Bang for the Buck: Midrange i5 Servers Versus the Competition
August 21, 2006 Timothy Prickett Morgan
In the prior two installments in this Bang for the Buck series of articles, I looked at how various i5 520 machines stacked up against similarly powered and configured Windows, Linux, and Unix servers. The first article looked at the i5 520 Value and Express Editions and their competition, and the second article did comparisons to the i5 520 Standard and Enterprise Editions. This week, I examine the relative performance and value that the i5 550 delivers compared to rival boxes.
As was the case with the entry i5 520 machines, the advent of dual-core X64 processors and aggressive, core-neutral software pricing on Windows and Linux software is giving the i5’s competitors a big leap in bang for the buck. The gap was large on entry machines with one or two processor sockets, and the gap appears to be even larger between the i5 and rival boxes that have either the same number of cores (if the per-core performance is roughly equivalent to the Power5+ chips from IBM) or more cores if they don’t quite stack up.
While many companies argue that the premium that they pay for the integrated i5/OS platform is well worth the money because of the operation of the machine is easier and less costly than Wintel, Linux, or Unix boxes and it takes fewer people to manage the systems, the size of the gap is now large enough that you could pay for many employees with the upfront savings you get by using a non-i5 platform. People costs are operational costs that are written off immediately, but capital equipment expenditures on hardware and software have to be depreciated over three to five years. In effect, by investing large sums in an i5 system, you are using future money and locking it up. This is something that many business owners do not want to do. IBM seems to be banking on the idea that the inherent value of the i5/OS and OS/400 platform is one that its 215,000 customers worldwide will be able to justify to their bosses, the owners of the business or the shareholders. IBM is also probably betting that the legacy lock-in, since RPG applications and DB2/400 databases do not really run on other platforms, gives it an opportunity to charge a premium.
I don’t happen to think this is wise. But I am not running IBM. However, it is my job to help i5/OS and OS/400 shops quantify the premium that IBM is asking you to pay with your continued investments in the platform. It is up to you to decide if it is worth it.
If the gap in the performance and price/performance of the i5 550 and its competition is larger than I would like it to be, the good news is that the vast majority of the installed i5/OS and OS/400 base buys an i5 520 system these days, where the performance gap is not as pronounced. However, customers who think they will grow into i5 550 and larger machines have to be conscious that these machines–at least at list price–get progressively more expensive for raw capacity, at least if you reckon the numbers using reasonable base configurations.
This has always been the case in the minicomputer, midrange system, and midrange server markets, as you well know. Entry machines and top-end machines provide the best bang for the buck, and the middle machines provide the cash that lets the vendor be generous at the low and high end of the line. All vendors have done this with their midrange products for decades.
The Metrics of Comparison
As I have done in prior Bang for the Buck stories, I have created a table outlining the feeds, speeds, and pricing of the midrange servers I compared. The machines in the table have the hardware features shown. I have tried to keep the configurations across server architectures and operating system platforms as similar as is practical based on the natures of the product lines. I realize that there are many ways to skin this cat, and I welcome your feedback.
I am aware that I am showing the estimated or actual OLTP performance of a given processor complex and comparing the cost of a base configuration to this estimated top-end performance for the machine. In this way, I am trying to isolate the base cost of a server and show its potential performance on the TPC-C online transaction processing benchmark. Yes, the Transaction Processing Performance Council frowns on this sort of thing. Someone has to do like-for-like comparisons, and it is either going to be you or me–and I figure you have better things to do, like read this story after letting me do the work.
On each server shown in the table, I have put a RAID 5 disk controller in the box, two 36 GB disks, and 2 GB of main memory for each processor core in the box. Each server also has a basic tape backup; the type I configured to the box is shown in the table.
In terms of the software stack on these basic servers, I have added an operating system and a relational database management system, and unlike in past years, I have thrown in virtual machine or logical partitioning hypervisors, since I think people are going to start using these in production. The i5 has had such software embedded for years, and to make it a fair comparison, this functionality should be added to X64 servers as well. For the heavy configurations of the Windows and Linux machines in the table, I added in VMware‘s top-of-the-line ESX Server 3 with all of the bells and whistles. And for the cheapo configurations, the Windows boxes have Microsoft‘s freebie Virtual Server 2005, while the Linux machines, which are configured with Novell‘s just announced SUSE Linux Enterprise Server 10, have the integrated and free Xen 3 hypervisor from XenSource. I put SQL Server 2005 Standard Edition on the Windows machines, and Oracle Enterprise Edition on the Linux and Unix boxes. The Unix boxes running HP-UX use HP’s own Virtual Server Environment partitioning, the IBM p5 boxes use the Virtualization Engine hypervisor (also used with the i5), and the Sun Microsystems boxes use Solaris containers. I know that the latter is not as sophisticated as some of the other hypervisors–since containers have a shared Solaris kernel and file system underneath virtual machines–but if you want, you could put VMware ESX Server 3 on the Opteron boxes and run Solaris 10 inside the partitions.
None of the configurations have any hardware or software support costs added in, and where vendors put these in as a base requirement–as IBM does with Software Maintenance on the i5 line–I have stripped these costs out. Pricing is just for system acquisition and basic installation support. Obviously, getting break-fix maintenance and real software support costs money. But that is a subject for a different day.
