The eServer i5 Versus Unix Servers
June 28, 2004 Timothy Prickett Morgan
Say what you will about Larry Ellison and the Oracle database company he founded, but they know how to compete and are not afraid to bloody their knuckles in competitive fights with IBM and Microsoft in the relational database market, which Oracle, along with IBM, created back in the late 1970s. Nothing made this message more clear to me than building the comparisons of the new eServer i5s with Unix servers from IBM, Sun Microsystems, and Hewlett-Packard.
Even though I used Oracle10g in the comparisons between the eServer i5s and Lintel servers last week (along with the open source MySQL database), the extremely aggressive pricing of Oracle’s new Standard Edition One (for two-way servers) and its impact on the bang for the buck on any server that runs it did not really sink in until I configured the Unix servers this week. Oracle is charging $4,995 per processor for the 10g database on two-way machines, effectively matching pricing for Microsoft’s SQL Server 2000 Standard Edition on Wintel boxes, which is $4,999 per processor. And because Windows Server 2003 has per-user charges of $40 per seat and Unix servers do not have these per-seat operating system fees, Unix hardware vendors can compete, and compete well, against Wintel iron. Unix shops can even get away with charging a slight premium for their Unix gear, as IBM and HP do for their respective Power and Itanium machines. Sun Microsystems, as my comparisons show, is hitting the right price points with new its Sun Fire servers, but it is still a performance laggard on a per-processor basis. It takes two Sparc cores to roughly match the performance of an Itanium chip from Intel or a Power chip from IBM.
When IBM launched the eServer i5 machine in May, the company said that it would keep rough parity in the pricing of iSeries and pSeries hardware (now i5 and p5 hardware). IBM cut prices on memory, disks, and processors in the iSeries line and launched the i5 Model 520 and 570 machines with prices that felt a lot like a Unix box. And even as IBM boosted the price/performance of the OS/400 server line by around 40 percent with the launch of the i5, that 40 percent is pretty evenly split between a price decrease for hardware and a boost in raw processor performance across the OS/400 server line.
Pricing in the Unix server market, which has seen IBM, Sun, and HP in a three-way knife fight for five years, is not just aggressive; it’s insane. It is hard to believe that any of these vendors are actually making money on their Unix boxes at the low volumes they are making and the relatively low prices they are charging in order to win deals. Each is content to push software, services, and peripherals sales as a way to make up for the fact that they are not making money directly on their servers. This is my opinion. None of the vendors will, of course, cop to this as a fact of business. But it is, especially when you take into account the vicious discounting that goes on in competitive bids.
Take a look at the eServer i5 versus Unix server table I made and see for yourself. IBM has not yet launched the Power5 processors and their “Squadron” servers in the pSeries line (presumably soon to be renamed the eServer p5), so my table compares the pSeries 615 and 630 servers using the Power4+ chips to the eServer i5 Model 520 and 570 machines. To begin with, IBM has hardware pricing in the pSeries line that aims to match the very low pricing HP has for its Itanium-based Integrity server line. AIX comes bundled on these pSeries machines (and will cost a few hundred bucks per processor on the i5s when it becomes available in August), and with Oracle 10g only costing $4,995 per CPU on two-way machines and $15,000 per processor on four-way machines, a pSeries 615 with no governors on performance and no high-cost 5250 protocol for supporting green-screen applications offers about twice the performance of the biggest Model 520 Express machine for half the cost in a like-for-like base configuration. (All of the Unix machines are configured with the processors, main memory, and disks shown, as well as an AIT3 tape drive from GST that is roughly analogous to the 30 GB tape drive IBM configures with the Express machines.) And because IBM is gearing down the performance of the Model 520 Express machines, which it does not do on the pSeries boxes, since they have no green-screen protocol to try to charge a premium for, the price/performance of the entry pSeries box is about five times better than that of the eServer i5 entry machines. I also still believe that, on a given piece of Power server iron, OS/400 plus DB2/400 will not offer the same performance as AIX plus Oracle. That has been the case for years, and I have no reason to believe this has changed.
With the Integrity server line, HP is supporting HP-UX, Windows, and Linux on the same iron, and the sporadic benchmark results it has released to date on these machines suggest that these platforms perform roughly equivalently. On any given benchmark, on any given day, HP-UX might be tuned to squeeze more transactions out of a machine, but soon enough, HP does another test on the machines on a different operating system and then it shows this advantage to not be due to the operating system at all. What the various benchmarks I reviewed for this comparison showed me is that the Itanium 2 chip running at 1.5 GHz has a performance advantage compared with the 1.45 GHz Power4+ and even sometimes to the 1.7 GHz or 1.9 GHz Power4+, depending on the workload, on a core-for-core basis. Particularly for two-way and four-way machines, the performance advantage is substantial.
