ARMed Servers Going Into Production

“you can now provision hourly, on-demand ARMv8 servers powered by 2 x 48 Cavium SoC’s on Packet. We’re starting with our EWR1 home (New York metro), as well as Sunnyvale and Amsterdam. We’ll add in Tokyo in early December when the facility opens for production customers.
 
The amazing thing about this server (which also features 128 GB of DDR4 RAM and 320 GB of m.2 SSD flash) is the incredibly low cost per core: at $0.50/hr for 96 cores, you’re paying about 1/10th the cost per core (per hour) versus our Intel-based Type 2 machine.”
 
See ARM’ing the World with an ARM64 Bare Metal Server
While some folks waste time and energy deriding ARM CPUs in servers at the same time billions adopt ARM for all kinds of workloads, others quietly go about installing many thousands of ARMed CPUs in their data-centres. It’s all good. The dinosaurs didn’t notice the mice either but we know who inherited Earth… 😉

About Robert Pogson

I am a retired teacher in Canada. I taught in the subject areas where I have worked for almost forty years: maths, physics, chemistry and computers. I love hunting, fishing, picking berries and mushrooms, too.
This entry was posted in technology and tagged , , , , , , , . Bookmark the permalink.

37 Responses to ARMed Servers Going Into Production

  1. dougman wrote, “That HP was a 90s machine, slow as heck, reliable to a degree, but now mainly just kept for the giggle factor.”

    Yes, most HP machines work well with GNU/Linux. However, TLW replaced a Xerox machine that I chose with another that I did not when the first one failed… I was in the North… Went from sanity to x86-only drivers. The thing scans to PDF and the FTP server nicely but only prints if spoken to in a “proprietary” language.

    “Hardware Environment:
     
    o Computer with x86 compatible 32-bit CPU, on which each distribution operates.

    Host-based, PCL® 6 emulation”

    Yep, not what I would have chosen…

  2. dougman says:

    “It’s an old Xerox printer. Works beautifully but it speaks an ancient language.”

    I have a HP LaserJet 4 downstairs, no need for a print driver. That HP was a 90s machine, slow as heck, reliable to a degree, but now mainly just kept for the giggle factor.

  3. dougman wrote, “Printers Servers? What is this the 90s??”

    Nope. This is about drivers. I want only one copy of an x86-only driver on my LAN when machines are AMD64 and ARM. It’s an old Xerox printer. Works beautifully but it speaks an ancient language.

  4. dougman says:

    Printers Servers? What is this the 90s??

    Nowadays people buy printers and plug them into the network via cable or wireless, as the printers already have the server built-in.

    I do still recall making a MSCE idiot mad, when I showed his client how to deploy a few linux computers and pull and install the Xerox ppd. Boy, was he made at me, I guess he thought I was pissing all over his work and network. Which I was and happily smiling while doing it. When he confronted me about it, the only response I could formulate was “Bro, do you even lift?” These Windows always going down.

  5. Further, check out Solid-Run’s “Use Cases” for the Armada 8040 Community Board: Notice NAS, Operators like me, and Web/Printers Servers…

  6. Deaf Spy, moving goalposts, wrote, “Robert, is nothing. Waste of time.”

    The real question is whether or not Odroid-C2 can be a server, not that it can serve Google. You, twit, are the waste of time.

  7. Deaf Spy wrote, “just having a USB 3.0 port soldered on a motherboard means nothing. It may be compatible with USB 3.0 devices, but it may, for example, never reach the speed advertised. Same goes for SATA, and Ethernet, and everything.”

    You have no evidence that the motherboard does not live up to specs. The block diagram suggests it does. Odroid-C2, for instance, has benchmarks that show it works at gigabit/s speeds. e.g. TLW’s Odroid-C2 can copy 93 MB/s over gigabit/s link. I expect the newer/faster boards will do even better. Oh, yes, TLW left a video looping all night with 100% CPU on one core and it still manages 93MB/s copy on NFS.

