The Power Systems 570 i Edition Versus Big Windows Boxes
October 6, 2008 Timothy Prickett Morgan
Now that Intel has its “Dunnington” six-core Xeon processor into the field, it would seem to be a good time to see how the Power Systems 570 i Edition servers stack up against Windows machines equipped with enterprise-class hardware features and systems software. It is a good time, with the fourth quarter under way and IT shops looking to buy–or not, as the case may be. Unless, of course, you were looking for some good news on the i pricing front.
Aye yi yi yi yi, Power Systems i.
As I pointed out in last week’s story comparing the System i 570s based on Power5+ processors with the Power Systems 570 i Edition machines based on the Power6 chips, customers will see, in like for like comparisons, some improvement in bang for the buck. Over the course of the two years between these two generations of machines, the improvement was about 35 percent, on average, for servers equipped with 5250 capacity on their processors. (See the story for details. I made a slew of configurations and comparisons, and the prices and value for dollar bounces around quite a bit.) I said last week that this is the kind of price/performance improvements that server platforms have been delivering per year for a long time. What I didn’t say–and what is certainly true as you will find out–is that Windows systems software is getting wickedly cheaper at a rate that far exceeds that of the ever-cheapening of server hardware.
What this means, in plain English, is that high-end Windows servers kick the tar out of the 570-class i boxes, even the new Power 570s.
The gap is quite impressive. As you know from reading other stories in this series, the Power 520 is absolutely competitive with Windows alternatives. User for user, transaction for transaction, a Windows box based on X64 processors is no better and no worse than a Power 520 once you add in the operating system, the relational database, a virtualization hypervisor, and a suitable configuration of memory and base disk and tape. Moving up into the midrange with the Power 550, the gap between Windows and the Power Systems i platform starts opening up, with the machine costing anywhere from three to six times as much as a Windows box. Incidentally, the JS12 and JS22 blade servers running i 6.1 are a lot closer in price to X64 blade servers, and this is a good thing. But the Power 550 gap is a caution.
With the Power 570s, the gap is ridiculous, and it is so for a couple of reasons. First, Intel and Advanced Micro Devices can cram four cores into four socket servers that have roughly the same feeds and speeds as the Power 570, but are more compact and cost a heck of a lot less for processor, memory, and disk capacity. And with Dunnington, Intel’s high-end, six-core Xeon MP chip, the company has put L3 cache on the chip and crammed on six cores running at 2.66 GHz. On IBM, NEC, and Unisys boxes with 16 processors slots (four per chassis, unlike the two per chassis IBM has used in the 570-class boxes since they were invented in 2003), that means customers can deploy 96 X64 cores in a single system image. Even using four-core Dunnington or “Tigerton” Xeon MP chips, you’re talking 64 cores. Even assuming poor scalability beyond eight sockets, these are nonetheless very scalable machines. This is not like a few years back, when Windows in the data center was an idea, but not a choice.
And, IBM charges per core for activations on Power boxes, while Intel is today adding four or six cores at a time into a box for a modest price. It costs $11,500 to buy a Power6 processor card for a Power 570 and then $23,000 a pop to activate the cores. IBM is charging $7,999 for a processor card using Intel’s Xeon X7460 six-core Dunnington, and all six cores are activated. That works out to $1,300 per core for the System x3950, but $28,750 per core for the Power 570–and that is just for the server processor capacity.
As significant as the differentiation in hardware pricing and the similarity in hardware configuration between Power and X64 enterprise servers is the similarly huge divergence in software pricing. IBM’s prices for OS/400 and i5/OS licenses have been raised over time, not lowered. Microsoft has been lowering pricing for Windows Datacenter Edition in the past eight years. It used to cost tens of thousands of dollars per processor back in the days when there weren’t Opterons or multicore chips from anyone but IBM, with Power4. But with Windows Server 2003, Microsoft chopped the price of Datacenter Edition to $2,999 per processor socket, even as Intel and AMD have been doubling and then quadrupling the number of cores in the chips to boost their performance. On database workloads, like the kind running on midrange gear all over the world, this is a huge difference.
