Entry Unix Servers: It's a Tighter Three-Horse Race Now

by Timothy Prickett Morgan

With all of the talk about volume Wintel and Lintel servers based on Intel Pentium 4 and Xeon DP and MP processors, it is easy sometimes to forget that a tremendous amount of innovation is still going on in the Unix server space. To be sure, these are not the heated days of RISC/Unix competition from years gone by, but the intense competition has not diminished among the big three Unix server suppliers even if their core technologies have, in many cases, changed dramatically.

It is just amazing what a difference a year can make in this part of the market--both good and bad, depending on if you are talking about IBM, Sun Microsystems, or Hewlett-Packard. Each of these vendors brings their own advantages to their Unix server lines, and they are all--without exception--facing some challenges. This is what makes the entry Unix server market interesting.

Entry Unix Server Market Overview

Without a question, IBM has been an increasingly dominant factor in the Unix server business, across all form factors and SMP scalability, since the Power4 processors debuted four years ago. IBM got into the Unix server and workstation business in 1990, and got a little traction with the "Deep Blue" RS/6000 SP PowerParallel commercialized versions of that machine, and started to gain a head of steam with its own "Star" family of 64-bit PowerPC processors beginning in 1997. As the dot-com boom got under way, IBM got aggressive about smaller Unix servers and became a contender in the market for pizza-box infrastructure servers that were pioneered and sold like hotcakes (and code-named "Flapjack" to reflect their popularity) by Sun Microsystems. By 2001, when IBM delivered the first dual-core processor in the world, the Power4 "Spinnaker" chip for its "Regatta" line of pSeries servers, the dot-com boom was over, the binge spending was all but done, there was a glut of server capacity in the data center, and IBM had surpassed both Sun and HP as the RISC/Unix supplier with the highest single-core performance. From this Power4 foundation, IBM has done three more revs on the processors--Power4+, Power5, and just last month with the initial Power5+ chips--and has revamped its server portfolio once (in the summer of 2005, with some new tweaks in more recent entry Unix boxes).

IBM has been relentless in driving up the performance of its pSeries Unix machines, but in recent years, both HP and Sun have had problems with their own processor technologies, which have made it a little easier for IBM to gain some market share and get a lot of mindshare.

Hewlett-Packard, for instance, has taken a tremendous amount of grief because of its decision to stop development of its homegrown PA-RISC processors and move its entire Unix server line over to Itanium processors. The 18- to 24-month delay in the delivery of the initial Itaniums, the "Merceds," plus their relatively weak performance meant that HP had to keep the PA-8700, PA-8800, and PA-8900 chips alive and continue to develop them for its HP 9000 and Superdome servers. If the second-generation of Itaniums, the 1 GHz "McKinley" chips, had come out prior to the Power4 announcements, as was clearly the plan, HP would have been sitting pretty. And the next generation "Madison" Itaniums, which clocked at 1.5 GHz when they were announced in 2003 and were expected to be pushed up to 1.8 GHz or even a little higher in the summer of 2004, would have been more than enough to keep pace with Power4+. The dual-core "Montecito" chips might have hit the market in late 2004 or early 2005, and been right there when IBM debuted its new Power5 processors and the "Squadron" p5 servers that came out in the summer of 2005. But the Merced delay pushed the whole roadmap out, the Madison chip topped out at 1.6 GHz with the 9 MB on-chip cache (instead of 1.8 GHz or 1.9 GHz), and now, as of October 2005, the dual-core Montecito processors that were expected to be finished by the end of 2005 (after several delays) and rolled into new servers in early 2006 are now expected in mid-2006--and that is if Intel doesn't slip again. Adding to HP's pain is the fact that Intel has scaled back the clock speed on the Montecitos, which will come out at 1.6 GHz with 12 MB of cache per core when the design was intended to scale up to 2 GHz or even 2.2 GHz. Moreover, the "Foxton" performance acceleration feature that was expected to debut in the Montecito chip is going to be deactivated, as will be the 667 MHz front side bus that would have been necessary for the higher clock speeds. While Intel is saying it can deliver on the 2X performance increase promises it was making to Itanium customers with Montecito processors, what seems clear from all of this is that Intel was shooting for a nearly 3X increase with Montecito compared to Madisons--an increase that would have finally, without question, put Itanium at parity or better compared to IBM's Power processor family.

