At Long Last, IBM i Finally Gets Power9
February 14, 2018 Timothy Prickett Morgan
The most important part of the Power9 rollout for the vast majority of IBM i customers has finally happened: Big Blue has lifted the veil on the Power9-based “ZZ” systems that we have been hearing about for a few months now and that are being delivered in the Power S914, Power S922, and Power S924 servers.
These are three of the six machines based on the ZZ design, and they are the only ones that can IBM i without any restrictions. Two other machines, which are called the Power H922 and Power H924, were created explicitly to run SAP’s HANA in-memory database atop Linux, but also have the ability to support AIX or IBM i on up to 25 percent of their aggregate capacity. The sixth machine based on the ZZ motherboard is called the Power L922, and as the name suggests, this is a Linux-only variant of the base Power S922 platform, but one running PowerVM and not the OpenKVM hypervisor or the OPAL microcode that was created by IBM and Google under the auspices of the OpenPower Foundation.
Here are the basic feeds and speeds of the six new machines in the Power9 Scale Out family:
This Scale Out family picture above is not complete, of course, because it does not include the “Newell” Power AC922 that IBM already announced in December as its flagship hybrid CPU-GPU machine for HPC and AI workloads. Nor does it include other scale out models, such as the forthcoming machines based on the “Boston” Linux-only systems that we know are still coming in the second quarter for a variety of workloads, including open source databases, hyperconverged storage, Hadoop and Spark, and key/value stores like Redis and Memcached. The Boston systems, which are distinct from the Newell and ZZ machines in a number of different ways, including the processors and I/O. (We will be trying to find out more about that.) The important thing as far as IBM i shops are concerned is that the boxes that the largest number of shops can deploy as their main database and application engine are being revealed. Those IBM i shops with Power7, Power7+, and Power8 iron with four or more sockets are going to have to wait until the third quarter for the “Zeppelin” and “Fleetwood” Power9 NUMA machines to be launched.
Then, the full Power9 family picture will be available, until a Power9+ generation comes out perhaps sometime 18 months to 24 months from now. Hopefully it will not take four years to see the Power10 chip emerge from the GlobalFoundries fab in Malta, New York. (Which I visited on a rare tour last week and which I will talk about in a future issue of The Four Hundred.) But I think it will take at least three years for Power10 to come out. Between now and then, we could see a Power9 variant emerge from Suzhou PowerCore or maybe another Chinese chip maker, but it is highly unlikely that IBM would allow for IBM i or AIX to be licensed for them – unless compelled by courts to do so. And that seems extremely unlikely to the point of being impossible.
I know, I know. The Power9 machines are not even shipping and I am thinking about Power9+ and Power10. That is my job, after all.
Let’s talk a little bit about the ZZ platforms and how they compare to their Power8 predecessors. (There were no Power8+ machines that ran IBM i or AIX, and there was only one that ran Linux, and that was the “Minksy” Power S822LC for HPC, which was really a proof of concept for the Power AC922 based on the Power9 and the Volta GPU accelerators.
Form Factors: The most obvious thing is that the form factors are essentially the same. There are variants with one or two processor sockets and with a 1U or 2U rack-mounted chassis. The Power S924 comes in a tower configuration, which is important for small IBM i shops.
Compute: There are two different flavors of Power9 chip, and each can come with two different architectural options. The “Nimbus” Power9 chip is aimed at machines with one or two sockets and all of the available I/O on the chip is allocated to PCI-Express 3.0, PCI-Express 4.0, CAPI, OpenCAPI, or NVLink buses. With the “Cumulus” Power9 variant, which is aimed at machines with four or more sockets, some of that I/O is used to provide the NUMA links that couple together the processors and share their memory across the processor complex. (IBM persists in calling this SMP, short for symmetric multiprocessing, but that is wrong.)
For each type, IBM has a skinny core with four threads per core – what it calls SMT4, short for simultaneous multithreading – and a fat core with eight threads per core – what it calls SMT8. All of the Power7, Power7, Power8, and Power8+ processors were SMT8 by default, and they could dial the threading model up and down on the fly. With the Power9 chip, as we detailed way back in August 2016, IBM decided that the Linux community with the OPAL/OpenKVM stack wanted a lower amount of SMT with more cores, so these variants of the Nimbus and Cumulus chips will have a maximum of 24 cores per die, while the PowerVM crowd wanted fatter cores with more threading and fewer of them to keep software costs (which is priced per core in a lot of cases) lower than they might otherwise be. The Cumulus chips support buffered memory, just like on the Power7 through Power8+ machines, while the Nimbus chips can use standard unbuffered memory, like most of the X86 industry excepting the high-end Xeon E7 processors that are basically end of life.