Microsoft is charging roughly one-seventh the price it used to with Windows Server 2000 Datacenter Edition with the new 2008 Edition, and it is charging based on sockets, not cores. So as Intel comes out with eight-core “Nehalem” Xeons next year, there won’t be a software penalty. IBM is and has always charged for OS/400, i5/OS, and i based on the core. On a 570-class box, IBM is charging $53,000 to activate the operating system on a core, once you take out Software Maintenance for that operating system and integrated DB2 for i database. Microsoft is charging $2,999 per socket plus just under $40 per user for Windows Server 2008 Datacenter Edition, and then $13,969 for a license that covers an entire server, plus $162 per user. For relatively low numbers of users–like from 40 to 1,200–the difference in software prices really adds up.
Take a look at the comparison table I built showing Power 570 i Edition machines up against IBM’s own System x3950 servers, which scale up to 16 sockets and which look very much like a 570 architecturally. Sit down first, though. Let me make this real simple for you. On a 570 with a relatively low number of end users, the cost of a comparable Windows setup is 10 percent to 15 percent of the Power 570 i setup; on machines with 600 or 1,200 users, the gap closes to 20 percent to 25 percent. What this means is you need to have a 570 supporting thousands of end users to get a machine that is comparable in performance and price to a similarly configured Windows box. And to be precise, you have to configure 7,500 users on the largest configuration in my table for the prices to be the same. IBM seems to be implying a very large number of users on its Power 570 systems; Microsoft assumes nothing and charges for each one individually. Whose business model is working?
As I have said many times, IBM needs to charge less for processor capacity and it needs to have user-based pricing for software across the entire i line, not just for entry boxes. Windows is a threat across the board, not just on relatively small machines.
Now, having told you the result of the comparisons, I will show you how I made them. Power Systems i boxes in the comparison table are set up with 3.5 GHz pr 4.7 GHz Power6 processors. The 3.5 GHz chips are the slowest Power6 cores IBM offers in the box, and not coincidentally, they are the least expensive processor boards and core activations, too. Unlike the prior Power 520 and Power 550 machines, I have not attached disk arrays to these enterprise servers. I did toss in two disks, a RAID disk controller, and a suitable tape drive, but I didn’t want the price of old and new disk arrays to skew the relative bang for the buck of these two generations of machines. I wanted to show you the merit of base server configurations, with an appropriate amount of cores and memory activated, waiting for storage to be hung on the box. Each configuration includes i 6.1 and the PowerVM Standard Edition hypervisor; the hypervisors allow basic logical partitioning. I also removed Software Maintenance fees from all of the systems, even though IBM requires customers to buy it, because I think customers should have the choice of maintenance services or not (or IBM or otherwise) and because I want to isolate system costs from support costs.
To help make your capacity planning a bit easier, I calculated the cost per transaction per minute for the machines based on IBM’s Commercial Performance Workload (CPW) benchmark and my conversion of that rating to estimated TPC-C online transaction processing performance. I also added user counts to the configurations, even though IBM does not price the machines based on users (just as I did with the Power 550 comparisons over the summer to older System i iron and to entry and midrange Windows boxes). I happen to think IBM should price all i5/OS and i licenses based on a basic (and low) per-server charge plus a modest fee, and adding a user component to the configuration allows per-user comparisons across the System i and Power Systems i lines and with other servers. I am not suggesting for even a second that the number of users shown is the maximum number of concurrent users a machine can host, or even an appropriate configuration. The user numbers are based on the user counts IBM has for the user-priced System i 515 and 525 servers from last year, then scaled up as processor counts go up. If a an entry Power 520 with a single core has 40 users, then an entry Power 570 with four cores should be configured with 160 users; if a large configuration of the Power 520 has 150 users, then a large configuration of a four core Power 570 should have 600 users.
On the Windows side, the System x3950 uses the four-core Tigerton or six-core Dunnington chips, and the machines have the memory and disk shown, plus a RAID 5 disk controller and a SAS LTO-4 tape drive, just like the Power 570s in the comparison have. The machines are configured with Windows Server 2008 Datacenter Edition and SQL Server 2008 Enterprise Edition, the flagship Microsoft platform. For virtualization, I through on VMware‘s Infrastructure Standard Edition stack, which includes the ESX Server 3.5 hypervisor and a bunch of other tools. I worked backwards from a real TPC-C benchmark result for an eight-socket x3950 that IBM tested using Red Hat Enterprise Linux 5.2 and its DB2 9.5 database; performance will not be appreciably different in the Windows stack, if past history is any guide.
Next up, I’ll take a look at how the Power 595 i Edition looks compared to earlier big boxes, and then it is time to case out how the i Editions compare to various Unix and Linux options.