The Four Hundred

Making comparisons on technology, pricing, and performance across any two architectures is always a tough piece of work. And there is a long line of people who are quick to point out all the inadequacies of any approach to doing so. But the fact remains that IT managers need to make such comparisons. This week, I am taking a shot at comparing my iDeal iSeries machines to the real IBM iSeries line and Intel-based servers. It gave me a headache, to be honest.

No pain, no gain, I suppose. But I love this kind of work. And the headache has more to do with the flu than spending days of my life building spreadsheets that I hope will show how to build a better iSeries.

For those of you who might have missed the past couple of issues of this newsletter, I have been designing and pricing my own OS/400 servers, which I am calling the iDeal iSeries. I started with a few simple ideas, and I am writing these stories in real time; I didn't write all of these stories and then publish them after knowing what the end result would be. I didn't know how my iDeal iSeries machines would compare to Windows configurations until after midnight last Friday. (That's not a typo. My wife was working late at the office, so I used the time while the kids were asleep. It's a damned good thing I love my work, with the hours I have to keep.)

The first rule is that the iDeal iSeries should use the best technology IBM has. I configured low-end machines akin to the Model 270s with IBM's forthcoming uniprocessor PowerPC 970, a 64-bit variant of the Power4 chip that IBM will be selling to Apple. For bigger machines, I just grabbed the current pSeries line from IBM. You can see a table of salient characteristics for the iDeal iSeries machines by clicking here.

I did a radical thing on pricing. I reviewed IBM's pricing on memory cards, disk drives, and disk controllers for the iSeries and compared it with the pSeries line and the high-end xSeries 440, IBM's eight-way Pentium 4 Xeon MP server (soon to be a 16-way). For the midrange and high-end iDeal iSeries, the rule on pricing for hardware was to charge exactly what IBM charges in the pSeries line. Pricing for the entry machines was more consistent with the high-volume, low-priced products that vendors of Intel-based servers kick out to the tune of millions of machines and tens of billions of dollars worldwide. I set pricing for memory, disk, and disk controllers for the iDeal iSeries so that the entry machines were consistent with the xSeries 440, the midrange machines were consistent with the pSeries 630 and 650, and the high-end was consistent with the pSeries 670 and 690. Out there in Unix land, these pSeries prices are discounted like crazy, but they are so low compared to actual iSeries prices that they would feel like the deal of the century to the iSeries base even without discounts.

Because the entry machines use very powerful PowerPC 970 processors--which actually rival the new Power4+ chips in raw processing power, but which would not scale well beyond four-way processing--I had to slug them, just like IBM does with its OS/400 governors in the real iSeries line. There would be no way to create a low-end machine with relatively limited processing capacity (in terms of the green-screen RPG and COBOL workloads) that many companies that comprise the OS/400 server installed base have. However, rather than isolate this capacity and do nothing with it, as IBM does, I want to do something radical with it, something that no Intel server or Unix server or mainframe server does. I want this so-called "shadow capacity" to be available--and only available--for high availability clustering. I want the iDeal iSeries to have high availability built in, configured from the get-go, utterly transparent to the end user. These iDeal iSeries machines make use of the integrated high-availability features, logical partitioning, and virtual LAN capabilities of OS/400 to create a super-reliable machine. (This shadow capacity is not available on the midrange or high-end iDeal iSeries machines, since these customers are already accustomed to doubling up their production servers, and, more important, at the prices I have set, buying two iDeal iSeries machines would be a lot less expensive than buying one real iSeries box.)

For OS/400 software, I want to shift to a mix of CPU-based and user-based pricing. No governors on 5250 performance. No interactive features. OS/400 comes in three editions--Standard, Advanced, and Enterprise--and so does DB2/400. These editions offer different scalability (1-4 CPUs, 2-8 CPUs, and 16+ CPUs respectively), and the pricing for the CPU-based elements reflects the fact that companies with big servers require more R&D to develop their products, and they should pay a little more for their OS/400 and DB2/400. But the biggest component of software is access to databases. Companies that want to use the 5250 protocol to link actual green-screen terminals or emulated terminals from PCs to their iDeal iSeries pay $225 a set; companies that access databases through ODBC or JDBC interfaces pay $50 a seat.

I reckoned CPW ratings for each of these machines--based on TPC-C online transaction processing benchmark test results. These CPW ratings were used to provide appropriate real-world estimates for base and maximum end users. I assumed that companies were running transactions that were a lot less dense and a lot less tuned than those used in TPC-C tests. Just to be able to make a comparison, I figured a base machine would support a number of users equal to CPW divided by 20, and a maximum practical machine would support a number of users equal to CPW divided by 5. Many of you have workloads that are a lot less dense and users who are not as efficient as simulated TPC-C users. I am not arguing that these are necessarily representative. At many companies, maximum practical users is more like CPW multiplied by two. I just want to start somewhere to show how this might look, to start you as iSeries customers and IBM as the sole iSeries supplier thinking about how to do things a little differently.

To see the table of iDeal iSeries base machines, including their hardware, OS/400, database, and connectivity costs, click here. (Note: when I say base configuration, I mean the smallest practical configuration IBM should sell for OLTP workloads, not a bare bones machine. This base machine does not include the cost of expansion chassis, tape backup, connectivity features, and so forth.)