But somewhere in HP’s architecture between the eight-way and 16-way machines, something happens, and the scalability curve flattens out considerably. IBM’s Power machines, by contrast, seem to scale more linearly on a lot of workloads. This means IBM’s 32-way pSeries 690 servers using the 1.9 GHz Power4+ processors can handle 1 million transactions per minute (TPM), while the HP Integrity machines need 64 of Intel’s 1.5 GHz Itanium processors to hit the same performance. But OLTP TPM ratings based on the TPC-C benchmark are not everything. On the SPECjbb2000 Java benchmark (which is roughly based on the TPC-C test but is designed to stress a server’s Java virtual machine more than its I/O and memory subsystems), IBM’s 32-way pSeries 690 with 1.7 GHz Power4+ processors is neck-and-neck with the 32-way HP Integrity machine using 1.5 GHz Itanium 2s with around 550,000 operations per second. But because HP’s Integrity machine can scale to 64 processors, it can push SPECjbb2000 to over 1 million operations per second. This big difference in the performance of IBM’s pSeries 690 and the HP Integrity server in the TPC-C test leads me to believe that with its latest run, using DB2 8.1 on AIX, IBM has figured out some tricks to make DB2 and AIX make much more efficient use of main memory than was possible in the past. These tricks are not necessarily bad; they move the state of the art in transaction processing forward. But whatever the secret sauce is, clearly HP and Oracle don’t know what it is yet.
Say what you will about Itanium not having a lot of commercial application support and HP-UX customers having to port their applications over to the new architecture. The move has allowed HP to keep pace with IBM in the Unix market in terms of performance and price/performance, and considering that Big Blue can prop up server profits with the zSeries and iSeries lines and, that HP only has a skinny Tandem fault tolerant server business and a rapidly dying AlphaServer VMS business, on which to charge a premium for its products, it is a pretty aggressive move. It would have been all the more stunning if the “Madison” Itanium 2s had been out the door in the current state in 1999 or 2000, as was originally planned. No one would have been ridiculing HP or Intel as they were (and mostly still are), and Opteron may not have attained the backing of IBM, Sun, and HP, as it has now. Itanium is holding its ground, even if HP is essentially the thin gray line holding back the columns of 64-bit Power, Xeon, and Opteron chips as they advance toward it. HP customers who still want PA-8700+ and dual-core PA-8800 chips can pay for them, but they will not yield anywhere near the same performance per core as the Itanium chips, and they cost a lot more. HP has a little profit room here as HP 9000 Unix shops see which way the market swings with Itanium.
Sun remains a wildcard in the Unix business right now. While it has finally ramped up production on its entry UltraSparc-IIIi, midrange UltraSparc-III, and high-end UltraSparc-VI processors (the latter being a dual-core chip, like IBM’s Power4 and Power5), the performance of these processors is not in the same league as IBM’s Power and Intel’s Itanium chips. And because it takes roughly two 1.2 GHz Sparc cores to match the performance of a Power4 or Power5 core or an Itanium core, Oracle databases cost more on Sun boxes for a given amount of performance. This is why Sun doesn’t run TPC-C tests on its Sun Fire machines, and hasn’t for three years. Still, everybody knows. What this also means is that a Sun box might have to run Oracle10g Standard Edition (which costs $15,000 per CPU) on four processors to match the power of an IBM or HP box that can have only two processors (and thus be configured with the much cheaper Oracle 10g Standard Edition One, which costs $4,995 per CPU). While Sun has boosted the performance and cut the price of its Sun Fire servers in the past year (and by about the same amount IBM just did as it moved from the iSeries to the i5), the pricing of the Sun Fire machines bears more of a resemblance to i5 pricing than to IBM pSeries or HP Integrity pricing.
I guess we know who has the vast installed bases of customers who are not likely to jump platforms.
It’s true. Sun has been the Unix leader for so long that even if it is only beating IBM and HP by a slim margin in quarterly Unix server sales, it has the biggest installed base of customers who are as concerned about the binary compatibility of their applications as they are with raw price/performance. To one way of thinking, Sun is charging a premium where it can because it can get away with it; to flip it around, IBM and HP are cutting prices dramatically with their Unix offerings because they need to in order to shake loose Sun accounts that do not want to migrate.
The same holds true for the eServer i5. IBM has shown more or less consistent price/performance improvement since 1988 with the AS/400, iSeries, and now i5 line. But somewhere in the late 1990s, when the Unix market went nuts, IBM’s AS/400 division decided that it was not going to race to the bottom on pricing with the Unix players. We could argue back and forth as to this being one of the primary causes for a decline in the number of OS/400 servers shipped each year. All I know for sure is that, from the numbers I can pull together, IBM is still not playing the game, but it is hanging back while keeping pace on price/performance.
As was the case when I discussed the comparisons with Wintel and Lintel boxes to the i5s, when you make comparisons you have to remember the superior logical partitioning and workload management capabilities of the iSeries and i5 machines, even compared with the Big Three’s Unix iron. IBM’s dynamic partitions on the pSeries are only granular to the CPU level; Sun’s domain partitions are only granular to the cell-board level; and HP’s hardware partitions are at the cell-board level, and its virtual partitions (vPARS, as it calls them) are only granular to the CPU level. The logical partitions on the iSeries and i5 machines can be any fraction of a processor and can span a whole machine. This makes them more flexible, and allows customers to get overall system utilization up in the 60 to 70 percent range. This is worth something in a Unix market, where processor utilization is lucky to be in the 25 percent range.
By the way, none of the Sun machines in the comparison, except the Sun Fire Enterprise 2900 server, has domain partitions. Sun doesn’t make entry machines that support domains, and will be relying on the “software container” partitions in the Solaris 10, due in October, to offer something akin to what IBM has created with the logical partitions in the iSeries line.