What Will IBM i Do With A Power10 Processor?
August 17, 2015 Timothy Prickett Morgan
The gap between what is needed and what is provided has been growing and growing for years, and we don’t talk about that very much in the IBM i base. As someone who loves big iron with lots of horsepower, I tend to look on each new Power chip rollout as a chance for the IBM i platform to capture more and more work. While this happens at a number of shops that mix IBM i and Linux workloads–and sometimes AIX, too–for the most part, the IBM i platform is a silo in the datacenter even if other applications tickle its databases over the Ethernet and the Internet.
The fact remains that customers running IBM i tend to be fairly modest in terms of the amount of compute they consume. This really started with the jump from proprietary CISC to PowerPC AS RISC chips back in 1995, when the amount of raw compute capacity tripled in one fell swoop. This was a boon to customers, who could buy lots of capacity, but their workloads grew more modestly than Moore’s Law and that meant they didn’t need to buy as frequently. Add in the efficiency gains from logical partitioning, which allows workloads that would have been on separate machines to be multiplexed across time on a single box, and it is no wonder that the Power Systems-IBM i combination generates far less revenue than its heydays back in the late 1990s. So do mainframes, Unix systems, and other proprietary machines–where they still exist.
The X86 server market grows some by virtue of competition–it has knocked out plenty of Unix and proprietary gear–but mostly it is growing because Windows on X86 is the platform for many small and midrange businesses and Linux on X86 is the platform for modern distributed computing, whether it is webscale infrastructure at Google or Facebook, risk management systems at a financial services firm, or simulation systems at a big manufacturer. The public cloud is also being built largely on X86 architectures, too. But the growth curves that the X86 platform has enjoyed in the datacenter won’t last forever, and that is why Intel has been keen on expanding into adjacent storage, networking, and telecommunications markets with its chippery. While there is insatiable demand for compute and storage capacity, that demand is tempered by budgets and by other factors. As I have said before, at some point, the workloads can only grow so fast as the population, once everyone has a smartphone and the Internet. You can only check your bank balance so many times a week, and there are only so many of us and we only generate so much data exhaust. We are not at that point, but we will get there. Perhaps even in my lifetime.
It is with all of this in mind that I contemplate the future Power chip roadmap, another sample of which recently came to light:
I am comforted, as you no doubt are, by the fact that IBM actually is showing customers and partners that it has a Power processor roadmap that extends out to 2020 and beyond. We need that kind of reassurance, and it is a wonder that it is not done more publicly considering that Big Blue is very eager to reinvigorate the Power platform and have it take a bigger bite out of the datacenter. This roadmap doesn’t say much, and if you want my detailed take on what this means technologically, take a look at the bit I wrote for The Platform. The roadmaps that IBM has put out, sometimes in conjunction with its OpenPower Foundation partners, are heavily focused on high performance computing, but the technology will obviously be adapted for more general purpose systems, too.
The question I have is if the Power roadmap will bear any reasonable relationship to the current and future IBM i landscape. If the future looks anything like the past, the kind of computing power and features that IBM will no doubt bring to bear may not map well to the requirements of the IBM i installed base. Let me show you a case in point.
Doug Fulmer, who is systems architect at KS2 Technologies in Grapevine, Texas, last week helped us understand what is going on in terms what machines customers are actually buying in the IBM i installed base. Like all of the dealers and resellers I know, the vast majority of the business that KS2 does is for four-core Power machines with only one core activated. This is in a server architecture that has 12-core Power8 processors that can scale all the way up to 192 cores in a single system image with 16 processor sockets addressing something crazy like 16 TB of main memory. That’s big iron, and ironically, it all fits into a portion of a single rack these days because, believe it or not, IBM is still a master at systems design and engineering.
According to Fulmer, who watches data like a hawk as those of us at The Four Hundred do, the average OS/400 customer back in 2006 was using somewhere between 285 CPWs and 385 CPWs of capacity based on IBM’s Commercial Performance Workload relative benchmark test. These days, on Power7+ and Power8 iron, which has only four cores and only one of them fired up running the IBM i operating system, the average customer has somewhere between 1,000 CPWs and 1,500 CPWs of capacity supporting IBM i–and this is on a core that can deliver somewhere around 9,900 CPWs of compute. So, the average machine for the SMB shops that comprise the vast majority of the IBM i installed base is running at somewhere between 10 percent and 15 percent of peak CPU.
So let’s just play a game here. Let’s assume IBM stays on the present course and tries to do the Moore’s Law game like Intel does. Some of the process shrink from 22 nanometers to 14 nanometers will go to cranking the clock, but most will go to cores. So maybe Power9 comes in at 16 cores or 18 cores. I think 16 cores with lots of eDRAM cache and a bunch of NVLink and CAPI ports is likely. The shrink to 10 nanometers might not get us that far, maybe 20 cores. In each case, we get microarchitecture improvements and better performance per clock enhancements, too. But on the order of maybe 5 to 10 percent unless IBM does something radical. (Which it might.) So we end up with a Power9 core that maybe has 20 percent more oomph, core to core, with the Power8, and the Power10 has maybe 20 percent more than the Power9. For the same of round numbers, that is something like 15,000 CPWs per core by 2020, and a two-socket server will have 40 cores for a total of 600,000 raw CPWs of performance, with NUMA overhead maybe shaving that down to 575,000 CPWs.
Now, let’s assume the workloads progress the same at IBM i shops. The average SMB shop running IBM i was burning 325 CPWs in 2006 and that rose to an average of 1,250 CPWs or so with modern machines over the nine years. That’s 18.5 percent growth in CPW use per year, on average, which ain’t bad mind you. The server business is not growing that fast. The trouble is, IBM i workloads are fairly modest even if they are economically vital. So if you project the same growth rate out from 2015 to 2020, an SMB shop will need something like 3,500 CPWs of capacity–in total–to run their workloads. That will be when IBM is delivering something like 15,000 CPWs per core. The gap will close a little, perhaps, but that is still only 23 percent utilization or so on that single CPU. IBM i SMB shops wouldn’t need a single Power10 core–forget a whole chip with maybe 20 cores–until 2028! And they would still have a few thousand CPWs left over for headroom. Let’s roll the clock back a little bit. A Power5 chip from 2004 with some of its cache crimped and running at 1.65 GHz could deliver 3,300 CPWs–twice as much as the average customer seems to need today. That was a P20-class machine, too, not a P05-class box like the Power S814 with only four cores in it.
Which brings me to my next point. When the Power8 entry machines were announced last year, this four-core machine was not part of the original announcements, and IBM was very clear to business partners that the company was not even sure it wanted to do a four-core variant. This machine, should it have come to pass, was expected in September, but was then was rushed out in June after partners and we presume customers made a big fuss. That IBM even thought it could launch only a P10-class machine loaded up with lots more cores than IBM i customers needed was a little disconcerting. A lot of the customer base is in the P05 tier, and truth be told, they could probably use a P02.5 tier based on how little compute they need.
IBM either needs to find more work for IBM i to do–and convince customers to do it, and do it fast–or make a better machine that fits the needs of IBM i customers. I will share some thoughts about what this machine might look like in the next episode of The Four Hundred. The answer I came up with will make you laugh, I think, and perhaps mesh well with what I think IBM needs to do for other customer sets as well.