Thanks For The (Higher Priced) Memories?

October 29, 2012 Timothy Prickett Morgan

IBM likes to talk sometimes about the convergence of the System i product line (running what was then i5/OS) and the System p product line as a relatively recent development, and dates the official name change to Power Systems back in April 2008. But we have longer memories than that here at The Four Hundred and we know that the first converged server was actually a “Northstar” server that came out in 1997 and that the convergence plans date from before then.

I have nothing against convergence, even though I have been annoyed over the past decade and a half by the homogenization of IBM’s AIX and OS/400-IBM i platforms, as if the hardware was all that mattered. (And you know how I love the hardware.) I think that sharing hardware platforms and key systems components (like Web servers or TCP/IP networking stacks) across AIX and IBM i is not only smart, but necessary because IBM has to squeeze profits out of the combined Power Systems business. And, very likely more profits each year as the Unix and IBM markets themselves decline. This is a tall order.

But like I said, I have a long memory. And I spent the late 1990s and early 2000s railing against what I think were unfair pricing practices in the Power-based systems line, where IBM essentially gouged on pricing for loyal AS/400 customers so it could pay for cheaper component prices and then even steeper discounts off list price for newbie RS/6000 customers or established AIX shops that were more than happy to bring the Sun Microsystems or Hewlett-Packard coffee mugs or the Red Hat Enterprise Linux brochure to the bargaining table. I remember very clearly that in April 2008, the intent with the Power Systems convergence of what used to be distinct AS/400 and RS/6000 product families, with wickedly unfair prices for the same components if they happened to be running OS/400 or i5/OS instead of AIX, was to bring pricing parity to the machines. Processing, memory, disk, and other peripherals were supposed to cost exactly the same, no matter if the machine ran IBM i, AIX, or Linux.

Now, I see the old ways of price discrimination between AS/400 and RS/6000 platforms beginning to creep back into the product line, and like many of you, I don’t like it one bit. Or in this case, gigabyte.

As I discussed in the wake of the PowerLinux launch of Linux-only Power Systems machinery back in April, IBM is charging a pretty hefty premium for memory and disk capacity on a regular Power 730 server on which the PowerLinux 7R2 machine is based. The only difference between the two machines is how the processor card, processor core activations, memory, disk, and Red Hat Enterprise Linux software licenses are priced on the machines. Here’s an updated table comparing the costs, which has been updated to reflect changes in disk prices on the PowerLinux 7R2 machine that came out as part of the October 3 announcements.

As you can see, this price differential is a big deal for customers who put a lot of memory and disk into their machines. Processing capacity is a bit more expensive on the plain vanilla Power 730, disks are somewhere between 1.3 and 2.6 times as expensive (I am rounding), and main memory is from 4.6 to 6 times as expensive.

Now, in a virtual server world, memory is everything, often more important than cores or clock speed, so this is not a small factor in overall system configuration, performance, and price performance. But as I showed when I walked through the full system setups, if you add relational database software to these machines, the price gaps on configured machines close because of the crazy prices vendors charge for enterprise-class database code. (That’s a different problem, and I am not going to try to solve it right now.)

So that’s problem number one, and given how Intel is coming after the low-end of the Power Systems business with its Xeon E5-based servers and IBM is not yet ready to double-stuff Power7+ processors into Power Systems machines, I can see why back in April, when the Xeon E5s were coming to market, Big Blue felt like it had to do something dramatic. Of course, IBM could have done what I think is the smart thing and simply lowered memory and disk prices on all Power Systems iron to compete with X86 iron head on. There is an absolute glut of main memory on the market, but IBM’s pricing does not reflect this. Except on the PowerLinux machines.

That’s burr number one under my saddle. Burr number two is Active Memory Expansion, a memory compression feature that thus far has only been enabled with AIX.

Back when the Power7 processors were announced in February 2010, IBM added a compression algorithm to AIX that allows for data stored in main memory to be compressed with a 2:1 ratio, thus effectively doubling the main memory available to AIX, the PowerVM hypervisor, and applications running on them. This Active Memory Expansion feature ran in the AIX operating system and therefore took away CPU cycles that might have been used for other purposes, so it was not entirely free.

As I explained back in August ahead of the Power 770+ and Power 780+ server announcements on October 3, IBM has etched the 8-, 4-, and 2-byte parsing that the memory compression algorithm uses right onto the Power7+ chip, with two units on each processor and allowing for compression and decompression of data going into and out of the memory controller to be done without using any of the expensive CPU cycles. And as a bonus, the version of Active Memory Expansion can increase the effective capacity of memory by a factor or 2.25 without CPU hits compared to the factor of 2 with the old software-only version that had an unknown (to us at least) amount of CPU overhead.

This is great, and exactly the kind of technology that all chip makers should be adding to their systems. And it should be available to all operating systems that run on Power7+ platforms, not just AIX 7.1. In fact, I would say that it should have been added to IBM i 7.1 Technology Refresh 5, also announced earlier in the month, precisely as a carrot to try to encourage customers to get on new hardware and software. IBM should have gone even further and explained that this hardware-assisted Active Memory Expansion would be available on all future Power7+ systems on all operating systems.

But that is not the plan, as of now anyway, Steve Sibley, director of worldwide product management for IBM’s Power Systems division, explained to me ahead of the October announcements. IBM has no plans to expand Active Memory Expansion to IBM i, and that IBM’s intent was to be able to boost the number of logical partitions an AIX machine could host. Well, as someone who has been advocating strongly for proper IBM i clouds with proper live migration, I can tell you that the memory compression is another one of those features that is necessary.

The economics certainly argue for it. Take the Power 770+ machine. On these boxes, you will recall, you add memory cards with a certain amount of memory on them and then you use capacity on demand features to active capacity on the cards. A feature 5600 card costs $1,960 and has 32 GB of latent memory; feature 5601 costs $7,720 and has 64 GB of raw capacity; feature 5602 costs $15,440 and has 128 GB; and feature 5564 has 256 GB and costs a whopping $44,114. If you do the math, which I just did, that’s $246 per GB for the 32 GB card when fully activated, $306 per GB for the 64 GB and 128 GB cards fully activated, and $357 per GB for the 256 GB card with all the chips turned on.

As you can see, that is crazy expensive memory compared to the price Power 730 customers pay, which ranges from $133 to $200 per GB for capacities that range from 8 GB to 32 GB.

Now, if you could turn on Active Memory Expansion for IBM i workloads, you could cut that price to somewhere between $109 to $159 per GB. While that is not the super-low $29 per GB that IBM is only charging for capacity on the PowerLinux machines, it is a heck of a lot closer than the price of raw memory without Active Memory Expansion. And this represents a new kind of IBM i tax. Call it the IBM i memory tax. And call your IBM rep and tell them to get to work on getting Active Memory Expansion fired up on IBM i. If we hadn’t been leaning on Big Blue, there would still be no live migration for IBM i partitions, which should have been available years ago. Likewise, in this one big happy Power Systems family, Active Memory Expansion should have been available for IBM i from day one, nearly three years ago now.

There is simply no excuse in a world where in-memory processing is becoming more and more of a factor and you have a perfect machine, with single-level storage and therefore automated tiering for data between memory and disk, for Active Memory Expansion to not be available on IBM i. I would go so far as to say that it should have been on IBM i first, in fact. But that would be playing favorites, and we’re not supposed to do that anymore, remember?