Doing The Two-Step To Get To Power8
May 12, 2014 Timothy Prickett Morgan
For most IBM i shops, the advent of the new IBM i 7.2 release and the Technology Refresh 8 for IBM i 7.1, both announced in the past several weeks, are more important than any shiny new Power8 iron, regardless of how much oomph it has. At any given time, only a relatively small percentage of customers are at a point where they absolutely have to upgrade their systems. For most shops, workloads grow more or less along with their revenues and even if they add new workloads, their existing Power6, Power6+, Power7, and Power7+ machines have enough latent capacity to tide them over for a year or two.
No one wants to hear this, least of all IBM and its reseller partners.
The days of needing to upgrade a machine every year or two, which was certainly the common practice during the AS/400 era in the late 1980s and through the mid-1990s, are long since gone. Like you, I miss the excitement of the yearly cadence and the wide variety of competing platforms in the midrange. This used to be a lot more interesting and, quite frankly, more profitable for all of us.
IBM’s Power processors have ridden the Moore’s Law curve to radically increase the throughput of an individual core and, when clock speeds hit a wall more or less a decade ago, the increasing core counts and the parallel nature of the database and middleware layers of OS/400, i5/OS, and IBM i have allowed for customers with modest workload growth to get by on progressively smaller systems–in terms of software tiers as well as physical form factors and pricing. The advent of virtualization has also allowed companies to consolidate machines and more efficiently make use of the processing capacity they buy over the peaks at end of week, month, and year. Everything runs more efficiently on Power Systems, and many machines have latent capacity in the form of spare cores that can be activated on a moment’s notice. And so the upgrade cycle for Power Systems shops, whether they are using IBM i or AIX, has stretched out to four, five, and sometimes six years. Small wonder, then, that the Power Systems business is showing revenue declines. It is a victim of its own success in some regard. (And, big state-sponsored corporations in China and several other emerging markets are looking at other platforms–notably Linux on X86 platforms–as an alternative to big iron from IBM, Hewlett-Packard, Oracle, and Fujitsu.)
The upgrade cycle for X86 systems is a bit different, and considerably shorter. For most enterprises, an X86 box will stay in use for around four years, and a quarter of the fleet–yes, I said fleet–is upgraded every four years. Their workloads tend to change fast, and large enterprises in particular tend to add new and large workloads at a fair quick pace. Some workloads, like risk management at banks or simulation and modeling for manufacturers or energy companies, are always hungry for more compute capacity. So not only are the machines swapped out every four years (after that, the increased hardware maintenance costs and the likelihood of component failure is costly), but the amount of computing tends to double every two years, again tracking, more or less, to Moore’s Law. These workloads are parallel in nature and can take advantage of both the increased core count inside of a single system, and thanks to tuning by software developers, and an increased number of nodes in a cluster. At the very high end of the market, companies who are licensing software on a per-socket or per-server basis are looking at a three-year cycle because they can consolidate software licenses onto successively fewer and fewer machines with each upgrade cycle. The reduction in software costs helps pay for hardware upgrades, and the fleet ends up being newer as well as having more performance.
That said, the vast majority of small and medium businesses whose applications run on Windows or Linux are more like IBM i shops, even if their platform is an X86 box instead of a Power machine. There just is not a lot of Linux in the SMB space for core applications, but that is changing ever so slowly as people become more comfortable with Linux and as Linux becomes the preferred deployment platform for Oracle and SAP applications. Oracle has its own Linux distribution, a clone of Red Hat Enterprise Linux, while SAP has anointed SUSE Linux as its sole platform for its HANA in-memory database. The point is, companies with modest X86 processing needs–something under a rack of iron, with maybe a dozen or two servers and affiliated storage–tend to stretch out their iron just like Power Systems shops do.
If this were not the case, then the world would be consuming 20 million servers a year instead of 10 million. Maybe 30 million, now that I think on it.
This bifurcation of the market into fast and slow upgraders is nothing new, but the ever-increasing performance in a socket has put a damper on revenue growth for most system vendors. The customer gets the benefit of more computing power, sometimes at the same price per unit of performance, but often at a lower price. When computing power on server chips (in the form of higher clock speeds) was hard to come by and workloads were not very parallel, server makers could charge a lot more for incremental clock speed improvements because customers were desperate for 15, 20, or 25 percent more oomph. You will find that Intel and IBM still do charge a wicked premium for their top-bin processor parts, which have all the cores working or which often have higher clock speeds to boost single-thread performance. The thermal envelope of a chip prevents chip makers from doing both at the same time, and boosting clock speeds 10 percent eats as much juice as boosting core counts by 25 percent.
The laws of physics and the laws of economics are thus conspiring against chip makers and their server partners, driving both revenues and profits from the supply chain. If it were not for cloud builders and hyperscale Web application operators (sometimes a company does both, as in the case of Google and Microsoft), after the virtualization downdraft and the efficiencies from cloud computing, the server business would be in far worse shape than it already is. And I happen to believe that the margin pressure is only just beginning. As SMBs shift to cloudy applications or run their own on utility-style clouds, the makers of machines will get yet a smaller portion of the revenue pie as minimalist servers are deployed and even less of the profits. Those vendors who own their own operating systems will continue to benefit, and as I have said before, IBM should long-since acquired its own Linux distribution. Canonical’s Ubuntu Server is a nice fit now that it supports the Power processor, but SUSE Linux Enterprise Server or Red Hat Enterprise Linux would do as well. RHEL is too expensive for IBM’s checkbook these days, and SUSE Linux seems to have less momentum (outside of SAP HANA and supercomputing) than Ubuntu Server.
Now would not be a good time to start a server business, and this is why IBM is trying to sell off its X86 server business to Lenovo Group and why HP has just partnered with contract manufacturing giant Foxconn to create a new line of HP-branded minimalist machines aimed at hyperscale companies. Perhaps this is what IBM should have done with its System x business as well, and if the $2.3 billion acquisition deal with Lenovo falls though, perhaps this is what Big Blue will end up doing with one of the contract manufacturers.
The important thing for IBM i shops to consider is that the Power8 machines are so different from their Power7 and Power7+ predecessors that Big Blue cannot do upgrades that preserve serial numbers–at least not on the entry machines we have seen. That means customers will not be able to do the normal MES upgrade, in the IBM lingo, and they will have to do a push-pull or what I have called a two-step upgrade, essentially buying a new machine and then finding a use for the old one or selling it back to IBM or a third party. The lack of serial number preservation also means that the value of the current machine will have to be fully depreciated once it leaves the premises, and that has implications for the accounting at many companies. This will be a factor in the upgrade process, for sure. IBM has promised that it will be able to preserve serial numbers for customers using its enterprise-class Power 770, Power 770+, Power 780, and Power 780+ servers, and therefore it will provide proper upgrades as we know and I guess love them.