Woulda, coulda, shoulda. Didn't.

The Montecito delays have meant HP was still trying to peddle hot (meaning temperature, not attractive like supermodels), single-core Itanium processors as 2004 came to an end, and it will now have to continue to do so through the remainder of 2005 and the first half of 2006. Any way you want to cut it, there's somewhere from three to four years of cumulative delays in the Itanium processor roadmap, and no one has been hurt by this more than HP. If the executives in HP's Enterprise Storage and Servers unit aren't furious, they must be on Prozac and fine single-malt scotch.

The reason is quite simple. HP's server platforms were ready to transition to Itanium a lot earlier than when Intel had Itanium ready, and that cost HP some money and some face in the market. And because the Itanium delays are cumulative and HP was not able to leapfrog IBM, it had to make up for missed Itanium performance with server price cuts. In a very real sense, those delays hurt HP's pocketbook a lot more than they hurt Intel's. It would be reasonable to question whether or not HP has lost as much money from the Itanium delays as it was supposed to save by partnering with Intel and getting out of the chip making business, but only HP and Intel know that for sure.

This is certainly true in the entry Unix server market, where HP's product line has not been able to show a very substantial performance improvement in the past year and HP has not cut prices very dramatically--not at list price anyway. In competitive bids, there is no doubt that HP has had to cut its server prices quite substantially over the past two years to compete with IBM. Those price cuts were not supposed to be necessary because those Integrity machines were supposed to have a lot more performance and, therefore, offer a lot better bang for the buck.

It gets more interesting now that Sun has delivered its long-awaited "Galaxy" line of entry Sun Fire servers--which also suffered from their own delays coming to market. HP and IBM will both have to contend with a revitalized and aggressive Sun, which knows the entry Unix server business that it created better than either Big Blue or Big Gray, which would be a good nickname for HP. (Does that give Sun the moniker Deep Purple?)

After years of its own delays and disappointments with the UltraSparc-III processors, which pushed out the UltraSparc-IVs and basically killed the UltraSparc-Vs as originally conceived, Sun has an interesting and competitive Unix sever offering with the initial Galaxy servers. These are much more interesting than the not-so-impressive performance of the cut-down UltraSparc-IIIi variant of that processor, which was used in some entry Sun Fire Unix servers. The Galaxy servers might be small, but for some workloads, they offer the same price/performance that market leader IBM delivers with its Power5 machines and they can compete head-on with Itanium-based Integrity servers from HP. And while the initial Galaxies do not offer the same top-end performance as either Power5 or Itanium platforms on all workloads, they also cost less and offer density and power and cooling savings. Sun can at least compete, which is more than anyone has been able to say for several years.

To put it bluntly, Sun is back in the game with Solaris 10 and the Galaxy servers, and the company had better be working on extending that line to four, eight, and 16 socket machines as fast as it possibly can.

And, to Sun's and HP's credit, their respective Galaxy and Integrity entry server lines also support and are certified to run Windows and Linux in addition to their respective Solaris 10 and HP-UX 11i operating systems. IBM can support its AIX 5L Unix and Linux on the new p5s, but it cannot offer Windows. And only Sun can offer 32-bit X86 Linux and Windows support that has any decent performance. Running 32-bit Linux or Windows on the Itanium-based Integrity line from HP means putting up with pretty substantial degradation; the Integrity machines do offer native Itanium Windows and Linux support, but customer applications have to be ported and recompiled from X86 platforms. IBM's Power5 machines, like all other Power-based servers, cannot support Windows--unless you want to try to hack an old CD of Windows NT 3.51 onto them, which was briefly available (and I mean briefly, like a matter of weeks) for 32-bit Power workstations and servers back in the 1990s.

The Metrics of Comparison

Entry Unix servers are used in a wide variety of applications, ranging from the core data processing box for small businesses (including applications and databases), as well as Web and other infrastructure servers for larger businesses. Increasingly, thanks to the integrated clustering capabilities in DB2 8 and Oracle 10g databases, these machines are also being used in clustered database environments, sometimes replacing big SMP Unix and mainframe boxes that support OLTP or data warehousing workloads. (This doesn't happen as much as many vendors would have you believe, but it does happen, and for sound economic reasons.) And, because of their superior number-crunching capabilities, these entry Unix machines are also used in high-performance supercomputing clusters.