    Being a little low on specs does not even support your argument. A server is a server whether it chokes networks or not. A server may well have to share traffic with other servers/clients/devices so even the best machine may not reach specs in the real world.

  8. Deaf Spy says:

    What makes those capabilities inept?

    That fact that they are glued together with some very different purpose in mind. You see, Robert, just having a USB 3.0 port soldered on a motherboard means nothing. It may be compatible with USB 3.0 devices, but it may, for example, never reach the speed advertised. Same goes for SATA, and Ethernet, and everything.

    Real life example. Many modern motherboards feature an M.2 slot for SSDs. That is handy, and performance exceeds SSD by at least 50%. But. There is always but, isn’t it? It greatly depends on the design of the motherboard, and how many PCI lanes are reserved for the M.2. Most have only 2, meaning you can’t get speed greater than 10Gbps. Which is great for most SSDs out there, today. But, put a high-end SSD, and you are busted. You won’t get more than 900 MB/s, while the SSD may do three times more. For that, you need an extra card that you plug in a PCIe x4 slot at least. Or, get a better motherboard.

    Does that starts ringing bells, Robert?

  9. Deaf Spy says:

    Well, the most popular web-serving software on Earth disagrees with you…

    Sigh. See, Robert, the tool you decided to use is ok. The problem is what you tested.
    Document Length: 2277 bytes…Total transferred: 237100 bytes
    You tested two megabytes of static content in total. You tested with a 2K static document.

    This, Robert, is nothing. Waste of time.

    To have a meaningful test, you need as a minimum:
    1. Dynamic content. Or, you don’t test you CPU.
    2. Content of variable size, not some tiny page.
    3. Dynamically change the number of concurrent connections to identify how your system scales.

  10. oiaohm says:

    PPS: Microsoft is already pressing steam ahead with stereoscopy and VR (only OS which has good support for both) and 3D immersive UI, while GNU/Linux can barely be trusted to support the latest OpenGL (depends on instance). Lol. Are we staying a big behind the times Pog?
    Not in fact true. Both stereoscopy and VR start on Linux. Both part of the closed source Nvidia driver. In fact Windows also has trouble with VR just like Linux in a lot of ways.

    http://www.vronlinux.com/articles/the-khronos-group-announces-vr-standards-initiative.43
    This shows the problem with VR applications to support everything have to support at least 6 different interfaces.

    https://www.gamingonlinux.com/articles/steam-vr-will-use-vulkan-on-linux-demo-shown-off.8316
    Also Valve has demoed VR on Linux and they have stated they will use Vulkan not opengl.

    When it comes to VR its a mess. Does not matter if it Windows or Linux. Opengl single threaded nature does not really agree with VR. Yes Opengl VR might have been the first but its nowhere near nice. Direct X multi threaded nature gave Microsoft some advantage of course Vulkan has that as well.

    not risk breaking the proprietary wifi, bluetooth and gps drivers every time you update the server because the driver ABI is table (a problem with Android and gnu/linux).
    Majority of android devices being phone and tables don’t use kernel space closed source wifi, bluetooth or gps drivers because it makes no sense.
    http://hackerboards.com/raspberry-pi-shields-offer-3g-4g-gprs-gps-and-xbee-support/
    This shows how the common versions of wifi bluetooth and gps hook up to a arm cpu in a phone/tablet most of the time. This does not matter if the phone is Windows Android, OSi or some other strange OS. Wifi, Bluetooth and GPS in most of the portable devices is sitting in a single serial port connect to the processor. A standard define multiplexed serial connection that is part of the Linux kernel and most operating systems in fact. Yes it is the old school AT dial-up modem with ppp on steroids.

    The drivers to blame most often for not being able to update Android devices kernels is display drivers(yes this is a hornets nest on arm of 20+ vendors) and audio(another 200 vendors).

