Entry Power System i Boxes Compete Well with Windows Boxes
September 8, 2008 Timothy Prickett Morgan
One of the key reasons why IBM merged the System i and System p product lines this year was to get the i platform more competitive with the rest of the system market. By converging the Power server family and stopping its practice of overcharging System i customers for hardware and software, Big Blue has made the entry Power Systems i Edition lineup–that’s the Power 520 and the BladeCenter S with JS12 blade servers–absolutely competitive with equivalent Windows boxes running online transaction processing workloads.
This is the second time in two years, with two generations of products, that this has been true.
When IBM announced the user-priced System i 515 and 525 servers in April 2007, the company made it clear that it was going to finally compete, dollar for dollar, transaction for transaction, with the Windows boxes that had come to dominate the small and medium businesses of the world. But IBM’s ongoing commitment to compete with regard to the AS/400, iSeries, and System i has always been a big question mark. Sometimes IBM did compete, and sometimes it didn’t and decided to pad its financials by gouging customers for processing capacity because the legacy nature of the workloads allowed it to do so. We could argue about the wisdom of this strategy all day and on into the night, but it is what it is. I am just thankful, on behalf of i shops, that IBM is keeping pace with Windows machines.
That assessment is based on the traditional OLTP spreadsheet I do every year or so to assess the relative bang for the buck that midrange shoppers can get on various hardware and software. And given the considerable amount of work Intel and Advanced Micro Devices have done to keep boosting performance and the increasing level of sophistication of entry X64 servers, and the competition from Microsoft‘s Windows Server 2008 stack, having the Power System i boxes keep pace is all the more remarkable.
To make it simple: X64 processors may now be quad-core, but the X86 and X64 architectures have generally required twice as many clocks or twice as many cores as the Power architecture to get the same amount of work done. And matching up dual-core Power6 systems against quad-core X64 chips available today from Intel and AMD, this situation has not changed. Microsoft might have a new Windows Server 2008 and SQL Server 2008 stack out there, but pricing hasn’t gone down; in fact, it went up a little bit for this software. IBM has cut prices on i5/OS V5R4 and i 6.1, which roll the operating system and database together, and moved to a base license plus per-user charges, allowing it to charge less for the software stack than it has in the past. And, at least as far as entry boxes are concerned, the Power Systems i platforms, whether they are towers or blades, are either meeting or beating Windows boxes of similar performance in either a tower or blade setup.
To reach that conclusion, I took the Power Systems i configurations I put together earlier in the summer as part of a comparison to the two prior generations of iSeries and System i machines (see The Power Systems M15 and M25 Versus Their Predecessors and The Power Systems JS12 and JS22 Blades Versus Other i Boxes) and consolidated these down into one sheet with just the Power 520 and BladeCenter S plus JS12 blade servers. I did a lot of different mixes of processor performance, memory and disk storage, and user counts to give you a sense of how these factors affect the overall price of a system and the price per transaction or user. Each configuration included i 6.1 and the PowerVM Standard Edition hypervisor, which allows basic logical partitioning. The i tower machines have the number of disk drives shown and RAID 5 data protection, while the i blades have two on-blade disks for local storage as well as access to in-chassis storage blades, which are configured with RAID 1 mirroring protection. I also threw in an enterprise-class SAS LTO-3 tape drive for all of the machines, which is perfectly reasonable. In the comparisons, 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.
On the Windows side, Hewlett-Packard is the undisputed dominant shipper of X64 ProLiant tower and blade servers in the world, and comparing Windows running on ProLiants is what matters at this point in the server market. IBM, Dell, Sun Microsystems, Fujitsu-Siemens, and others all have to compare their speeds, feeds, and prices against the ProLiant line, which is the touchstone for workhorse X64 computing. Dell has its fans, and Sun has nifty designs, but HP is the market leader. And so my comparison is this time around limited to HP ProLiant rack servers of equivalent processing capacity to the Power 520 i Edition and BladeSystem c3000 blade boxes with ProLiant BL class blades that have roughly the same oomph as the IBM BladeCenter S chassis using JS12 entry blades. The HP boxes are all configured with Windows Server 2008 Enterprise Edition, the roughly analogous version of the Windows platform that has the features of i 6.1, plus the new SQL Server 2008 Standard Edition database, the reasonable choice for SMB-sized OLTP workloads.