To see the table of iDeal iSeries maximum configurations, including their hardware, OS/400, database, and connectivity costs, click here. (Note: when I say maximum, I mean maximum practical configurations for OLTP workloads, not a fully loaded machine. One more thing: I found a typo in my spreadsheet from last week's issue that built those maximum configurations. I pointed to a wrong column and accidentally boosted the pricing on the maximum iDeal iSeries configurations. The table above, and the comparisons I make to Windows platforms, are correct. My apologies for the error. The prices didn't change that much, as you will see, but the iDeal iSeries machines look even better than I expected at the low end because of this error, and the great news is that the pricing scheme I have come up with can compete toe to toe with anything the Intel-Microsoft duopoly can throw at the midrange. I was hoping this would be the case, and it was a tremendous relief to me. Headache relieving, in fact.

In last week's issue, I compared the configured base and maximum configurations of the iDeal iSeries with IBM's two Green Streak machines. The iDeal iSeries machines were better deals in some respects, much better in others. (The deal aspect of these imaginary machines is the whole point, as I have explained. I want customers to want to buy them, rather than have to be sold them.)

Beating the existing iSeries is easy, and while beating Green Streak configurations was nice and a little surprising, pricing for Windows on Intel is the thing to beat in the entry and midrange market. We already know that the 64-bit Power chips from IBM are better--in terms of efficiency and elegance of design--to the 32- and 64-bit chips that Intel kicks out. But Wintel boxes set the pace in pricing--at least until Lintel machines get in the driver's seat a few years from now, as Linux gets entrenched as a commercial data-processing alternative to Unix and Windows.

To compare Wintel and iDeal iSeries servers, I had to translate TPC-C OLTP performance estimates back into CPWs, the current iSeries relative performance metric, and the one I chose for the iDeal iSeries. I went to the TPC Web site and grabbed all the relevant benchmark results for ProLiant servers from Hewlett-Packard, which is the volume leader in the Wintel space. I built a table showing what the estimated CPW ratings were for every modern machine on the list, which included machines running Windows 2000 or the future Windows .NET Server 2003 (Whistler) operating system from Microsoft and SQL Server. The current HP Wintel lineup includes Pentium III, Pentium III Xeon, Xeon DP, and Xeon MP servers. I have built a table of HP's TPC-C configurations used in its tests, calculated CPW ratings for these Wintel machines, and added estimated TPC-C and CPW ratings for two small servers, which were necessary for a comparison with an entry iDeal iSeries machine. This table also gives ratings for the 1 GHz and 1.6 GHz Integrated xSeries Server cards, which are used in some of my iDeal iSeries configurations, as I will explain below.

That was the easy part. The hard part is reconciling the fact that OS/400 has a sophisticated workload manager, logical partitioning, and Integrated xSeries Server (IxS) cards that allow database, application, and infrastructure workloads to be isolated from one another. Wintel machines do not have this capability, and are back somewhere in 1990 (or earlier) in this regard. That means, for a given set of applications, a Wintel setup has to be comprised of isolated machines supporting isolated workloads. This is what contributes to server sprawl. Most modern Unix operating systems, by the way--and Linux, too--do not have very good workload managers, which is why all the Unix vendors have been trying to create logical partitions that rival that in the iSeries line. (So far, they have all fallen short of that goal, even though all the big vendors now offer some form of partitioning.)

In the interest of saving time, I only created three possible scenarios--a small customer needing an aggregate of 1,500 CPWs to support databases, applications, and Web, print, and file infrastructure; a midsized customer needing about 5,000 aggregate CPWs; and a large customer needing 15,000 aggregate CPWs. I specified that half of that CPW capacity would be dedicated to the database, 20 percent to the application servers, and 30 percent to the infrastructure servers that feed into the application servers or sit beside them doing print and file serving.

The HP-Windows setups consist of networked machines with workloads distributed to different servers. I created two different iDeal iSeries configurations for each scenario. The first is a single machine using logical partitions that consolidates all of the necessary database, application, and infrastructure tiers on a single machine. The second configuration puts the database and application tiers on a single iDeal iSeries machine, then uses IxS cards running Windows 2000 to support Web, file, and print serving.

In all three scenarios, a convincing economic argument can be made for choosing the iDeal iSeries setup over a pure Wintel setup. The technical merits of the iDeal iSeries, if such machines were brought into existence, would put them over the top. I will walk you through the first scenario, then you can look at the other two to see what I mean.

The first scenario requires an aggregate of 1,500 CPWs to support 200 users of a database-driven application suite, who also need to access the Web and do print and file serving. (See this table for the specs on the first scenario.) I chose four HP ProLiant DL320 machines using 1.26 GHz Pentium III chips--basically, the least powerful rack-mounted machines HP offers--to support those tiers. I grabbed current prices for the configurations--including processors, memory, and storage--from HP's online store. That HP network results with an overcapacity of more than 1,200 CPWs, but at $17,072, the hardware is so cheap that no one would care, until they thought about it for a second. (Wintel buyers don't think about it because CPU capacity is seemingly so cheap.) Adding Windows 2000 to all of the machines, and putting SQL Server on the database server, drives the total price up to $59,397. If customers want to make the platform components redundant, to increase the availability of the servers, they have to double up on this hardware and software, which drives the cost up to $118,794. That's not exactly a small piece of change, is it? But those entry Wintel hardware prices and CPU-based and user-based software prices sure make it seem like it on the front end.