Gauging relative performance and price/performance for the entry Unix server market is a bit tricky. No vendor runs a complete set of benchmarks that describe all manner of workloads across its entire product line. And vendors try to show as many different benchmarks as possible without allowing direct comparison to competitor's boxes. So, for instance, IBM and HP will show results on the TPC-C online transaction processing test or the Linpack Fortran benchmark for supercomputing workloads, but Sun will use SPECjbbApp and SPECfpRate to show off the OLTP and number-crunching capabilities of its boxes. What this means is that there is a lot of guesswork coming up with a consistent set of metrics on which to compare the performance and therefore the price/performance of servers. IT managers and vendor salespeople have to do a lot of math on the back of envelopes.

To help make comparisons a bit easier, I have ginned up a set of comparisons for entry Unix machines running the TPC-C OLTP benchmark. (See the bottom of this story for links to these tables.) Wherever possible, I have used actual benchmark results, and where not possible, I have made the best estimates I can for how a particular machine would perform based on past benchmark results, other relative performance metrics out there, and general guidance provided by vendors. In the hypothetical TPC-C comparisons shown in the table accompanying this article, I have included the cost of a reasonable base configuration of a server, including memory, disk drives, a Unix operating system, a version of the Oracle 10g database, and an AIT-3 tape drive for backup since data backup is not an option. The prices shown are for initial acquisition costs at list price, and the bang for the buck is that price divided by the estimated performance. For each vendor and in each of the four general configurations shown (Small, Medium, Large, and Extra Large), I have a configuration from each vendor's product line in July 2004 (which was the last time I did a similar comparison) as well as a machine from the vendor's October 2005 product line. This 2005 machine is in the same general class as the July 2004 machine, but obviously is not necessarily delivering the same performance. Moore's Law moves on (well, most of the time it does), and a small machine today has more oomph than a small machine did 12 to 18 months ago.

While TPC-C is an interesting metric for comparing relative performance as well as relative changes in performance and price/performance over time, it is certainly not a suitable benchmark for all workloads. So this time around, I have added a new twist to the comparisons, and I have ginned up Linpack Fortran benchmark results for the same exact machines. As was the case with the TPC-C performance estimates I have shown in one set of tables, the Linpack numbers are based on actual results where possible and estimated where they are not. I have taken the cost of the Oracle 10g database and tape drives off these machines because when they are used in clusters, they do not have these features. There are a number of different Linpack tests, and the one that I think is most relevant is called Linpack Theoretical Peak Performance (TPP) "Best Effort" test, which is the result in megaflops that a machine posts crunching through a matrix on the order of 1,000 (meaning, it has 1,000 by 1,000 elements). Some vendors use the newer Linpack HPC test, which allows the size of the matrix to scale, but there is far less information available on this test. The Linpack TPP is a bit of a misnomer, since it is not really the peak theoretical performance (which means the clock speed times the number of floating point instructions a chip can execute in each cycle), but rather how close a machine comes in attaining that peak performance while running the Linpack TPP test on the n=1,000 matrix.

For each set of machines (July 2004 and October 2005 in each of the four general Unix server classes), I have calculated the change in performance, the change in price, and the change in price/performance. In the case of performance increases, a positive number is good, and the bigger the better. For price and price/performance--at least as far as customers are concerned--a negative number is good, showing the reduction in the cost of the box and the cost per transaction, whether it is TPC-C transaction per minute (TPM) or Linpack TPP Best Effort megaflops. (Again, the tables showing the performance of IBM, Sun, and HP entry Unix boxes on Linpack benchmarks, both real and estimated, are at the bottom of this story.)

How Entry Unix Servers Stack Up

IBM: Depending on the platforms compared, IBM is offering anywhere from 39 percent to 153 percent more performance running the TPC-C workload on entry Unix servers with its latest Power5 and Power5+ machines, and it offers the most bang for the box among the three Unix vendors in most cases. IBM's performance increases on the Linpack TPP Best Effort workload are less impressive, but all machines saw a big gain, ranging from 86 percent to 115 percent compared to the July 2004 product line. The price/performance gains on the TPC-C workload are not just due to the improved performance of IBM's pSeries Unix iron, but in some cases to the pricing practices of Oracle with its 10g database on dual-core processors.

