Thanks for the extra info! I was in fact having trouble finding a CPU with < 65W TDP. Yes, real nodes, makes sense.
If the H2 can't move up, can it be more flexible? Suppose ODroid made it with an L-shaped heatslug instead of a heatsink, like in the Willy Tarreau build farm. Then multiples could be attached to a heatsink or other cooling system for cluster computing (maybe also reduce the height by putting some ports on headers). For a single node, provide a heatsink which is easily attached to the slug. Now the whole thing is longer, which is actually good because it is a better match for the size of a hard drive, so then a simple case and backplane for two drives. No need for different boards for every usage model (e.g. XU4/HC1/MC1). Just thinking out loud...

Post Reply

Return to “The Ideas”

Who is online

Users browsing this forum: No registered users and 3 guests