Hacker News Reader

Comment by ilyt | Hacker News Reader

I always wanted such thing for various "plumbing" services (DHCP/DNS/wifi controller etc) but lack of ECC and OOB management kinda disqualifies it for anything serious.

>He's running forty Blades in 2U. That's:

    >
    >      160 ARM cores
    >      320 GB of RAM
    >      (up to) 320 terabytes of flash storage
    >

>...in 2U of rackspace.

Yay that's like... almost as much as normal 1U server can do

Edit: I give up, HN formatting is idiotic

View on HN

But does anyone remember the Beowulf trope(1) from Slashdot? Am I a greybeard now?

(1) https://hardware.slashdot.org/story/01/07/14/0748215/can-you...

I do. And cool research OSes that did process migration.

Ah, Plan 9.

I'm not sure that Plan 9 does process migration out of the box. It does have complete "containerization" by default, i.e. user-controlled namespacing of all OS resources - so snapshotting and migration could be a feasible addition to it.

Distributed shared memory is another intriguing possibility, particularly since large address spaces are now basically ubiquitous. It would allow users to seamlessly extend multi-threaded workloads to run on a cluster; the OS would essentially have to implement memory-coherence protocols over the network.

If not Plan 9, then likely Inferno. (A pretty different system, of course.)

And God help us, OS2/Warp.

With much better tooling for OO ABI than COM/WinRT will ever get (SOM).

It's too bad Apple never bought Gerry Popek's LOCUS (https://en.wikipedia.org/wiki/LOCUS), which could do process migration between heterogeneous hardware!

I do but in all fairness, I have an entirely grey beard.

Grey! Mine is white.

To go along with "Imagine a Beowulf cluster of those!", don't forget "Take my money!"

You can get off my lawn now.

I do and you are. I’m also imagining one covered in hot grits…

New to me - found the source article in the Wayback Machine:

https://web.archive.org/web/20010715201416/http://www.scient...

Found the definition:

Sterling and his Goddard colleague Donald J. Becker connected 16 PCs, each containing an Intel 486 microprocessor, using Linux and a standard Ethernet network. For scientific applications, the PC cluster delivered sustained performance of 70 megaflops--that is, 70 million floating-point operations per second. Though modest by today's standards, this speed was not much lower than that of some smaller commercial supercomputers available at the time. And the cluster was built for only $40,000, or about one tenth the price of a comparable commercial machine in 1994.

NASA researchers named their cluster Beowulf, after the lean, mean hero of medieval legend who defeated the giant monster Grendel by ripping off one of the creature's arms. Since then, the name has been widely adopted to refer to any low-cost cluster constructed from commercially available PCs.

But does it run Doom?

Crysis

What does Natalie Portman need to imagine a Beowulf cluster of Dooms running Crysis? Grits?

Natalie Portman says yes, and instructs you to put some hot grits down your pants.

Don't forget the Hot Grits & Natalie Portman.

Beowulf clusters were those lame things that didn’t have wireless, and had less space than a nomad, right?

It remember building a 4 node Beowulf cluster out of discarded compaq desktops and then having no idea what to do with it.

75mhz, yeah! Stacked on top of each other! With 10mbit ethernet! I think we got OpenMosix going even.

But then 5 years later I was working on them for a living in HPC, but they were no longer called Beowulf Clusters then.

Did kinda the same thing but with Raspberry Pis. Neat, a cluster of r-pi's... now what?

If you want to continue the chain of specific goals in service of no specific purpose: run Kubernetes on it.

Then add ArgoCD for deployment and istio for a service mesh!

While you are at it, also setup Longhorn for storage. With that solved, you might as well start hosting Gitea and DroneCI on the cluster, plus an extra helm- and docker repo for good measure. And in no time you will have a full modern CI/CD setup to do nothing but updates on! :-)

Seriously, though, you will learn a lot of things in the process and get a bottom up view of current stacks, which is definitely helpful.

I did this. I am still dead inside. Thank goodness all my production shit has a managed control plane and network.

Alternatively, if you dont like the kube's learn yourself some erlang, and make a super fault tolerant application.

Step 1: run a page ranking algorithm using Naive Bayes on a bastardised HPC framework

Step 2: add advertising

Step 3: make more money than God.

You and me both. The funny thing is, I wound up writing a program that would benefit from clustering, and felt my way around setting up MPICH on my zoo. I laughed out loud when I realized that, after all these years, I'd built an impromptu Beowulf cluster, even though the machines are scattered around the house.