    Also if you are after Wifi/Bluetooth performance you will stay clear of anything closed source driver. Making wifi fast project point out the testing and design errors in closed source wifi/bluetooth drivers. If you are having trouble with closed source wifi and bluetooth drivers under Linux you are using hardware not designed for Linux and don’t care about performance these days. There are tones of choices of machines avoiding these problems.

    kurkosdr so this is you bring up and example of something that was an issue that is not any more. Was a issue until it was found out that closed source really could not design formal for traffic transfer in secret that worked.

  11. Deaf Spy wrote, without any reason, “What you also fail to grasp is that all the peripherals you choose so far (motherboard, extension slots, etc.) are totally inept for a server.”

    10Gb-E, USB3, Sata 3 are all commonly found on servers. Why not mine? What makes those capabilities inept? Should not a server by sized rightly? Should I use a floor-to-ceiling stack to keep my reloading notes, medical appointments and correspondence? No. A tiny motherboard with a good processor, RAM and peripherals is all that’s needed as has always been the case.

  12. Deaf Spy wrote, “This is one weird benchmark here”.

    Well, the most popular web-serving software on Earth disagrees with you…

    See ab, Apache Benchmark…“ab is a tool for benchmarking your Apache Hypertext Transfer Protocol (HTTP) server. It is designed to give you an impression of how your current Apache installation performs. This especially shows you how many requests per second your Apache installation is capable of serving.”

  13. kurkosdr says:

    every time you update the server = every time you update the kernel (autocorrect)

  14. kurkosdr says:

    Newsfeed: Windows 10 is SmallCheapComputer-compatible(tm), and now with the ability to run win32 apps on those SmallCheapComputers too:
    http://arstechnica.co.uk/information-technology/2016/12/windows-10-arm-x86-details/
    And since Windows 10 is free for computers with 9-inch monitors or less, it is also cheap on the software side.

    So, Windows 10 gives you all the Smallness and Cheapness you want in your computing, along with the advantages of Windows, such as patches not being withheld by OEMs (a problem with Android), the OS not withholding root from the buyer and requiring exploiting vulns or flashing a new ROM to get root (a problem with Android), better power management (a problem with Android and gnu/linux) and not risk breaking the proprietary wifi, bluetooth and gps drivers every time you update the server because the driver ABI is table (a problem with Android and gnu/linux).

    So, considering Android’s momentum, iOS’s momentum and Windows 10 running win32 on ARM, this leaves client-side GNU/Linux the advantage of… the advantage of… ehm… can somebody help a bit here?

    Let’s be honest for a moment: Squezzed between Android, iOS and Windows 10, is there a market for GNU/Linux, beyond the 1%er market of people who demand to “see the source codes”?

    PS: I am already salivating about that surface phone that could dock to my Bravia and run win32 apps.

    PPS: Microsoft is already pressing steam ahead with stereoscopy and VR (only OS which has good support for both) and 3D immersive UI, while GNU/Linux can barely be trusted to support the latest OpenGL (depends on instance). Lol. Are we staying a big behind the times Pog? Thinking that setting up a little low-end server is something worth bragging about, while us Windows users are already enjoying stereoscopy and marching on to VR?

  15. Deaf Spy says:

    How is it possible to run a demanding desktop

    Who says a demanding desktop, Robert?

    What this news and analysts basically say is that ARM is good to run a lowly desktop. Which doesn’t translate to a lowly server.

    What you also fail to grasp is that all the peripherals you choose so far (motherboard, extension slots, etc.) are totally inept for a server. They have always some very specific application in mind, and server is not one of them.

    What you do is repurpose a lowish some purpose device into a server.

  16. Deaf Spy wrote, “Windows 10 now officially runs on ARM with full Win32 support”.

    …but you naysayers have just been telling the world that ARM isn’t good enough to run a lowly server. How is it possible to run a demanding desktop?!??

  17. Deaf Spy says:

    This is one weird benchmark here, but I will let it pass.

