Why the Microsoft Gods may own the Mount Olympus of how data centers develop
- Summary:
- You don’t normally think Microsoft when it comes to server hardware design, but its Project Olympus development platform may hold the key to how servers, applications, services and data centers develop. It has already become a de facto standard for the processor chip makers to meet.
That last bit is perhaps the most interesting aspect because it looks like the server hardware market is about to hot up at least a bit, with new processors looking to muscle in on a market that, since the turn of the decade, has largely been an Intel Xeon monopoly. By the middle of this year, however, that is set to change, and by the end of the year there could be blood on the carpet from some hardware contenders. Microsoft, however, will be providing the tea and sympathy to the wounded.
For example AMD, which all but withdrew from the server market – well, the leading edge development end of it, holding on to a corner of it with a largely outperformed Opteron range of processors – is now about to come charging back with a new processor family currently code named Naples, which will come with 32 cores in a new processor designed specifically for large, scaling server applications.
Its all-but nemesis, Intel, is also set to launch a new design of Xeon processor, codenamed Skylake, around the middle of this year. This will have 28-cores, a number which no doubt doffs a small cap to the long legacy of applications that still run on the classic x86 processor architecture.
This obliges a trade-off in processor design between the number of cores and the clock speed of the chip. Better throughput performance comes from more cores and a slower clock speed within a given power envelope, but there still exist important single-threaded applications that perform better the higher the clock speed used. AMD pitching Naples at a higher number of cores would suggest it feels these applications may be reaching the end of their lifecycle.
There has also been a fair bit of talk of ARM-based processors making a pitch for a share of the server market, where the famed low power capabilities of the technology could win fans amongst heavy data center users. Here, energy consumption is a major component of operational running costs. Now Qualcomm, Cavium and others now have processors based on the ARM64 architecture, all aimed at the server market.
Finally, Nvidia, is pitching in with a hyperscale accelerator, known as HGX-1, aimed specifically at one of the hottest of hot ticket items on many user wish lists - high performance AI systems. This uses the company’s latest generation of CPUs, codenamed Pascal, together with its NVLink high speed multi-CPU interconnect technology. This can connect four HGX-1s together to provide interconnectivity for up to 32 CPUs.
Looking at these developments in isolation it might even be possible to surmise that the uniformity of the commodity technology server marketplace was in danger of becoming fractured, with different server architectures emerging for different applications types. That, however, is unlikely, for there is a common factor tying them all together, giving the prospect of a higher level of performance and service optimisation in an environment of continued uniformity of applications development.
Olympian heights
That common factor is Microsoft, or to be more specific, the growing use of Microsoft’s Project Olympus, its cloud hardware design platform for open source hardware development. This became the focus of attention at last week’s Open Compute Project (OCP) Summit, held in Santa Clara.
This demonstrated that Olympus is likely to be the next OCP hardware standard as server vendors gear up for mid-year upgrades and revamps of their products. And every processor maker used the event to show just how much they want to be compatible with it.
Olympus defines both server chassis and rack designs, as well providing a Universal Motherboard that leaves the processor to be used unspecified. So all the processor makers want to ensure that their new offerings make that grade.
And the really interesting bit about what was on show at OCP was not that it marks the start of an either/or battle for processor supremacy in the server marketplace. Instead it could mark the start of a new era of complementary server options that Cloud end users can opt for. And because of that, it can create a new range of service options that the Cloud Service Provider community can consider offering those end users can specify.
That, in its turn, means the opportunity for them to charge more for such services, knowing that they will be offering users better value by providing hardware resources that are intrinsically optimised for specific applications types. Yet at the same time they will be agnostic about the actual applications themselves.
For example, the new AMD Naples processors are being pitched at the heavy workload Cloud applications, such as big data analytics, where the ability to scale rapidly and smoothly is important. Arguably, this maps well onto the <marketing hype of the year> theme now emerging, that this is the year that Hybrid Cloud really takes off (whereas, of course, all Cloud has only ever been Hybrid Cloud all the time).
It is likely, however, that Naples will not be the best choice for users with a more generalised mix of old and new applications. Here, the Skylake Xeons from Intel may prove a better option as it seems they may provide the greater clock speed required by established single-threaded applications.
Big Cloud services users, the ones that soak up rack after rack are server resources, may well find themselves tempted by the low power capabilities of ARM-based systems. After all, unless something dramatic happens in the realm of nuclear fusion in the next few months, energy costs are only ever going to climb inexorably higher. Data centers already consume over 3% of the world’s total energy production, and the demand is expected to triple over the next decade.
What is also important here is that Microsoft has ported Windows Server to both Qualcomm’s and Cavium’s ARM-based processors. This is said to be for 'internal use only’ by Microsoft, but as it was talked about at the OCP Summit, it is only reasonable to expect much more from this in the near future.
And with AI now becoming `flavour of the decade’ for every corner of the IT industry, the Nvidia GPU approach to AI acceleration is likely to be, at the very least, an important add-in service capability for Service Providers to have available.
All of these can be complementary to each other, in any combination. But as has been common with much of IT over the years, what is complementary in one direction can often prove totally incompatible in other ways. Here, the use of a Microsoft hardware development platform may overcome this problem, for the most obvious tests to run are based on Microsoft applications, with Windows Server seeming to be a `standard’ evaluation tool. By implication, this also means other third party applications are likely to run.
There are also some possible by-products for Microsoft as well. For example, it is expected that its in-memory processing architecture, known as Project Hekaton, will be making its formal appearance in the next version of SQL Server, which is expected to be ready for introduction around mid-year. This could coincide well with the potential arrival of new server hardware running, say, AMD Naples processors.
My take
This year looks likely to hold a bunch of new server developments that advance the potential of big data analytics, hyper-convergence, in-memory processing, AI, and machine learning. It also holds out the chance for more optimised resource options to be available, while retaining software compatibility.
But it also seems like Microsoft may have hit on the sweet spot of commonality when it comes to the design and implementation of servers that overcome that potential for incompatibility. That could put it in a strong position when it comes to influencing how the Cloud Service Provision business develops over the next few years.