How the i5 550 Measures Up
If Linux and Solaris lead the price/performance on entry boxes, with Windows close behind, Windows machines are leading in this midrange category primarily because Microsoft charges a lot less for its SQL Server 2005 Standard Edition database than Oracle charges for Oracle10g Standard Edition. With Microsoft, it is simple: You pay $5,999 per socket, and Microsoft doesn’t care how many cores you have. With Oracle, you count the cores, multiply by 0.75, round up to the nearest whole number, and pay $15,000 times that number. On a four-core box, SQL Server costs $11,998, Oracle10g costs $45,000. This really starts to add up.
Of course, Oracle doesn’t have anything on IBM when it comes to database pricing. While Big Blue cut the i5 520 customers some slack in late January, when it cut the cost of i5/OS on these machines to $21,000 per core, it still costs $40,000 per core to add i5/OS to an i5 550. So to add the same four cores, you are talking $160,000 of effective pricing for the database (a base machine comes with one i5/OS license and 5250 capability, called Enterprise Enablement, for that core). And if you want to use the 5250 protocol, it costs $50,000 per core on top of that. While IBM bundles in the costs in the base machines for i5/OS and DB2/400, when you work it backward, the basic i5/OS software stack represents about half of the cost of the i5 550 systems running Enterprise Edition shown in the table. And on the Standard Edition machines, i5/OS and DB2/400 represent between 75 and 80 percent of the cost of the configurations shown. IBM doesn’t say it that way, but that’s how the math works. IBM says you buy all this expensive hardware and you get some free software. I can assure you of this–you are most certainly paying for that software, even if IBM wants to book the money in its Systems and Technology Group as hardware to make Wall Street happy. Or, to be more precise, to make it less unhappy than it might otherwise be.
The other thing that is immediately obvious from these comparisons is that IBM is not tuning i5/OS and DB2/400 to the same degree that it is doing so with AIX and its own DB2 V8 and V9 databases or Oracle10g. The performance gap between similar i5 and p5 hardware running these two different stacks–i5/OS and DB2/400 and AIX and Oracle10g–is getting larger. I don’t know what the story is, to be honest. But I do know that the CPW benchmark test that IBM uses to gauge relative performance of the iSeries and System i5 line is based on the TPC-C benchmark (shhhhhh, don’t tell the Transaction Processing Performance Council), and so is the rPerf benchmark that IBM uses for a similar purpose on the System p5 machines. CPW is approximately the peak TPC-C performance of the i5 machine divided by 9.95, while rPerf is the peak TPC-C performance of the p5 machine divided by 10,250. These ratios have held more or less for years.
The gap between the i5 and the p5 is large. A four-core i5 550 using 1.9 GHz Power5+ processors has a peak performance of about 139,300 transactions per minute (TPM), but a p5 550 machine using four cores of 2.1 GHz Power5+ processors is rated at approximately 254,800 TPM. Some of that difference is clock speed, of course. About 10 percent, based on clock speed. That leaves 90,700 TPM unaccounted for except in raw software performance or benchmark gaming. Considering that IBM is using CPW and rPerf to tell people what size computer to buy–and how much performance they will get from upgrades–I don’t think it is “benchmarketing.” I think the DB2 team working on Unix and Linux has been doing a lot of stuff to get the database to hum on OLTP workloads. Whether these techniques translate into DB2/400 or not, I cannot tell. What I do know is that the AS/400 and iSeries division used to be a big proponent of the TPC-C test, and when the machine started to look bad compared to the RS/6000 and pSeries, suddenly there were no more AS/400 or iSeries TPC-C tests formally announced. The gap between an AS/400 running DB2/400 and an RS/6000 running Oracle used to be around 10 percent. Now, with the i5 and the p5, the gap is 65 percent.
There’s a reason for everything someone does and doesn’t do in this world. I think IBM doesn’t want you to think about this too much. Which is why I brought it up.
The other obvious thing that you will see from the table is that Intel’s “Paxville” MP processors, which are used in four-socket servers, are not so good compared to Opterons. Which is why Advanced Micro Devices has taken a quarter of the server business and more than half of the high-end X86/X64 business in under three years. Sun’s brand new 1.8 GHz UltraSparc-IV+ processors are about as powerful as Opterons, give or take, and Intel‘s new dual-core “Montecito” Itanium 9000s can go toe-to-toe with a p5-based Power5+ chip, too. HP is, yet again, charging too much of a premium for its HP-UX boxes–for both the HP-UX software and the underlying hardware–compared to IBM p5 machines and its own Opteron boxes. But having said that, HP’s pricing has come way down as it moved to the dual-core chips in the forthcoming rx6600 machines (which have not even been announced yet), and they look like a real bargain compared to the i5 550 running i5/OS Standard Edition.
Don’t even get me started about comparisons to i5/OS Enterprise Edition. (Too late.) A four-core i5 550 with Enterprise Enablements and i5/OS Enterprise Edition costs almost half a million dollars and delivers 35 percent less performance than an eight-core, four-socket Opteron box from HP running the Windows stack. The Windows-Opteron setup, including Client Access Licenses for a whopping 600 end user seats–costs one-sixth as much. So instead of paying $3.57 per TPM, you pay 39 cents per TPM. To put that another way: You can afford to pay the salary of five full-time programmers on the difference in the cost between these two setups. If you include overhead costs on personnel, its 3.5 programmers, or three really good ones who make six figures a year and have big bonuses.
You have to really, really like the i5/OS operating systems and your legacy code to pay such a premium.
Up next, the bigger midrange boxes: the i5 570, where it is harder for Windows and Linux boxes to compete because the volume servers based on X64 processors in two- or four-socket configurations just don’t scale any higher without a lot of engineering. Few companies have done such engineering because it is expensive and the market is relatively small for such big servers.