The i5 515 and 525 Versus the Unix Competition
June 18, 2007 Timothy Prickett Morgan
Last week, I walked through a performance and price/performance comparison that pitted the new user-priced i5 515 and 525 System i servers against various Windows-X64 servers. Windows, as we all know, is the main competition in the small and medium business space when it comes to server platforms. Windows, of course, only runs on X64 and Itanium platforms–even though it was originally intended to run on IBM, MIPS, and NEC RISC processors. These alternative RISC platforms spawned the Unix revolution nearly two decades ago, and while Windows may dominate the SMB space, there are plenty of Unix shops among these smaller businesses.
Comparing the new i5 515 and 525 servers running a Java workload like the SPECjbb2005 benchmark to Unix servers is not just useful to get a sense of how the machines stack up. It will illustrate that IBM is once again trying to compete in a Windows world with Unix prices when it comes to the System i line. I say will and not does since I have not yet built the comparison table with prices for configured systems and their performance on the SPECjbb2005 test. But based on the nearly two decades of price/performance comparisons I have done putting AS/400, iSeries, and System i5 servers against other platforms on various workloads, I am pretty sure that I will see pricing parity of some sort here, with perhaps some outliers where Unix pricing is becoming more aggressive.
It is kind of exciting to write the story this way, at least for me, since even I do not know what will happen yet. Perhaps we will all be pleasantly surprised, and the i5 515 and 525 will kick the tar out of similarly powered Unix boxes . . . .
While I am building the comparison table, let me remind you of what these comparisons are all about. The SPECjbb2005 benchmark, which is administered by the Standard Performance Evaluation Corporation, is in essence a Java version of the Transaction Processing Council‘s TPC-C online transaction processing test with the substantial I/O requirements of the TPC-C test removed. Basically, it converts a system-level benchmark test into one that stresses the processors and main memory–which is exactly what modern computers, with their big gobs of main memory, have become.
When IBM tested the i5 515 back in March on the SPECjbb2005 test–just prior to the release of the user-priced machines, in fact–it only tested the box with one of the 1.9 GHz Power5+ cores in the server activated. The single-socket i5 515 can have two cores activated, and the i5 525 can have one or two cores as well. The i5 525 is a larger configuration of the user-priced machines that can have lots more users hanging off of it but which does not necessarily deliver more performance. The i5 525 is a more expandable box, but the most important characteristic is that it can support more than 40 users–the cap on the i5 515–and can, in fact, be converted into a box that has no limits on the users attached to it. As I said last week, IBM did test a System p5 520 running AIX with exactly the same hardware as the i5 525, and in the comparisons I make, I am using this p5 performance on the SPECjbb2005 test to reckon what it might be on the i5 525 with two cores activated. The i5 could do better or worse than this, of course, depending on the tweaks in the operating system and Java Virtual Machine (JVM), but this is a ballpark estimate that I feel pretty comfortable with.
So, to compare the i5 515 and 525 to Unix servers, I went to the SPEC site and gathered up performance information for entry servers that are roughly in the same power class as the new System i machines. You can see the SPEC results I distilled down in this table. The table includes not only the results for the i5 515 and my estimates for the i5 525, but also for a mix of Windows, Unix, and Linux boxes. (This is the same table I ran last week.) This week, I only plan to cover comparisons with Unix platforms. (See last week’s story for the Windows comparisons.) In future editions of this newsletter, I will cover how these user-priced machines compare to Linux boxes, as well as to prior generations of System i servers.
As I said last week, I know the limitations of benchmark tests. This is only one test, and I know that IBM does not want people to price the i5 515 and 525 based on performance at all, but rather on user count. But a user count without some kind of reckoning of the performance available to end users is only half the story. And even after making the comparisons that I could–given the thinness of data for the SPECjbb2005 test–I wish there was more data. To keep it relatively simple, I am examining Unix boxes from Hewlett-Packard, Sun Microsystems, and IBM, providing the actual performance on the SPECjbb2005 test or an estimate. (The performance estimates are marked in bold, italic, red.) SPECjbb2005 gauges performance in terms of business operations per second, or BOPS.