    The big news today is that Windows 10 now officially runs on ARM with full Win32 support. All analysts and OEMs look overjoyed, and speak of new desktop possibilities. No one even mentions that some Linux distros might be runnable on ARM, too.

    Wintel is dead. Long live Winarm, or Winqual, whatever you prefer.

  18. We use vsftpd on Beast. The worfer thing was just an experiment to show Odroid-C2 could do something. Unstable does not mean unusable. It just means the version is not fixed and could change before release. That’s part of the Debian packaging process. Stuff starts out in “experimental” flavour, moves to unstable, then to testing and finally to stable. It has nothing to do with the stability of the software in use, just the version/edit/copy. I’ve used testing in production quite regularly. It’s a bit like Ubuntu. You know, a lot of nice features mixed in with bugs but if the bugs are not in the package you use you don’t care. This is one of the nice features of Debian. If you don’t want to fool around with software, just use stable and everything is about two years old and as reliable as Hell.

  19. dougman says:

    I have FTP running on my UNRAID box. https://imgur.com/a/XkNXE Big deal rolls-eyes

    That webork thingy is not even at version one yet, so its alphaware. The ftp server I am using is https://security.appspot.com/vsftpd.html, and is stable.

    Interestingly, your webdork software you use is not even mentioned here: https://en.wikipedia.org/wiki/List_of_FTP_server_software

  20. I just installed “weborf” on Odroid-C2 and pointed it at a thin user-directory. Apache Benchmark from Beast says it’s a server:
    Server Software: weborf/0.13
    Server Hostname: odroid-c2
    Server Port: 8080

    Document Path: /
    Document Length: 2277 bytes

    Concurrency Level: 4
    Time taken for tests: 0.145 seconds
    Complete requests: 100
    Failed requests: 0
    Total transferred: 237100 bytes
    HTML transferred: 227700 bytes
    Requests per second: 687.65 [#/sec] (mean)
    Time per request: 5.817 [ms] (mean)
    Time per request: 1.454 [ms] (mean, across all concurrent requests)
    Transfer rate: 1592.21 [Kbytes/sec] received

    Connection Times (ms)
    min mean[+/-sd] median max
    Connect: 0 1 0.2 1 1
    Processing: 1 5 2.6 5 11
    Waiting: 1 2 1.0 2 7
    Total: 2 6 2.6 6 12

    Percentage of the requests served within a certain time (ms)
    50% 6
    66% 7
    75% 8
    80% 8
    90% 9
    95% 10
    98% 11
    99% 12
    100% 12 (longest request)

    Oh, yes, that’s while it was running a desktop, GUI, browser, etc. Chuckle. The Armada 8040 Community Board can be a server because it will blow the Odroid-C2 away in performance.

  21. dougman wrote, “Your definition of a server is unsavory at best.”

    “Unsavory” certainly is an ugly word. My definition comports fairly well with others. Does yours? Do you actually have a definition or are you just posting strawmen?

    Free On-line Dictionary of Computing: “A computer which provides some service for other computers connected to it via a network. The most common example is a file server which has a local disk and services requests from remote clients to read and write files on that disk, often using Sun’s Network File System (NFS) protocol or Novell Netware on PCs. Another common example is a web server.”

    Are you suggesting the Armada 8040 Community Board or the Cello cannot do all of those? I assure you the Odroid-C2 certainly can and it is a tiny box compared to these other two. They are flexible, general-purpose computers with added features for specific tasks like virtualization and networking.

    Do you agree Cello and Armada 8040 Community Board are much more resourceful than Odroid-C2? Then you’ll appreciate this guy who got Docks of cloud-servers running on Odroid-C2. You’ll enjoy following his problems of finding a GNU/Linux that worked smoothly. As I did he went with Armbian.
    “I downloaded and deployed the Raspberry Pi busybox httpd (web) Server image in the docker container and it worked. Next, I instructed Docker to use my web content instead of the default content”

    So, if one can install that flexible virtual system on Odroid-C2 you can install just about any service you want. I don’t intend to do that. I will use Armbian on the Odroid-C2 as a thin client OS and run my processes on the new Beast. TLW is doing fine with the Odroid-C2 as a thick client. Since Linux 4.4 will run on the Armada 8040 Community Board, I don’t expect much trouble in porting all my stuff over to it.

