Let’s Take Another Stab at Power7 Blade Bang for the Buck
May 10, 2010 Timothy Prickett Morgan
In last week’s issue of The Four Hundred, I created a slew of configured Power7-based blade systems, loading up i 7.1 and AIX 6.1 with two relational databases, and I said that the premium that IBM was charging was small enough to be a sign of progress. Surprising, even. What was even more surprising to me was finding out I made a big error configuring the Power System 701 blades and a small one setting up the AIX blades. And those errors make the i blades look downright awful compared to the AIX blades.
First, the big mistake. IBM is charging $14,995 per core for an i 6.1 or i 7.1 base operating system license on the PS701 blade, not the much lower $2,245 per core it is charging for the PS700 blade. Both are single-socket blades, but the PS700 has only four cores and only eight memory slots, while the PS701 has eight cores and 16 memory slots. The PS700 configurations I set up to compare to prior Power6-based JS12 and JS22 blades and to prior Power6+-based JS23 and JS43 blades were correct; as were comparisons that pitted the double-wide PS702 blade against these earlier machines in the story I did two weeks ago.
The second mistake, and IBM has to take at least some of the blame for this one, is the online system for the PS700, PS701, and PS702 blades puts i and AIX pricing up as single radio buttons on the Web, not explaining that the AIX licenses are per blade, not per core as on the i editions of the blades. Because IBM no longer seems inclined to publish a complete price list of its systems software, it did not strike me as unreasonable that AIX Express Edition would cost $1,200 per core and AIX Standard Edition would cost $2,000 per core. As it turns out, on this class of entry machine, AIX Express Edition costs $300 per core (but only scales for four cores), Standard Edition costs $500 per core (it can scale as far as you want), and Enterprise Edition costs $790 per core (and has all the bells and whistles).
Obviously, the $12,750 swing in per-core licensing for i 6.1 or i 7.1 makes a huge difference in the overall price of a configured PS701 blade. The difference in AIX pricing is noise in the data.
To help you sort out the price/performance of old versus new Power-based blades running i 7.1, I have recast the original table from the story two weeks ago comparing the three generations of Power blades running the i operating system. You can see the updated table here. I think I have it right this time (wink, wink), and my apologies for the errors. But it is fair to say that if IBM provided the pricing up front and in the announcement letters for the software, people would have an easier time getting this stuff right. And it might be a good idea to put the Quick Pricers the i side of the Power Systems house with all the pricing explained in plain American. When the AS/400 finally disappeared into the gaping maw of the RS/6000 business in late 2008, we never saw those again. And business partners and resellers have to go through their master resellers’ online configurators to get any idea what anything costs.
This is patently stupid, and more important, it ought to be illegal. And in IBM’s case, as a twice-convicted monopolist, it used to be illegal and should have remained so even after Big Blue wiggled out of the consent decree more than a decade ago. The only time a vendor doesn’t publish list prices is when they have something they are trying to hide. And people like me will always, always, always, always do everything in our power to figure out what things cost and reckon what they are worth. I won’t always do it perfectly, but I will not stop. And when I find an economic or technical injustice, I am sure as hell going to point it out.
Like now, for instance.
Why in heaven does it cost $14,995 per socket for an i 6.1 or i 7.1 license, plus $250 per user, for the operating system and database? These entry Power blade machines are supposed to be aimed at competing with rack and blade systems running Windows, Linux, and Unix systems, right? I can see a modest fee for the OS, like the $2,245 that IBM is charging on the PS700 blade, and I can even live with the $250 per user charge because, as my mistaken tables from the past two weeks show, this is about the right price per core and per user, when you add them together, to at least put the i 7.1 platform in the running.
Let’s do the i-to-i comparisons first before I get into the AIX comparisons. As you can see, the new PS701 servers running i offer nothing much in the way of a price/performance advantage over the JS23 blades they replace. The prices are almost the same, so on a per-user basis, you only save a bit money. What the PS701s do offer is considerably more oomph for slightly less money, so on a dollars per CPW basis, they are a better deal. But don’t get too excited. We’re talking somewhere in the range of 20 to 30 percent. (CPW being short for Commercial Processing Workload, the variant of the TPC-C online transaction processing test that IBM uses internally to gauge the relative performance of AS/400, iSeries, System i, and Power Systems.) That is about where it should be, in terms of relative gains, but the JS22, JS23, and JS43 blades were too expensive to entice customers for the past two years, so that Moore’s Law improvement–remember that X64 servers have doubled in power at their same price points in those same two years–didn’t allow the i platform to gain any ground on its rivals.
I’ll say this again: If you have modest CPW and memory needs and you don’t want to lay out a lot of cash, the two-year-old JS12 blade is tough to beat because IBM tossed in the i 6.1 license for free. While the new PS700 blade in a light or modest configuration can meet the JS12 blade, on a per-user basis for a similar number of users, the JS12 wins hands down, user for user. It looks to me like that $2,245 license cost for i 6.1 and i 7.1 was designed to do just that. People will look at the cost per CPW and see they are the same, but when they look at the cost per user and see the PS700 is nearly twice as expensive as the JS12, they will decide they can get by on less CPWs. Given this, I would think that IBM announcement will have the unfortunate effect of propping up JS12 values, despite the I/O issues these blades had. (They are not speed demons).
