OpenPower Could Take IBM i To Hyperscale And Beyond
March 23, 2015 Timothy Prickett Morgan
Years ago, back when the IBM i platform was called the iSeries and IBM had just caught the Linux bug pretty bad, many of us had the idea that IBM should open up the OS/400 operating system and let it be driven more directly by a community of end users. The idea was to emulate the open source community that had fostered the maturity of the Linux kernel and the many thousands of other projects that make their way into a Linux distribution. As it turns out, Big Blue is starting out with opening up the hardware and from the looks of things at the OpenPower Summit last week in San Jose, alternative Power-based platforms are set to take off.
The platforms that were on display at the OpenPower Summit were largely based on the “Turismo” family of Power8 chips, which are the merchant variants of the processor that the company created in the wake of establishing the OpenPower Foundation in August 2013. While IBM has not open sourced the specifications and files that would allow someone to make a clone or derivative Power8 chip, IBM is licensing the chip specs to those who join the OpenPower Foundation and who pay fees to do so. Thus far, Suzhou PowerCore, a spinoff of an existing Chinese company that made variants of IBM’s low-end PowerPC processors for embedded devices, is the only company that has agreed to license the Power8 chip. Suzhou PowerCore, as we have reported in past issues of The Four Hundred, wants to eventually build its own version of the Power chip and sell it to various OEMs and ODMs who in turn will build servers and storage for datacenters based on its chip. Suzhou PowerCore’s first chip, called the CP1 and presumably short for China Power 1, was on display at the event, and according to Brad McCredie, who is president of the foundation and also an IBM Fellow and vice president Power Systems development at IBM, says that the chip is mostly the same as the Power8 chip that IBM has created except that it has tweaks in the security features of the processor that the Chinese government is keen on. (Precisely what these tweaks are is a state secret, but it does not take a genius to reckon that the Chinese government wants a means to secure that no outside government has backdoors into CP1 systems and possibly to make sure that the Chinese government does.)
The main thing is that Suzhou PowerCore is learning how to use IBM’s chip development tools and to coordinate with GlobalFoundries, which took over IBM Microelectronics last year, to get this tweaked Power8 chip etched and packaged for production use. In a pinch, if all else fails, there is now a second source for 64-bit Power8 processors, and this is important to IBM i customers even if IBM does not formally support IBM i on these chips. IBM i will run on them, and will likely also run AIX as well as various versions of Linux that are formally supported Linux.
There will also be multiple sources of machines available and multiple motherboard makers, who are focused on one-socket and two-socket machines right now because these machines are most commonly deployed in the hyperscale and supercomputing markets that IBM and its OpenPower partners are initially targeting with their machines. The good news is that the vast majority of IBM i shops are using machines with one or two sockets, so again, any of these OpenPower clones would provide more than enough oomph to run RPG, Java, PHP, COBOL, or Ruby applications and the DB2 for i relational database. I am not saying this will happen, because IBM i and AIX are dependent on the original IBM microcode and run on top of the PowerVM hypervisor (even if you don’t use logical partitioning, it is there), so porting IBM i and AIX to these other Power systems might take more than just a few driver changes.
In the long haul, I think IBM could outsource chip and hardware manufacturing entirely, as I have said before, and work with several other companies to create future Power processors. IBM could capture all of the revenue from the sale of future Power machines this way and avoid some of the costs. So a future Power Systems machine running IBM i might not only be made in China, but it could be designed in China with a Chinese variant of the Power processor running open source microcode that will allow companies in Europe and the United States a chance to check for any security backdoors themselves. This will be a different kind of Power Systems market to be sure, should it come to pass, but it will be a hell of a lot better than the tough rows that other proprietary platforms such as the Hewlett-Packard HP 3000 running MPE and Integrity machines running OpenVMS have had to hoe.
As we have discussed many times before in this publication, the Power8 chip has some advantages over Xeon processors from Intel, specifically when it comes to SMP scalability, number of threads per core, memory bandwidth, and low-latency connectivity and virtual memory sharing with peripherals through the Coherent Accelerator Processor Interface, or CAPI. These advantages are going to be leveraged at first by the supercomputing communities and perhaps by hyperscalers like Google and its peers who control their own software stacks and who, more importantly, the new open source microcode called OPAL was written for. (In fact, Google has done a lot of the work on this open source microcode as a means of participating in the OpenPower community.) The HPC and hyperscale communities, as well as some large enterprises that have very sophisticated IT and deep benches of system programming talent in their own right, And as always, they will be the early adopters of opened up Power technology. In all cases, the Power platform will have to demonstrate sustainable advantage over X86 iron, as it has over the past two decades on the performance front but which it will have to now do on the price/performance front.