  22. dougman says:

    Your definition of a server is unsavory at best.

  23. DrLoser wrote, “not a single one of us who think you have any idea what you are talking about”.

    I guess that’s due to a generally low standard of reading. I’ve written plainly about my need for a new server.

  24. dougman says:

    “Next post please.”

    Yes, this is becoming tiring. We need more stimuli so as to prepare new mockery, to counter illogic and pertinacity on your part. Trying to elicit an emotional response from you is the resulting joy and sadness we seek.

  25. DrLoser says:

    You’re on a bit of a loser on this one, apparently, Robert. Taking all the conversations down the line into account. Ain’t not a single one of us who think you have any idea what you are talking about.

    Next post please.

  26. The Wiz wrote, “Neither the desktop or the server market is being pursued”.

    “Gateway – Paired with Marvell's industry-leading wireless and wired networking solutions, the ARMADA® 8K Family is perfectly suited for Enterprise and Home Gateway applications delivering the right balance between performance, power and price, along with key features such as security acceleration and virtualization.”

    Home Gateway suggests the routing, LAN infrastructure, and caching role that Beast does. “performance, power and price” suits me and my needs just fine. In what way is the role of a small server not encompassed? The whole idea of Cortex A processors is powerful application processors of which A-72 is nearly state of the art in ARM. You guys were saying A-57 was whimpy. Compared to that, this thing rocks. About the only features of it I don’t like is that it only has one memory slot and it’s 28nm. Otherwise it’s about perfect.

    “ARM Cortex-A processors are at the heart of the most powerful and compelling technology products. They are deployed in mobile devices, networking infrastructure, home and consumer devices, automotive in-vehicle infotainment and driver automation systems, and embedded designs.”

  27. Wizard Emeritus says:

    “They are thinking like I’m thinking…”

    Really?

    yet on the data sheet you cited, we find as their envisioned uses:

    “Control Plane Card
    – Paired with Marvell’s industry-leading wired networking solutions, the RMADA® 8K Family is perfectly suited for next generation contro
    l plane cards, delivering the right CPU performance and
    memory throughput, along with key features such as 64b processing and virtualization.

    Gateway
    – Paired with Marvell’s industry-leading wireless and wired networking solutions, the ARMADA® 8K Family is perfectly suited for Enterprise and Home Gateway applications delivering the right balance between performance, power and price, along with key features such as security acceleration and virtualization. ”

    Neither the desktop or the server market is being pursued, Robert Pogson
    These people are going after the network appliance market.

    nothing to see here.

  28. “the industry’s desire for an alternative to Intel remains. I’m hopeful that the years of investment in the ARM ecosystem can translate into positive market momentum over the next couple of years”.

    Well, the technology is all anchored to ARM so it can be sold/bought by someone actually interested. Apparently Marvell is interested. They are thinking like I’m thinking

  29. dougman says:

    FIfi, go back to sleep.

    “ARM-based servers have been hyped in the market for 6-plus years, with little to show for it in terms of real customer adoption,”

    http://www.theregister.co.uk/2016/12/07/broadcom_arm_processor_vulcan/

    nm wall?? Don’t you mean death of Moore’s Law? LOL…there’s a law about Moore’s law, it goes something like this. “The number of people predicting the death of Moore’s law doubles every two years.”

