Advanced Computing in the Age of AI | Thursday, March 28, 2024

Canonical To Kick Ubuntu Server Into Hyperscale 

Canonical, the maker of the Ubuntu distribution of Linux, does not have the same kind of footprint in enterprise datacenters that rivals Red Hat and SUSE Linux do. But the company is trying to do an end run around them, and has set its sights on the future of hyperscale computing in the enterprise.

Red Hat is the undisputed leader in commercial Linux business, with $1.33 billion in sales in its fiscal 2013 ended last February and on track to break $1.5 billion this fiscal year. Red Hat has grown so large and become the largest company providing support for open source software by focusing on the big middleware and database jobs on the back-end systems at major enterprises. In many cases, these are jobs that used to run a variant of Unix on RISC systems. Red Hat, of course, also is the platform of choice for whole new classes of jobs that went first to X86 iron and Linux and never touched a proprietary or Unix system.

SUSE Linux, which is privately held after being sold to Attachmate a few years back, has done best among supercomputing customers and as the default platform for SAP's HANA in-memory clusters and also has a much smaller slice of the enterprise market than Red Hat.

Canonical has its enthusiasts for Ubuntu on the desktop as well as aspirations on smartphones and tablets and a toe-hold in the enterprise. Hyperscale web application operators, media companies, telcos, and service providers sometimes opting for Ubuntu Server over Red Hat Enterprise Linux or SUSE Linux Enterprise Server.

Christian Reis, vice president of hyperscale computing at the commercial Linux distributor, joined Canonical last year to chase the opportunity for Ubuntu Server among those building clusters to support parallel applications on bare metal or create clouds to run virtualized applications side-by-side on nodes in clusters. Canonical was an early and enthusiastic supporter of the Eucalyptus cloud controller and the KVM hypervisor, and got a fair amount of traction with concept clouds a number of years ago. KVM's star has continued to rise, but the big commercial Linux suppliers have by and large hitched their wagons to the OpenStack cloud controller. Canonical is at the very bleeding edge with the OpenStack and KVM combination and is also very aggressively supporting current 32-bit and future 64-bit ARM processors as well as continuing to support Intel Xeon and AMD Opteron chips for servers.

"Every telecommunications company or service provider out there that is thinking about a public cloud has been approached by us," Reis tells EnterpriseTech. "We have worked with lots of them on hundred-node proof of concepts that will move into thousands of nodes in production. I think that OpenStack, like hyperscale, is still in its infancy. Who has announced that they have launched a large public cloud running OpenStack? I don't think you will find many. The fact is that today we are submarine in all of the large telcos and financial services companies that are experimenting with how they move to the next platform that is in a similar vein to the large web properties. The early adopters are large enough for the scale to matter to us."

To that end, Canonical's Landscape management system for Ubutnu Server (and its sibling version for clients) can scale to tens of thousands of server nodes and, significantly, already knows how to manage both X86 and ARM machinery. The Landscape update from last September allowed the management tool to provision and mange Ubuntu Server instances on Calxeda ECX-1000 processors (now made more or less irrelevant by Calxeda's closing its doors in December) as well as Marvell's Armada XP chips as well as the Xilinx Zync-7020, which mashes up two ARM cores and an FPGA on the same die. (This is the processor that Adapteva is using in conjunction with its massively multicore Epiphany RISC coprocessor.)

Reis was formerly vice president of engineering at Linaro, the group that is working to get the various Linuxes ported and running well on various ARM processors. Like many in the open source community, he is enthusiastic about the possibilities of low-powered servers in general for running hyperscale workloads and about ARM variants in particular. But like the companies who will eventually deploy hyperscale systems based on, Reis is practical about the fact that companies have developed a lot of code for 64-bit X86 systems and that low-powered Atom, Xeon, and even Opteron chips will also be a part of the hyperscale landscape in enterprises.

In April, the next big release of Ubuntu Server will be launched by Canonical. This one, nick-named "Trusty Tahr" after the sure-footed shaggy Himalayan mountain goat, is a Long Term Support variant. What that means is that Ubuntu Server 14.04 will be supported for customers for five years. This is the release that enterprises wait for and deploy widely, unless they absolutely need to be on the cutting edge, in which case they deploy the regular Ubuntu releases that come out every six months.