Installing MPICH from source instead of from your distribution is best if you can't have all your cluster members running the same version of the same distro and/or have multiple architectures to contend with. But it takes forever to compile, even on a fast machine.

well since you already have a cluster you could distcc the compilation. I remember doing that on a bunch of machines when we were building gentoo

That's more a COW, than a (Beo)wulf, no?

<https://www.oreilly.com/library/view/high-performance-linux/...>

A cluster of workstations (COW) is usually opportunistic exploiting existing systems, and lower density than a dedicated (usually rack-based or datacentre-based) cluster.

In practice, COWs usually turn out to be not especially useful, though there are exceptions.

yeah, wanted to replicate something like that by proposing a hardware vendor who visited my uni. decades ago. didn't go nowhere because i was intimidated by the red-tapes.

I like user big.ears' speculation on what someone could possibly do with that much parallel compute:

  I don't think there's any theoretical reason someone couldn't build a fairly realistic highly-complex "brain" using, say, 100,000,000 simplified neural units (I've heard of a guy in Japan who is doing such a thing), but I don't really know what it would do, or if it would teach us anything that is interesting.

Simplified neural unit? Less capacity than a human brain? Lame.

Lame, that just gives the opportunity to release for a higher price the Neural Unit Pro, Neural Unit Max, and Neural Unit ProMax. You consider it lame because you're acumen in business nuance is lame ;-)

Sure do! It is interesting that these technologies evolve more slowly than it seems, sometimes.

On the graybearding of the cohort, here’s a weird one to me. These days, I mention slashdot and get more of a response from peers than mentioning digg!

In 2005, I totally thought digg would be around forever as the slashdot successor, but it’s almost like it never happened (to software professionals… er, graybeards)

So, I once met a guy named Don.

We were hanging out in the garage of a mutual friend, chatting. Got to the "what do you do" section of the conversation, and he says he works in massively parallel stuff at XYZ corp. Something something, GPUs.

I make the obvious "can you make a Beowulf cluster?" joke, to which he responds (after a pregnant pause), "you... do know who I am?"

Yep. Donald Becker. A slightly awkward moment, I'll cherish forever.

Ew. "Do you have any idea who I am?" is never a good look.

I do think this is sort of fool's gold in terms of actual performance. Even though the core count and RAM size is impressive, those cores are talking over ethernet rather than system bus.

Latency and bandwidth is atrocious in comparison, and you're going to run into problems like no individual memory allocation being able to exceed 8 Gb.

Like for running a hundred truly independent jobs then sure, maybe you'll get equivalent performance, but that's a very unique scenario that is rare in the real world.

I built such a toy cluster once to see for my self and gave up. It is too slow to do anything serious. You can be much better off by just buying older post lease server. Sure it will consume more power, but conversely you will finish more tasks in shorter time, so advantage of using ARM in that case may be negligible. If it was Apple's M1 or M2, that would have been a different story though. RPi4 and clones are not there yet.

I overall think people tend to underestimate the overhead of clustering. It's always significantly faster to run a computation on one machine than spread over N machines with hardware of (1/N) power.

That's not always a viable option because of hardware costs, and sometimes you want redundancy, but those concerns are on an orthogonal axis to performance.

Well, a complete M1 board, which is basically about as large as half an iPhone mini, is fast enough. It's also super efficient. So I'm still waiting for Apple to announce their cloud.

They're currently putting Mx chips in every device they have, even the monitors. It'll be the base system for any electric device. I'm sure we'll see more specialized devices for different applications, because at this point, the hardware is compact, fast, and secure enough for anything, as well as the software stack.

Hello Apple Fridge

Fast enough for what?

Lines gets blurred when you are on a supercomputer interconnect and a global address space or even rdma

the fastest practical interconnects are roughly 1/10th the speed of local RAM. Because of that, if you use interconnect, you don't use it for remote RAM (through virtual memory).

I don't think anybody in the HPC business really pursued mega-SMP after SGI because it was not cost-effective for the gains.

Cray did some vector machines that were globally addressed but not coherent. That’s an interesting direction. So is latency hiding.

The really important thing is that the big ‘single machine’ you’re talking about already has numa latency problems. Sharing a chassis doesn’t actually save you from needing to tackle latency at scale.

Both Single System Image and giant NUMA machines were and are still pursued because not everything scales in shared-nothing message passing well (some stuff straddles it by doing distributed shared memory over MPI but using it mostly for synchronisation).

