With the Mobile World Congress kicking off next week in Barcelona, the IT market will for a few days be focusing a little more than usual on the infrastructure that underpins the mobile services that we all increasingly depend on in our day to day lives. Dell is jumping the gun a bit to be heard above the cacophony.
This infrastructure is vast, and changing fast, and companies that have built large-scale compute and storage for hyperscalers like Google, Microsoft, Amazon, and Facebook in the United States or Alibaba, Baidu, Tencent, and China Mobile are hoping to cash in and peddle iron to the next tier down of telcos and service providers. And in a number of cases, the cloud builders are taking direct aim at some of the traditional hosting and fledgling cloud businesses aimed started by the telcos in recent years, and now the telcos are looking for something that they can use to compete back.
This trickle down of technology is precisely why Dell formed its Datacenter Scalable Solutions, or DSS, group last year, which The Next Platform profiled in detail. As we pointed out, hyperscalers buy machines in lots on the order of tens of thousands and tend to have multiple datacenters each with 100,000 or more servers in them. DSS is not aimed at them, and there is no rule of thumb that can so simply characterize them. But generally speaking, they are not buying at the node level, but rather at the rack level, and usually multiple racks per order and many racks per year per datacenter is a good guess. While the hyperscale market, which is served by the company’s bespoke system unit called Data Center Solutions as well as Hewlett Packard Enterprise and a number of other ODMs like Quanta Computer and suppliers like Supermicro, generates somewhere between $7 billion and $8 billion a year in revenues, Ashley Gorakhpurwalla, vice president and general manager of Server Solutions at Dell, estimated last summer that this next band down of 1,000 or so service providers would generate maybe $6.6 billion a year in sales, with traditional enterprises making up maybe $25 billion to $26 billion in sales.
Chasing that next tier down is important, and the telcos and service providers have seen the rise of Amazon Web Services, Microsoft Azure, and Google Compute Platform and they know they have to do something to keep their hosting customers. They also know they need efficient infrastructure on which to host their own applications, for the very same reason that their own customers are looking for hyperscale-class efficiencies.
Enter the DSS 9000 rack-level system.
Last fall, Dell put the DCS and DSS groups under a single management team, called Extreme Scale Infrastructure, and Stephen Rousset is a distinguished engineer and director of architecture for this combined unit. Rousset walked us through the feeds and speeds of the DSS 9000 system and talked about why it was something that telcos and service providers will now be hankering for.
“These customers are looking at the hyperscalers and they see that they have all of the cool toys,” says Rousset. “And they see all of the models and architectures that they are using to execute their IT infrastructure and this is where they want to go with Telco 2.0.”
The DSS 9000 is based on a set of rack-level machinery that was developed in parallel to the systems created by Facebook in 2010 and open sourced by the social network in April 2011 through the Open Compute Project. The G5 systems that Dell’s DCS unit has been selling to hyperscalers in competition with the Open Compute designs have some similarities, including using 21-inch racks that allow for full, half, and third width server and storage nodes that better accommodate 3.5-inch disk drives that are still common in hyperscale datacenters and will remain common in telco and service provider datacenters for years to come, too.
One of the big differences is between the G5 systems and the Open Rack alternative is that Dell has been selling G5 systems for several years and has worked out the manufacturing and customization kinks. Many Open Compute customers have to do this for themselves to fit their own environments. This is why the G5 system, with some modifications suitable for telcos and service providers, can make the hop from DCS to DSS, chasing more customers that do lower volumes. This indeed is the game for all system suppliers who aim at large enterprises, HPC centers, and cloud builders – they need to broaden their base.
The same mantras for hyperscalers are increasingly applying to telcos and service providers as they move from specialized appliances for every kind of network function to virtualized network software running in virtual machines or containers on generic Xeon-based server clusters. Not all of this gear can easily be moved over – no one has figured out how to turn a rack of servers into a core switch yet – but a lot of it can. The use of data analytics on a massive scale and the rollout of 5G networks will hit infrastructure hard. And, as we point out, these same telcos and service providers who want to dense iron for their networks want the same thing for the public clouds they build for customers and the private clouds they use to develop and host their own customer-facing applications.
The big thing that a system like the DSS 9000 delivers right off the bat is density. “These customers want power density and they despise stranded power. So they have 24 or 30 or 35 or even 50 kilowatts per rack,” Rousset explains. “Power density is huge, and we do our Modular Data Centers to allow this. You can have a 20-rack MDC that has a 1 megawatt footprint. Well, to do that, you have to have the density and accommodate the airflow to cool that. If you look at standard PowerEdge servers, you have maybe 20 or 40 of those in a rack and you are going to be stranding a lot of power. With the DSS 9000, we can get 80 nodes in a rack with half width sleds, and in some cases, 96 nodes with third width sleds.”
Similarly, rather than using a sophisticated and somewhat expensive iDRAC server controller on each machine to manage it individually, the DSS 9000 nodes employ a simpler baseboard management controller and the Redfish system management API stack to control the nodes on a rack and cluster level.
As for comparing the costs of the DSS 9000s versus standard PowerEdge servers, Rousset was not at liberty to say, but said that 80 percent of the components in a set of servers were the same – processors, memory, disks, flash, adapters, and so on – so focusing on the bill of materials misses the point. He added that Dell was “getting some pretty thin margins” on the sales to hyperscalers and presumably at lower volumes it can get better margins with the telcos and service providers.
(See? Now we understand the game of technology transfer from on high. Just like with supercomputing systems, so goes it with hyperscale iron.)
When pressed on pricing, Rousset says that cost is obviously a concern, but steady supply chain and the ability to meet demand for an integrated, rack-scale system that is ready to turn on and load up with software as it enters the datacenter is equally important.
“The infrastructure is going to be a portion of the savings, but where you have to add it all up is the integration of management and the time to live is quicker. You have to turn the conversation around and say, ‘How much is it worth to have your systems in production two weeks earlier than you do today?’ We don’t know their TCO models, so you have to pose the question back to them so they can quantify it in their own heads.”
The G5 rack-level systems are already sold using Intel’s current “Haswell” Xeon E5 v3 processors through DCS using what Rousset calls an “engineer to engineer” engagement, which is the kind of streamlined engagement you can have when the customer is an expert. While a number of hyperscalers used to buy G5 and predecessor rack-scale systems from Dell, we got the impression that Facebook and their friends are doing their Open Rack and Alibaba, Baidu, and Tencent were focusing on their own Scorpio rack-scale iron and not buying more G5 systems as they have done in the past. So maybe a move down to the telcos and service providers was inevitable.
With the DSS 9000s, a 4U chassis is the unit of configurability. The DSS 9000s will come equipped with the “Broadwell” Xeon E5 v4 processors that Intel is expected to launch sometime this spring if the rumor mill is correct. (We had heard February or March, but given that we are halfway through February, it is looking like March, which is a traditional launch time for Xeons anyway.) And they will also have a certain amount more hand holding, which stands to reason and will justify, to a certain extent, what we see as a higher margin that Dell will be able to command for the DSS 9000s versus the G5 systems that are their predecessors. For now it looks like Dell is only selling the full and half width sleds with the DSS 9000s.
Dell is putting its manufacturing teams, salesforce, and support organizations through the training necessary to ramp up the production of the rack-level systems in the second half of this year. The company has over 20 manufacturing sites in the world and another 16 configuration services centers, and can drop machine into 148 different countries. This is definitely required for chasing the telcos and service providers, who literally span the globe.