In the ZZ systems, IBM is offering Power9 chips with 4, 8, 10, or 12 cores, and speeds range from a low of 2.9 GHz on one of the 10 core chips to a high of 4 GHz on one of the eight core chips. There is a wide mix of processor options that come in 190 watt, 225 watt, or 300 watt power envelopes for the Power9 chips. It is not obvious yet how this all maps out from the materials IBM has delivered, but we will be figuring it out. The other twist is that IBM is going to be giving maximum clock speed ranges for each chip, and says it is doing this because it has automatic frequency turbo boosting for the Power9 chip that tries to crank the clocks as far as they can go for each workload, as they run, and dynamically adjusts to get all the work done within a power budget for the chip itself or, if needed, in conjunction with the surrounding environmentals.
What we can tell you is this. Steve Sibley, vice president and offering manager for the Cognitive Systems division and basically the go-to person for Power hardware, and Brad McCredie, an IBM fellow who has headed up Power Systems development and the OpenPower Foundation, told me that, as we had heard, the IBM i Commercial Performance Workload (CPW) and AIX Relative Performance (rPerf) benchmark test results across all of the different variations of the ZZ systems would not be done until February 27. But, generally speaking, based on early test results, customers should expect somewhere between a 30 percent to 50 percent boost in performance moving from a Power8 system to a like-for-like Power9 system.
Memory. The ZZ systems have eight memory controllers per socket, which are located on the Power9 chip, and up to two channels of DDR4 memory can be hung off of each controller. IBM is offering memory in 16 GB, 32 GB, 64 GB, and 128 GB capacities and running at 2.13 GHz, 2.4 GHz, and 2.67 GHz, depending on the performance needs and the thermal envelope in the box. If you do the math, that works out to 2 TB per socket using 128 GB memory sticks. That matches what AMD can do with its “Naples” Epyc server processors and bests Intel’s “Skylake” Xeon SP chips, which top out at 1.5 TB per socket. IBM offers a nearly 2X benefit on memory bandwidth compared to the Xeon SPs, which is one reason it could ditch the buffered memory (which is lot more expensive) and still compete.
I/O And Storage. The new Power9 machines have varying amounts of 2.5-inch disk and flash drives, and some of the flash drives are capable of supporting the NVM-Express direct connections over PCI-Express ports, which bypass chunks of the operating system kernel and I/O stack to talk more directly to the compute complex, radically improving throughput and reducing latency on flash devices. These NVM-Express flash drives cost a pretty penny, but when it comes to performance, they are worth it and all server makers are embracing them. The only trouble is that flash prices have skyrocketed like main memory prices in the past year and a half, and the memory makers (particularly Micron/Intel and Samsung) are in no hurry to ramp up capacity and collapse the prices. They are getting rich right now.
The ZZ systems have the usual gang of I/O expansion drawers and disk and flash expansion drawers that hang off of the main system chassis. As in the past, the low-end machine – in this case the Power S914 – has a variant with four cores, only 64 GB of memory, and no I/O expansion that is aimed at very small, P05-class IBM i shops. We will see how this stacks up to the Power S812 Mini, a single-core machine IBM announced last Valentine’s Day for IBM i shops, and the Power S814 in a future edition of this newsletter. IBM says that it plans to keep selling the Power S812 Mini through 2019, so no worries if you are looking for one.
Pricing. When the Power Systems business gets in a pinch, IBM gives the market it is chasing lower prices on components than the parts of the market it has captive. In the early days of the AS/400 and RS/6000 convergence, IBM charged a lot more for AS/400 features than for RS/6000 features, and after a lot of crabbing, it converged the features and their prices. And then when it wanted to go after Linux workloads, IBM i and AIX had higher prices and the same features on Linux systems had lower prices.
With the Power9 ZZ systems, which span all three operating systems with an underpinning of PowerVM, IBM is setting memory, disk, flash, and I/O feature prices back to parity across the lines. And that does not mean it is charging Linux customers more, but rather IBM i and AIX customers less. How much remains to be seen. IBM is also fully activating the cores on all Power9 processor feature cards and it is similarly activating all memory that is put into the system. No more core activations, no more 1 GB memory chunk activations. This, says Sibley, is a lot simpler. Correct. It probably also means more initial cash outlay for customers because there is no latent capacity that is not activated and therefore not charged for. Hopefully the net price is lower; we will figure that out in due time.
Bang For The Buck. Speaking very generally, Sibley says that the Power9 ZZ machines running AIX or Linux should offer somewhere between 20 percent to 30 percent better bang for the buck than their Power8 predecessors, including the cost of a configured hardware system and the operating system and also including the fact that PowerVM is now bundled into the deal for free on the Power9 machines. For IBM i, that price/performance will be a little bit lower because IBM i includes an embedded relational database management system, and this cost is often larger than the cost of the underlying hardware.
We will be gathering up the slots and watts and feeds and speeds and prices and dices and doing all the math. So stay tuned.