It's just that there's a range of very well paying problems that scale quite well in message passing systems, and this means that even if your problem scales very badly on them, you might have easier time brute forcing the task on larger but inefficient supercomputer rather than getting funding for smaller more efficient one that fits your problems better.

It probably lends itself to tasks where CPU time is much greater than network round trip. Maybe scientific problems that massively parallel. Way back in the 90s I worked with plasma physics guys that used a parallel system on "slow" Sun boxes. I can't remember the name of the software though.

>I do think this is sort of fool's gold in terms of actual performance.

It’s a fun toy for learning (and clicks, let’s be honest).

It’s not a serious attempt at a high performance cluster or an exercise in building an optimal computing platform.

Enjoy the experiment and the uniqueness of it. Nobody is going to be choosing this as their serious compute platform.

In TFA, isn't Jetbrains using it as a CI system?

Tangential, but it is so funny to me that “TFA” has become a totally polite and normal way to refer to the linked article on this site. Expanding that acronym would really change the tone!

I'm not sure it is 'totally polite'? I usually read it as having a 'did you even open it' implication that 'OP' or 'the submission' doesn't. Maybe that's just me.

Maybe it isn’t totally polite, but it IMO it reads in this case more like slight correction than “In the fucking article,” which would be pretty aggressive, haha.

I always thought it meant “the featured article”; I must be a lot more wholesome than previously indicated!

Unless they need something Pi specific I don't understand why this would be preferable versus just virtualizing instances on a "big ARM" server. I'm sure those exist.

It's actually 3U since the 2U of 40 pis will need almost an entire 1U 48 port PoE switch instead of plugging into the TOR. The switch will use 35-100W for itself depending on features and conversion losses. If each pi uses more than 8-9W or so under load, then you might actually need a second PoE switch.

If you are building full racks, it probably makes more sense to use ordinary systems, but if you want to have a lot of actual hardware isolation at a smaller scale, it could make sense.

In some colos, they don't give you enough power to fill up your racks, so the low energy density wouldn't be such a bummer there.

> Yay that's like... almost as much as normal 1U server can do

Hyperscale in your Homelab. Something to hack on, learn, host things like Jellyfin, and have fun with.

I agree but can't you get the same effect with VMWare ESXi? If I just wanted to "have fun" managing scores of tiny computers, and I emphasize that this sounds like the least amount of fun anyone could have, I can have as many virtual machines as I want.

I can understand why some people want something physical/tangible while testing or playing in their hobby environment. I'm still a fan of virtualization - passmark scores for an RPi4 (entire SOC/quad core) are 21 times less than a per-single-core comparison in a 14-core 15-13600k (as a point of reference, my current system) and while am running 64GB RAM, can easily upgrade to 128GB or more on a single DDR4 node.

Hard to see to an advantage given obvious limitations, although it may make it more fun to work within latency and memory constrictions, I guess.

Haha jellyfin would eat through all your memory and cpu time transcoding or remuxing on a SBC. I'm seriously thinking of getting another home server just to run that.

It's ~~four~~ two spaces to get the "code block" style.

    like
    this

and asterisk for italics (I don't think there is a 'quote' available, and I'm not sure how they play together.

* does this work? * Edit: No! Haha

    *how*
    *about*
    *this*

Edit: No, no joy there either.

I agree, it's not the most intuitive formatting syntax I've come across :)

I guess we're stuck with BEGIN_QUOTE and END_QUOTE blocks!

It’s two spaces. https://news.ycombinator.com/formatdoc

The code block was workaround

I wanted to do this

    > * list element 1
    > * list element 2

without the indent.

I don't get why it just doesn't use common mark or something, it's just some inept, half-assed clone of it anyway.

> but lack of ECC and OOB management kinda disqualifies it for anything serious.

> Yay that's like... almost as much as normal 1U server can do

It’s a fun toy. Obviously it isn’t the best or most efficient way to get any job done. That’s not the point.

Enjoy it for the fun experiment that it is.

> I always wanted such thing for various "plumbing" services (DHCP/DNS/wifi controller etc)

You don't need a cluster for that, even a 1st gen Pi can run those services without any problem.

I can only speak for Raspi 3B+, but I agree.

I have multiple services running on it (including pihole, qbittorrent, vpn) and it's at about 40% mem usage right now.

I need multiple nodes for redundancy, not because it can't fit on one. And preferably at least 3 for quorum too.