I also threw VMware‘s Infrastructure Standard edition hypervisor stack, which includes the ESX Server 3.5 hypervisor plus the VirtualCenter management, the Virtual Machine File System (VMFS), Symmetric Multi Processing (vSMP, which allows a virtual machine to span multiple cores), and tools for updating, backing up, and clustering virtual machines. This is not a cheap software stack, even on a small box. And there are cheaper Windows editions and freebie hypervisors, but these are not the ones that real companies will deploy for their core ERP systems. Finally, I removed HP’s various tech support options from the box, and ditto for the Windows software stack. There is no way to separate the first year of support from VMware’s stack, however.
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 benchmark and my conversion of that rating to estimated TPC-C online transaction processing performance. For the HP machines, I worked backward from real TPC-C tests on X64 iron and estimated what each HP setup would do. I would be the first to agree that there is a lot of estimating there. But someone has to do it.
Without further ado, here is the entry i and Windows comparison table that stacks these entry boxes up. The setups outlined in the tables in the beige background are what I consider the most typical examples. I am comparing single-core Power6 machines to dual-core Xeon boxes. (Opterons do not have a statistically significant performance or price advantage–not like they did two years ago–and I did not see any point in making such a comparison. HP’s prices, core for core, for Opteron machines are identical to Xeon machines.) For those of you who like pretty pictures instead of analyzing data, here’s how the smallest configurations (single-core Power6 machinery versus dual-core, single-socket HP iron) stack up on a per-user basis:
As you can see, there is a little wiggle here and there because of different software pricing as users are added to the machines, but no box is outside of the discount zone of equivalency. By which I mean the Power Systems i machines are competitive with Windows-X64 iron. Now, look what happens when you double up the processing capacity and start bulking up with users:
Two things to notice here. First, the cost per user on machines with more processing power is lower, and this stands to reason since you don’t buy two whole servers to double the workload–you double memory, you double disks, and you plunk in a second processor. And you don’t have to pay for base operating system or database licenses; you just add a nominal per-user fee to the software on the box. And the VMware virtualization on Windows machines is for a two-socket license (although these are all for single-socket machines), so the incremental cost of virtualization would be zero if capacity were quadrupled from one dual-core X64 chip in the first set of data to two quad-core X64 chips in a fully loaded ProLiant two-socket box. The Power 520 and JS12 blade servers have no expansion room–at least not until IBM doubles up the sockets on the Power 520, which it can do–while the ProLiants do.
The other obvious thing to notice is that the cost per user goes down a lot as you add users to a configured machine. Which means customers should always buy the smallest configuration that will support their typical concurrent user load, and then you are only paying for software licenses for your total user base instead of paying for extra hardware capacity for each user as well as software licenses. Most companies put a little extra capacity in there for peak workloads, but the key here is not to go crazy.
One other thing that is not obvious from these comparisons is the relatively high cost of hypervisors on X64 iron–even after VMware has cut prices in response to intense competition from Citrix Systems, which owns Xen, and Microsoft, which is now shipping Hyper-V. (Virtual Iron, which has a Xen-based hypervisor, too, has been the price leader for several years.) By switching to an open source MySQL database from Sun, a low-end hypervisor from Citrix or Microsoft, and a Small Business Edition of Windows 2008 (when it is available), the price of the Windows stack can be brought down a lot.
Of course, you could argue that a database engine and ERP box doesn’t need to be virtualized, and therefore adding the extra costs of the virtualization hypervisor and associated management tools that make it useful is not a fair comparison at all. But in 2008, I think this is becoming an increasingly invalid argument. More and more companies are deploying virtualization, and I think that many companies are going to be deploying their n-tier application architectures on smaller numbers of virtualized boxes–the kinds of things that AS/400, iSeries, and System i customers have been doing for years with database and application servers using subsystems (a precursor to logical partitions and akin, in some ways, to virtual private servers for Windows, Linux, and Solaris today). I think data center managers are being increasingly pressured to simplify their infrastructure, and not for green reasons–unless you are talking about money. The green benefits, in terms of energy efficiency, are only a concern for the very largest data centers with exploding workloads, and maybe a bunch of midrange shops in urban areas where energy and office space are priced at a premium. And while I am thinking about it, the OLTP performance estimates in these comparisons do not take into account any processing overhead that the hypervisors impose. I have no idea how much capacity gets chewed up by the current crop of hypervisors. That is a story for a different day.
Next week, midrange Power 550 boxes and JS22 blades versus Windows machinery. Let’s see if IBM can keep competitive here, too.