This summer, Oracle announced a pricing scheme that offers customers Oracle 10g at 25 or 50 percent off, depending on the number of cores and sockets in the box. (To be precise, you multiply the number of active cores running Oracle 10g by 0.75, then round up to the nearest whole number, and then multiple by $4995, $15,000, or $40,000 depending on if you are using Standard Edition One, Standard Edition, or Enterprise Edition.) Prices for the basic hardware as so low for these entry Unix boxes that the price of Oracle can skew the price of the entire box. Ironically, the changes in the Oracle pricing allows IBM to mask hardware price increases for configured machines, which are obviously worth more money because they do a lot more OLTP work in most cases. On the Linpack comparisons, which factor out the database and tape drive costs, each machine showed different price/performance gains, but as a group, they were pretty impressive, ranging from a reduction in TPP megaflops costs of between 35 percent and 63 percent.

Sun: The most interesting aspect of these comparisons is that the Sun Galaxy boxes have more or less caught up to the Power5 and Power5+ entry servers from IBM in terms of bang for the buck--and in some cases, the Sun boxes beat out IBM. Of course, these comments have to be taken with a grain of salt, since they involve real pSeries TPC-C benchmark test results and estimated pSeries results from published relative performance figures for OLTP workloads being compared to estimated Sun Fire benchmarks that are based on very thin data--and not on any current TPC-C tests, either. The fact that Sun obviously knows its machines can perform well on the TPC-C test and yet stubbornly refuses to prove it just shows that Sun has some ways to go to fully remake itself.

By moving to the Opteron processors in the Galaxy machines, Sun has been able to double or triple the OLTP performance of its Sun Fire machines in some cases while cutting server prices by a factor of two to three. And because it can cut back on the number of processor cores in the servers since the Opteron chips are more powerful, core for core, than the UltraSparc-IIIi and UltraSparc-III processors used in entry Sun Fire servers, the costs for database software also go down in many cases. While IBM has a performance advantage compared to Opterons, it is charging a lot more for Power5 servers than Sun is charging for Galaxy boxes. Specifically, in the configurations I examined, IBM is charging anywhere from 25 to 40 percent more money than Sun, and it only does 25 to 40 percent more work, so the price/performance of these two Unix lines are essentially the same. On the Linpack comparisons, the Sun Fire Galaxy boxes certainly didn't do as much work as the Power5 and Power5+ machines, but they offered a significantly lower cost per megaflops when all the math was done. Something on the order of 50 to 65 percent lower costs per megaflops.

I checked the numbers three times because I didn't believe my eyes, either.

HP: To its credit, HP has been able to make do with the minimal enhancements in the Itanium processor line in the past year and rejigger its pricing to show at least some performance and price/performance gains. And, because the Itanium processor has lots of performance, even the current HP Integrity product line can and does go toe-to-toe with the IBM pSeries and Sun Galaxy server lines on OLTP workloads and even pretty much ties with IBM on the Linpack comparisons, besting Sun. But in the bang for the buck department, HP-UX pricing is way out of whack with Solaris 10 (which is free) and AIX 5 (which costs a few hundred bucks on these machines). HP list price for HP-UX 11i Enterprise Edition is $6,105 per processor. That doesn't make sense, but that is the price. While HP's pricing was at a slight premium compared to IBM in July 2004, it is way off the mark (again, at list pricing) compared to IBM in October 2005. IBM has cut server costs in half by reducing prices and substantially boosting performance. HP really needed Montecito to do the same thing. And Sun's Galaxies are offering the same value as IBM's boxes on OLTP workloads and even better bang for the flops on number-crunching workloads.

Those customers who have HP-UX workloads are obviously not going to be happy to hear this. No one wants to hear that they are now paying double or triple per unit of work than a competitive machine. And about the only thing HP can do is appease those customers who complain with deep discounts on the iron and try to convince those customers who can wait until mid-2006 for Montecito to do so.

RELATED TABLES

Comparison of Entry IBM Unix Servers, OLTP Workload

Comparison of Entry Sun Unix Servers, OLTP Workload

Comparison of Entry HP Unix Servers, OLTP Workload

Comparison of Entry IBM Unix Servers, HPC Workload

Comparison of Entry Sun Unix Servers, HPC Workload

Comparison of Entry HP Unix Servers, HPC Workload