IBM is committed to getting its merchant silicon prices for Power8 chips down below the price of Xeon chips with roughly equivalent raw compute performance, basically using its memory bandwidth, thread count, and CAPI I/O coupling advantages as a sweetener on the deal to entice new customers to the Power platform. And with upstart server makers in China and Taiwan leading the way to create low-cost platforms, there is a good chance that the OpenPower collective will be able to compete against makers of Xeon-based servers (which IBM is no longer one) at the system level. It is early days for OpenPower systems, but it is significant that IBM will be deploying the “Habanero” system from Tyan on its SoftLayer public cloud, not its own Power S812 or Power S824 machines. That also tells you that, at least as far as we know, the portions of the SoftLayer cloud that will be using Power server nodes from the OpenPower partners will not be support AIX and IBM i, although it is possible that IBM will roll out support for AIX and IBM i on such machines at a later date if the porting of AIX and IBM i to OPAL and the OpenPower machines is in fact happening. (I will try to get a firm answer from IBM about this.)
If you want to get the feeds and speeds of the various new OpenPower systems and see the conversation I had with Gordon MacKean who runs server and storage development and who is chairman of the OpenPower Foundation, you should check out a new publication called The Platform. My co-editor, Nicole Hemsoth, and I launched The Platform a few weeks ago, and we will be tracking the convergence of HPC, hyperscale, and enterprise technologies, among other things. You can see my overview of the various OpenPower machines at this link and my interview with MacKean at this other link.
What is clear from the talk and the nascent hardware is that the OpenPower partners are dead serious about making a business out of this and they plan to do so by making machines that have some advantages over Xeon iron. The movement is so focused on China now because the Middle Kingdom likes big iron and it likes to have intellectual property and security that it can control. As China ramps up its usage of Power-based systems and IBM rolls out a few generations of machines leading up to the Summit and Sierra supercomputers it is developing in conjunction with Nvidia, Mellanox Technologies, and possibly Tyan for delivery in 2017, the Power ecosystem will grow. It is already far safer to be on the Power platform than a few years back when the state and regional governments in China radically cut back on spending on Power-based systems from IBM at the same time the Unix market was going deeper into its malaise. IBM is not hot to trot for Linux on Power because this directly benefits most IBM i shops, but because an expanding customer base of Linux users on Power platforms makes the whole Power ecosystem healthier.
I have said it many times now, but anything that makes the Power ecosystem stronger lets IBM i last longer. The key thing now, I think, is for Big Blue to get IBM i onto OPAL and those cheaper OpenPower platforms. Anything that will get the 100,000 or so OS/400 and IBM i shops who are not on Power7 or higher hardware to move forward is a good thing.
And if IBM can’t sell customers machines directly to help them modernize, then maybe it is time to build a global SoftLayer cloud based on OpenPower iron and get those IBM i shops moved over. How many PowerKVM virtual machines running IBM i could it take to support these 100,000 vintage iSeries and System i customers? Probably no more than around 500,000 threads to 1 million threads. A single Power8 chip has 96 threads with all 12 cores fired up, so we are talking about on the order of 5,000 to 10,000 systems, tops. (If you use machines with fewer cores per CPU, maybe double it again, but I would not advise on that choice. I would go with a dual-chip module with two six-core Power8 chips in it.) Hyperscale companies eat that many machines for lunch. And even if each Power8 machine in this prospective cloud cost $10,000 as the foundation of such a hyperscale IBM i cloud, that is still only somewhere between $50 million and $100 million.
IBM could make it up on the sale of IBM i subscriptions on that cloud. Let’s say the average core on an IBM i platform in the field had a license price of $7,500, which is a mix of P05, P10, and P20 fees skewing toward the low end. Add in three years of Software Maintenance and you are up to close to $12,000 per core over three years. Divide it by the number of threads on average per machine (call it six or so) and divide by 36 months. That works out to around maybe $50 per thread per month, which works out to $25 million to $50 million per month. Or something on the order of $300 million to $600 million per year.
Yeah. That’s the math, people. And it turns the IBM i platform from a business where resellers and partners are pushing an upgrade every three or four years to one where everyone gets some dough every month. That would be a much healthier IBM i business, for sure. And it could be a material part of IBM’s cloud business, and therefore make IBM i as relevant to IBM’s future as the system is already to those 100,000 customers who are stuck on old iron. As for the 30,000 or so who keep more or less current already, this turn of event–hyperscale-inspired, low-cost iron and subscription-based IBM i licensing–could only lower their costs, too.