With the support of the Linux 3.13 kernel, the next release of Ubuntu Server will support Applied Micro's eight-core X-Gene 1 and sixteen-core X-Gene 2 processors as well as the "Thunder" 64-bit chips from Cavium. Reis said that Canonical was still trying to decide about continuing support for the current Calxeda ECX-1000 chips and the future ECX-2000s that Calxeda said it would continue to supply and support for existing customers. Reis is unable to comment on whether Ubuntu Server 14.04 will support AMD's eight-core "Seattle" Opteron A1100 ARM chip, which will ship later this year, and he was similarly unable to confirm that the operating system would run atop the KeyStone-II hybrid ARM-DSP chip from Texas Instruments. The X-Gene-1, Thunder, and KeyStone-II chips are all going to be available on server cartridges for the in the Moonshot 1500 hyperscale chassis from Hewlett-Packard. The Calxeda chips were already to be dropped in, too. HP is already shipping cartridges based on AMD's "Kyoto" Opteron X2150 CPU-GPU and Intel's two-core "Centerton" S1260 and eight-core "Avoton" C2000 Atom processors. All of these chips have 64-bit processing and memory addressing except the KeyStone-II, which used the 40-bit Cortex-A15 design from ARM holdings. Incidentally, Ubuntu Server was the default server operating system for the Moonshot machines running the Atom cartridges. For now, Moonshot is the big hyperscale hardware play for Canonical, but other platforms will follow.

While core operating system support is a foregone conclusion with any server platform with lots of options, as Moonshot certainly has, the application development toolchain has to be brought up to pace for ARM machines. On the Java front, Ubuntu Server 14.04 will ship with an OpenJDK version of the Java Virtual Machine, which Reis says will provide acceptable performance. But, he says, many enterprises will be looking for the full commercial port of its Oracle port of HotSpot JVM, which is due in 2015. Companies will deploy OpenJDK to get started and then probably move to HotSpot as soon as it is running well on ARM chips. The rest of the toolchain, especially the GNU compiler set, will also take some time to be tuned up. (This can't really be done now because 64-bit ARM server chips are not themselves fully cooked.) Reis says that tweaks to the development tools alone should provide somewhere between 10 and 15 percent performance improvements each year for several years after ARM server chips are out.

Canonical also embeds all of the key new technologies like OpenStack, Memcached, Hadoop, MongoDB, and Ceph into its Ubuntu Server release and offers support for them and deployment through its Juju templating system and Landscape management tool. And for those who want separate support for Hadoop, Ceph, or MongoDB, then Canonical has arrangements with MapR Technologies, Inktank, and MongoDB to go that extra support mile. (This is not an exhaustive list of the software stack, but just a taste.) Canonical is a big contributor to the OpenStack project and does its own support, and there is no appropriate commercial entity to support Memcached – at least not yet, and it would probably be Facebook if it had the inkling to do so. All of this software, like Ubuntu Server itself, will run on ARM or X86 server chips.

"Canonical's contribution is to make Google-style management of a large cluster available to people who do not have Google's deep pockets. It is to make the technology that you would need if you were managing hundreds to thousands to hundreds of thousands of nodes. These people understand that in ten years, you are not going to think about an individual server when you deploy workloads. They are going to be thinking about complex systems from the beginning."

Incidentally, in case you are wondering, Canonical's two largest customers that it can talk about publicly are Wikipedia, the online encyclopedia, and Baidu, the Chinese search engine giant.

Wikipedia has thousands of server running Ubuntu Server to host the encyclopedia, and these are presumably normal X86 workhorse machines. (Details were not available.)

baidu-cloud-server

Baidu is interesting in that it has created a storage platform for its Baidu Cloud personal storage service, and that it built this storage cluster using Marvell's four-core Armada XP MV78460 ARM server chip. Baidu was a little vague about the configuration of the storage array behind the Baidu Cloud service, but it has now scaled up to multiple petabytes, EnterpriseTech hears. The system behind this storage cloud is a 2U chassis that fits six server sleds, each with four 3.5-inch, 4 TB SATA drives in it. The chassis held 96 TB of data when it was announced this time last year. A rack of these machines would hold 2 PB of capacity. The storage nodes all run Ubuntu Server and a home-grown cold storage management system that Baidu calls Pan. Baidu said this time last year that the ARM-based storage servers were 25 percent cheaper than prior X86-based machines it had been using for its cloud storage. That comparison may or may not hold up today since processor technology has moved on.

EnterpriseAI