Indeed, that box here next to my desk draws 50W of electricity continuously despite being mostly idle. Why? Because it has ECC.

Having some affordable low power device with ECC would be a game changer for me.

I added affordable to exclude expensive (and noisy) workstation class laptops with ECC RAM.

Xeon-D series?

Epyc Embedded and possibly some Ryzen Embedded devices.

> Why? Because it has ECC

Sorry if I am missing the obvious here, but why would ECC consume so much power?

It's not that ECC consumes power, it's that systems that support ECC tend to consume more idle power (because they're larger etc.)

There are Intel Atom CPUs that support ECC. I had a Supermicro motherboard with a quad core part like that and I used it as a NAS. It was not that fast, but the power consumption was very low.

Do you remember how many Watts it was using with idle disks?

It was this board: https://www.supermicro.com/en/products/motherboard/a2sdi-2c-...

I think it was idling at something like 30-40W with four HDDs and a UPS. I didn't have an especially efficient PSU and the UPS must have taken some power too. The motherboard alone would draw as little as 15W, I suppose.

I personally have at 43-45W idle…

    >Corsair SF450 PSU
    >ASRock Rack X570D4U w/BMC
    >AMD Ryzen 7 Pro 5750GE (8C 3.2/4.6 GHz)
    >128GB DDR4-2666 ECC
    >Intel XL710-DA1 (40Gbps)
    >LSI/Broadcom 9500-8i HBA
    >64GB SuperMicro SATA DOM
    >2 SK Hynix Gold P31, 2TB NVMe SSD
    >8 Hitachi 7200rpm, 16TB HDD
    >3 80mm fans, 2 40mm fans, CPU cooler

That was an at the time modern “Zen 3” (using Zen 2 cores) system on an X570 chipset. The CPU mostly goes in 1L ultra SFF systems. TDP is 35W, and under stress testing the CPU tops out around around 38.8-39W. The onboard BMC is about 3.2-3.3W of power consumption itself.

Most data ingest and reads comes from the SSD cache, with that being more around 60W for high throughput. Under very high loads (saturating the 40Gbps link) with all disks going, only hits about 110-120W.

By comparison, a 6-bay Synology was over double that idle power consumption, and couldn’t come close to that throughput.

thanks for the parts list, especially because I think ASRock Rack paired with a Ryzen Pro offers better performance than a Supermicro in the same price range.

There’s reasons for that though.

I could drop a few more watts if ASRock could put together a decent BIOS where disabling things actually disables things.

SuperMicro costs what it does for a reason.

—- ————-

If you’re looking for a chassis, I’m using a SilverStone RM21-308, with a Noctua NH-L9a-AM4 cooler, and cut some SilverStone sound deadening foam for the top panel of the 2U chassis.

Aside from disks clicking, it’s silent, runs hilariously cool (I 3D printed chipset and HBA fan mounts at a local library) and it’s more usable storage, higher performance (saturates 40Gbps trivially) and lower power consumption than anything any YouTuber has come remotely close to. That server basically lets me have everything else in my rack not care much about storage, because the storage server handles it like a champ. I really considered doing a video series on it, but I’m too old to want to deal with the peanut gallery of YouTube comments.

If you don't mind me asking, how do your other workloads access the storage on it, NFS? The stumbling block for NFS for me is identity and access management.

Wow I just picked up an ASRock Rack X570D4U and put my 5950X into it.

Do you know how to make the BMC not a laggy mess when using the “H5Viewer”? I’m getting basically unusable latency when the system is two yards away compared to a RDP server 1,000 miles away.

That's impressively low, considering the amount of storage capacity and the performance potential for the time you need it. It goes a long way towards paying for itself if you replace some old Xeon server with it.

Most AMD desktop platforms support ECC, and if you don't use overclocking facilities, they are pretty efficient (though their chiplet architecture causes idle power draw to be a good fraction of active power draw, still much less than 50W though).

How much RAM is that with? My home server idles at ~25-27W, but that's with only 16GB (EEC DDR4). However, throwing in an extra 16GB as a test didn't measurably change the reading.

Didn't AMD announce a 96 core processor with dual socket support?

As usual this is either done for entertainment value or to simulate physical networks (not clusters).

Intel also has now up to 480 cores in an 8-socket server with 60 cores per socket, though Sapphire Rapids is handicapped in comparison with AMD Genoa by much lower clock frequencies and cache memory sizes.

However, while the high-core-count CPUs have excellent performance per occupied volume and per watt, they all have extremely low performance per dollar, unless you are able to negotiate huge discounts, when buying them by the thousands.