Now let's look at the comparative iDeal iSeries machine. I think an iDeal iSeries Model 382 with dual 1.4 GHz PowerPC 970 processors, 6 GB of main memory, and 450 GB of disk capacity can do the 1,500 CPW job on a single machine using four logical partitions (OS/400 for the database and application tiers, Linux for the two infrastructure tiers). At my pricing, the hardware cost for that setup is $33,743. That is twice the cost of the base Wintel iron, which made me jumpy. But then I configured the software. Adding OS/400 Standard Edition costs $800 (very inexpensive) and adding DB2/400 Standard Edition with ODBC/JDBC user access to databases costs another $15,000 at the prices I have set. The total iDeal iSeries Model 382 configuration costs $49,543 using Microsoft's own protocols to access databases. (How sweet that is!) For those customers who want to use the 5250 protocol to access their databases, the total cost is $84,543 for 200 users on the Model 382. This is considerably higher than the $59,397 of the Wintel solution, but remember, the iDeal iSeries entry machines are craft and sly. This machine, which I have slugged, has a shadow excess capacity of over 3,000 CPWs just sitting there. Activating just half of this shadow capacity in high-availability-only partitions would give a high availability setup for exactly the same money, minus the cost of the high-availability software solution, of course. The Wintel high availability solution costs $118,794 (minus clustering software as well), but the iDeal iSeries high-availability solution costs between $49,543 and $84,543, depending on how many users come into the applications, using what database protocols. Even with $25,000 to $40,000 for high availability software costs, the iDeal iSeries would be competitive with clustered Wintel boxes--and a heck of a lot easier to manage, to be sure.

See how this works? This is not that hard. This is not rocket science. I've studied rocket science, and this ain't it.

In the small customer scenario outlined above, shifting infrastructure workloads to IxS cards results in slightly higher prices for the overall setup. The shadow iDeal iSeries capacity can be used for a high availability setup for the database and application layers, but customers have to buy redundant IxS cards to have failover clustering for their Windows infrastructure tiers. In the midsized and larger customer scenarios, this second approach results in a lower cash outlay than the all-iDeal iSeries solution. The savings associated with staying in the entry iDeal iSeries line (those with the PowerPC 970 chips) can be substantial for high availability configurations, since the bigger iDeal iSeries machines are based on the pSeries line and do not have shadow CPW capacity to dedicate to high availability (or, indeed, any workload that might be thrown into that space).

You can see how the midsized scenario for supporting 700 users on a 5,000 CPW configuration works out for the iDeal iSeries against a Wintel network by clicking here.

You can see how the large scenario for supporting 2,100 users on a 15,000 CPW configuration works out for the iDeal iSeries against a Wintel network by clicking here.

I'm not going to waste a lot of mental oxygen going through each scenario in detail. I will point out that I nailed the pricing just right, compared to Windows in the midsized scenario--meaning that an iDeal iSeries configuration costs about the same as the Windows setup when both are using ODBC access to databases. However, the Model 851 in the midsized scenario does not have shadow CPW capacity, so high availability configurations are pricey (but no more than for redundant Windows setups). Pricing for 5250 users is high on this machine, mainly because it requires OS/400 and DB2/400 Advanced Editions. (This is the price you pay for expandability). But using the iDeal iSeries Model 384, and a bunch of IxS cards to support the workload drops, the high availability price, even for 5250 users, to less than it costs to do high availability on the Windows platform, and the ODBC high availability configuration using the Model 284 costs just a few grand more than the non-high-availability setup for Windows.

On the large configuration, the 16-way Model 871 is pricey against the new ProLiant DL760 eight-way "Gallatin" Pentium 4 Xeon MP server that was formerly known under its code-name, "F8." This Model 871 solution requires the Enterprise Editions of OS/400 and DB2/400, and it is priced like a big Unix box, not a Wintel server. The Model 871 setup costs about twice as much as the Wintel network that includes the F8 server and seven other machines feeding into it. However, a Model 851 server to support databases and applications equipped with IxS cards for infrastructure workloads costs half as much as the Wintel solution using the ODBC protocol to access databases, and it costs about the same as the Wintel solution when using the 5250 protocol. This is, I think, not only acceptable, but a great deal.

Finally, because I know you want to try to figure out how these iDeal iSeries machines compare against the entire iSeries and more recent AS/400 lines, I created a spreadsheet that lays it all out, including current IBM list prices for new and vintage equipment. This is a large spreadsheet, but you will find it valuable and you can see it by clicking here.

Next week, in the final installment of this story series, I will examine the impact that machines such as the iDeal iSeries will have on IBM's revenues and profits. This is the key factor in determining if IBM will ever do something as crazy or sane (depending on your viewpoint) as I have been suggesting.