  30. oiaohm says:

    http://www.anandtech.com/show/10353/investigating-cavium-thunderx-48-arm-cores/20
    Low price per core… Great, if I need to cache static content and data. Poor, when I need computational power.

    http://www.anandtech.com/show/10353/investigating-cavium-thunderx-48-arm-cores/13
    Deaf Spy the issue here is per core processing power of even the older thunderx is 1/3 of the Xeon. Problem is the thunderx can run 8 cores for the power usage of 1 Xeon core. Remember the arm cores are not hyper-threading. So a single threaded task on a Xeon can out do the arm but multi threaded its not big lead at all.

    Remember due to nm difference Xeon is twice as fast due to it. So when the arm and x86 chips are on the same nm things get a lot more interesting.

    Even with the older thunderx different workloads out run the Xeon while using less power.

    So computational power is questionable if arm64 or x86 is faster. Multithreaded workloads arm64 wins small core size means more cores for same area of silicon and this results in event that the are cores are slower the xeon is run into ground.

    I guess when you say need computational power you need single threaded power. Arm64 does well in the server roles because that is mostly multi threaded workloads that is advantage to arm vs x86 at the same nm.

    Even in desktop multi threaded is quite a important workload. MIPS cores are also smaller than x86 and have the same thing where they end up in multi threaded workloads competitive and not so much in single threaded.

    Deaf Spy remember any thing majority processing big data like 3d rendering or video rendering multi threading is quite important.

    I do expect intel to be able to hold lead for a while yet. Problem is there is the nm wall. When we get to the nm wall it going to become a very interesting dog fight.

    Power, Mips, Arm and x86 all will go at each other throat when the nm wall is hit.

  31. DrLoser says:

    Precisely how much time and energy does it take, Robert?

    Here’s my guess. About ten seconds, and 200 Joules. Which is roughly what I have spent posting this.

    Somehow I get the feeling that you are the one wasting your time and energy. Which is fine, except that you have no idea what you are talking about.

    And you are rather ancient, which means that your time is increasingly limited. Your energy might be better employed elsewhere in the FLOSS ecosystem, too.

    How about getting back to that Pascal-based GEBC substitute on the Web? Now, there’s a useful target for what little energy you have left.

    I’m absolutely sure you will never do it, though, so here’s a different proposition. How about documenting a moan on a Bugzilla board somewhere, possibly featuring a complaint about fonts (your main complaint against GEBC, I believe), or then again … I don’t wish to limit you.

    Pick something really important, and get back to us when somebody on said Bugzilla board listens to you.

    Give! Give! Give back!

  32. Deaf Spy says:

    Low price per core… Great, if I need to cache static content and data. Poor, when I need computational power.

    But hey, how can I expect you know anything about server-less computing. 🙂

  33. dougman says:

    “While some folks waste time and energy..”

    Spoken with authority, eh?

    “The dinosaurs didn’t notice the mice either but we know who inherited Earth”

    LOL…mammals did not conquer the reptiles. The standing theory was that a cataclysm decimated the large reptiles and made it inhospitable, allowing mammals to thrive. Today, there are some direct descendants of the large reptiles still alive today.

  34. dougman says:

    “The bandwidth through my ISP is too small for that. I have storage and bandwidth here.”

    What are you on dialup? Your cable modem offers more than enough bandwidth, you’re just letting that Pogson stubbornness affect your decision making process.

    Instead of wasting hours on a useless thing, you could be writing code and improving Linux.

  35. The Wiz wrote, “if you wish to have your ARM server sooner rather than later, purchasing time on packet.net might be a way to do it. Perhaps you might even be able to host mrpogson.com on, eh?”

    I could do that, but I still want local services. The bandwidth through my ISP is too small for that. I have storage and bandwidth here. I don’t need it there.

  36. Wizard Emeritus says:

    A Cloud vendor opens up a new service and has some takers. Of course you will note that this company still has the services of its intel server farm for sale. At this point its just another new service offered for some of its clients.

    It strikes me Robert Pogson that if you wish to have your ARM server sooner rather than later, purchasing time on packet.net might be a way to do it. Perhaps you might even be able to host mrpogson.com on, eh?

Leave a Reply