Using multiple servers with Ryzen 9 7950X can provide a performance per dollar many times higher than that of any current server CPU, i.e. six 16-core 7950X with a total of 384 GB of unbuffered ECC DDR5-4800 will be both much faster and much cheaper than one 96-core Genoa with 384 GB of buffered ECC DDR5-4800.

Nevertheless, the variant with multiple 7950X is limited for many applications by either the relatively low amount of memory per node or by the higher communication latency between nodes.

Still, for a small business it can provide much more bang for the buck, when the applications are suitable for being distributed over multiple nodes (e.g. code compilation).

This is the exact space I’m in, high cpu low network. By my estimates it’s about 1/4 the cost per CPU operation to use consumer hardware instead of enterprise. The extra computers allow for application level redundancy so the other components can be cheaper as well.

One problem with 480 cores in single node: 480 cores is a shitload of cores, who needs more than a single node at this point? The MPI programmer inside me is having an existential breakdown.

I think the hardware isolation would be a selling point in some cases. Granted, it's niche.

In most cases it is wrong choice unless your business is selling raspberry pi hosting I guess

> Yay that's like... almost as much as normal 1U server can do

...but the normal server is much cheaper.

From the FAQ: https://news.ycombinator.com/formatdoc

its a nice hobby project, but of course a commercial blade system will have far higher compute density. supermicro can do 20 epyc nodes in 8u, which at 64 cores per node is 1280 cores in 8u, or 160 in 1u, so double the core density, and far more powerful cores, so way higher effective compute density.

Also not noted: 320 TB in 40 M.2 drives will be extremely expensive. Newegg doesn't have any 8 TB M.2 SSDs under $1000. $0.12/GB is about twice as expensive as more normally-sized drives, to say nothing of the price of spinning rust.

> Yay that's like... almost as much as normal 1U server can do

What about cost, and other metrics around cost (power usage, reliability)? If space is the only factor we care about then it seems like a loss.

What about them? 1U servers from vendors are reliable and efficient - people use them in production for years. As for the cost, those hobby-style board are very expensive for dollars/performance. Indeed I'm not getting why would one want a cluster of expensive, low spec nodes?

Just the Pi’s are $35 a pop, right? So that’s $1400 of Pi’s, on top of whatever the rest of the stuff costs. Wonder how it compares to, I guess, a whatever the price equivalent AMD workstation chip is…

It seems they're the ones with 8 GB of ram, so probably closer to $75 each.

I’d be interested to see if anyone had any application other than CI for Raspberry Pi programs, I really can’t see one.

What are you talking about "lack of ECC?" The Pi4 has ECC.

Didn't knew that! Would still prefer one that actually reports the errors tho

That would be 40x (Rpi4 8GB $75 + 8TB nvme $1200 + psu and others) ~ $51000.

> lack of ECC and OOB management kinda disqualifies it

Can you expand on this please?

ECC RAM is more robust (fewer crashes due to random bit flips), and OOB management means if a server has issues, you can remotely view it as if you were jacked in, and force reboot, among other things (like installing an OS remotely).

The 1U server is however likely to use more than 200 Watts of power that the 40 Blade 2U setup would use.

> The 1U server is however likely to use more than 200 Watts of power

Q: Why would a 1U server need more than 200W if you're doing nothing more than basic network services?

I have mini tower servers that draw a fraction of that at idle.

I have some idle Dell R650's that draw 384W. A couple drives, buncha RAM, two power supplies, 2 CPU's (Xeon 8358)

> Dell R650's that draw 384W

Umm, I'm not sure I can afford the electricity to run kit like that :)

I'm currently awaiting delivery of an Asus PN41 (w/ Celeron N5100) to use as yet another home server, after a recommendation from a friend. Be interesting to see how much it draws at idle!

Hrm, interesting to see how the TDP of those 8358s drives overall power consumption. I'm looking at the idrac consoles of a couple R720XDs with 12 3.5" hdds, >128gb ram, two E5-2665s per, and they're all currently sipping ~150W at < 1 load average. The E5s have a TDP of 115W to the 8358's 250W, so I assume that's what's most of it. I admittedly do some special IPMI fan stuff, but that only shaves off tens of watts.

The Pi's will be using those 200Watts at near full tilt. The main use here would be larger computational tasks that you can easily split up among the blades. Or you run a very hardware-failure tolerant software service on top.

Reply on news.ycombinator.com