So why is IBM charging a lot more for the i 7.1 operating system and its integrated DB2 for i database? I don’t know, and IBM sure never explains it. But what I can tell you is that the pricing is seriously out of whack with the AIX-Oracle combination on the PS701 single-wide (and single socket) and PS702 double-wide (and two-socket) blades. As I explained many times in the past and in last week’s original i-AIX PS7XX blade comparison, there is a huge performance gap in the OLTP ratings between the i and AIX boxes as implied by their CPW on i and Relative Performance (rPerf) on AIX benchmark ratings. I can’t tell what is going on, but I suspect that AIX is doing things to push the performance envelope that the i folks are not–things that the Transaction Performance Council hasn’t outlawed but do not necessarily reflect real-world performance. (One thing it might be doing is de-randomizing the TPC-C transactions, sending data to particular CPUs so it is ready before it is even asked for.) I give you estimated OLTP transaction per minute (TPM) figures based on my best guess of how CPW and rPerf translate, but I think it is probably best to use per-user comparisons.
In that case, the AIX and Oracle combo just beats the tar out of i-DB2 on the same Power iron. IBM’s AIX charges are very, very low–lower than what it costs to get a Linux license in many cases. And Oracle’s Standard Edition One database, which is restricted to servers with one or two sockets, is very aggressively priced at $180 per user or $5,800 per core. Oracle understands that it is competing against Microsoft and to a certain extent, the MySQL database it now owns thanks to its acquisition of Sun Microsystems. And it prices its database accordingly. Oracle 11g SEO has all the same features of the Standard Edition and Enterprise Edition, it is just limited to two sockets. Standard Edition, which costs $350 per user or $17,500 per core, is limited to four cores, while 11g Enterprise Edition has no core scalability limits on a single SMP machine and costs $950 per user or $47,500 per core.
Oracle is the market leader for databases on Windows, Unix, and Linux, which is why I configured it on the PS7XX blades. In the revised i-AIX comparison table I built for this story, I did a few things besides fixing the two errors. One, I fixed the AIX configurations in terms of the AIX pricing and left 11g SEO on the PS7XX blades. Then I added a new set of AIX configurations that bumped the machines up to AIX Enterprise Edition and on the PS701 and PS702 machines plunked on Oracle 11g Standard Edition (even though it is not required) just to see how moving to the more expensive database changed the relative price/performance compared to the i-PS7XX blades. Lo and behold, that helps get things closer into whack. But it is wrong. You don’t need 11g Standard Edition on these machines, and it is stupid to put it on there. The competition is 11g Standard Edition One, as it has been for many years.
Other things to note: The PS700 running i 7.1 holds its own against the same blade running the AIX-Oracle 11g SEO pairing, with only a 10 percent premium over the AIX setup supporting 10 users. On setups supporting 20 and 30 users, the i premium rises a bit, and on heavier hardware configuration supporting 40 or 150 users, the premium gets up to 18 and 26 percent, respectively. This is too much, but within discounting range.
Forget about it with the PS702. By putting the correct per-core $14,995 license fee in my comparisons, even looking at it on a per-user basis, machines with 40 users running i 7.1 have an 87 percent premium over the AIX-Oracle setups, and with 150 users it gets a bit better with only a 70 percent premium. As you activate more cores and add more users, the disparity just grows and grows, with the i setups costing two to three times as much as the AIX-Oracle setups.
The heavy PS702 configurations have an even larger gap. Personally, I think that if you tried to get $384,531 to activate a 16-core PS702 blade, you’d be laughed at, and if you somehow managed to get it done and you got caught, you’d be fired. The i 7.1-DB2 combo on the PS702 blade is just obscenely more expensive.
So why would IBM do this? Two reasons.
First, $25,000. That seems to be the number of a base, single-core i configuration of any usefulness for the past five years or so, and I think it is a price that resellers need to stay in business and so IBM always shoots for it. It will cost you around $25,000 per core for all the PS7XXs. The thing is, IBM really only expects you to activate i on one or maybe two cores. There are not that many shops that need the 16 cores of a PS702 fully loaded. And at those prices, of course, there never will be.
Second, DB2 on AIX is not cheap and that means the i-DB2 duo can’t be, either. The DB2 Express Edition that is available for Windows, Linux (on X64 or Power servers), and Solaris (on X64 servers only) is not available on AIX on Power. Yeah, I know. Whaaaaa? DB2 Workgroup Edition has a governor limiting it to four sockets (with a maximum of 16 cores total) and 16 GB of memory; it costs $15,000 per socket or $442 per user. That price sound a little familiar? But DB2 Workgroup Edition users don’t have to pay $250 per user on top of that fee. So it is not exactly the same as the i 6.1 and i 7.1 licensing fees on the PS701 and PS702. (That DB2 price includes a year of maintenance, and the i license includes Software Maintenance.) The core DB2 Enterprise Server has a $1,040 per user fee or a bunch of different prices based on Processor Value Units, but does not seem to have a per-socket price. Various add-ons to DB2 EE are priced on a PVU basis.
Here’s the big difference: In the Unix market, IBM discounts hardware and software like crazy to beat Oracle. In the i space, where there is no alternative, the discounts are relatively skinny. At least until you show this data to IBM and explain that you want to pay the AIX-Oracle price for i 7.1-DB2 running on the PS7XX blades. There is no reason on Earth why you should have to pay one red cent more.