As you can see from the first table, there is no shortage of Java horsepower in the entry and midrange Unix boxes that have been tested on the SPECjbb2005 test. IBM’s System p5 line (all based on the Power5+ chips) spans a decent range, and so does Sun’s entry “Niagara” Sparc T1 and “Galaxy” Opteron series of servers. IBM’s entry Unix boxes have plenty of memory and disk scalability, and so do HP’s entry Integrity boxes, which are based on the dual-core Itanium 9000 chip from Intel. Sun’s T1000/T2000 and X2100/X4200 machines have lots of processing oomph (for Java workloads, at least) and have lots of memory expansion, but they come up a little thin on the disk drives. Moreover, in the heavy configurations of the T2000 Sparc boxes, you have to use very expensive 4 GB DIMM memory to hit the 64 GB memory configuration, which is prohibitively expensive. (You can see the price and performance comparisons in the second table I created comparing the i5 515 and 525 to entry Unix boxes.)
The table illustrates how the pricing on Unix boxes is very aggressive–with the possible exception of HP’s Integrity boxes. I am annoyed at HP right now, since the company is not publishing list prices for the Integrity boxes and has not published pricing data for its entry machines in public benchmarks, either. So heaven only knows what these things cost. And on a Friday afternoon after a hard week of cranking it out, I am just not in the mood for vendor shenanigans. So, HP, a word of advice: stop hiding whatever you are charging–or trying to charge–for these machines. It ain’t gonna work. It’s a bad marketing strategy. In any event, I am happy to show the performance of the entry and midrange Integrity boxes on the SPECjbb2005 test, but as for price/performance, HP is in the doghouse.
Sun is also heading towards the doghouse, since it has not tested its X2100 and X4200 servers using the dual-core Rev F Opterons and its Web site says it has pricing for the older Rev E boxes that have been tested, but in actuality, that part of its site isn’t working. If I had data on the newer Opteron-based Galaxy servers from Sun, I would have run it an gladly used the pricing. Which would show even better bang for the buck than the machines in my table. With the quad-core “Barcelona” Opteron Rev F chips due any week now, Sun would be wise to show some great SPECjbb2005 numbers with these machines.
The one thing to consider as you look at the comparative prices between the i5 user-priced machines and the IBM and Sun Unix boxes is that the cost of the database management system, even on a per user basis, quickly overwhelms the cost of the operating system and even the hardware. IBM is charging a mere $150 per core for AIX 5.4 and Sun bundles it for free on its own servers. And because Oracle 10g R2 Standard Edition One, which is the version that runs on two-socket servers as well as the Sparc T1-based boxes from Sun, costs only $149 per user, the software stack on the Unix boxes I examined is very inexpensive. To the point that Unix machines are 55 percent to 75 percent less expensive than the i5 515 and 525 servers on a cost per BOPS basis and from 15 percent to 65 percent less expensive on a cost per user basis. (As was the case with Windows boxes last week, I cannot keep the amount of processing capacity per user constant using the benchmark data and server configurations that have been tested. But boy, it sure would be nice to be able to do that.)
Only by shifting the Unix machines to the Standard Edition of Oracle 10g can the cost of the Unix machines be brought more in line with the user-priced i5 machines. That’s because Oracle 10g Standard Edition, which runs on machines with up to four processor sockets, costs $300 per user. When you change this one factor in the comparisons, it really closes up the gap between the i5 and Unix machines. But not completely.
Which brings me to the point I made about Windows platforms and the Standard and Enterprise Editions of Microsoft‘s SQL Server 2005 database. If IBM wants to attack the SMB market, it needs to have a much cheaper edition of i5/OS and DB2/400. The hardware is not the big problem–excepting the exorbitant prices IBM charges for memory and disk drives. We need i5/OS Entry Edition and DB2/400 Standard Edition One–and we need it fast. And, it would not hurt to have Power6 chips in entry servers, too. That would close up the performance gap with quad-core X64 boxes, Sun’s eight-core Sparc T1 boxes, and dual-core Itanium boxes rather nicely. Waiting until next year to get faster chips in the i5 line is a bad idea, even if the shift to user-based pricing does cushion the blow a little.
In the end, I was not pleasantly surprised with the user-priced i5 and Unix comparisons. Next week, Linux boxes and open source databases and how they compare.