The eServer i5 Versus Linux Servers
by Timothy Prickett Morgan
There are two great things about IBM's AS/400 and RS/6000 midrange servers: they have very good hardware and software technology, and Big Blue goes out of its way to demonstrate the performance of its machines with internal and public benchmark tests. Sometimes IBM may not offer the best deal in town, but you can always qualify and quantify the deal. This is not the case with Linux servers running real commercial workloads, however. You move to Linux, you are on your own.
Or, more precisely, you have to rely on people like me to make estimates of Linux server performance (or make them yourself) if you want to compare and contrast the value Linux servers offer against that of the new eServer i5 machines.
I follow benchmark test results like many people follow sports scores. And I can tell you that, with the exception of two Linux-based TPC-C benchmark tests on Itanium 2 machines and one SAP SD benchmark on a dual-processor Xeon blade server, there is not a lot of data about Linux performance. We are in uncharted territory. But there are some stars to guide us. Hewlett-Packard has run a TPC-C test on a cluster of 16 four-way rx5670 servers with four 1.5 GHz Itanium 2 processors, which ran Oracle's Oracle10g database on Red Hat's Enterprise Linux AS. Late last year, NEC ran a benchmark test on its 32-way Express5800 Itanium 2 servers running Microsoft's Windows 2003 Datacenter Edition operating system and its SQL Server 2000 database and then again running Novell's future SuSE Linux Enterprise Server 9 and Oracle's 10g database. And as I told you last week, Fujitsu-Siemens has run the SAP SD benchmark on SuSE 8 on a two-way Xeon blade server.
These are not a lot of stars to steer by, to be sure. Still, customers evaluating servers want to have at least a rough idea of where Linux will perform and what it will cost. In the estimates I have put together for this week's comparison of the eServer i5 Model 520 and 570 servers against Lintel machines, I had to guess what the performance of Oracle10g would be running the TPC-C benchmark test. Right now, Oracle has the lion's share of commercial database sales on the Linux platform, so choosing to put Oracle 10g on Red Hat (also the volume Linux operating system leader) on HP ProLiant servers (the volume Linux server maker) seems obvious enough. And guessing the performance is not all that tough: it should be about the same as a Wintel machine, plus or minus 10 percent. I decided to go conservative and lop 10 percent off the performance of a Wintel machine, just because there is simply not as much experience in the market tuning Linux as there is in tuning Windows or OS/400, for that matter.
But just to make things interesting, I decided to throw in a second comparison, with MySQL's MaxDB database, a product of joint development between SAP and MySQL. MySQL is a privately held Swedish company that sells commercial and open source versions of its eponymous database, and it is wickedly popular on Linux. About 10 million copies of MySQL are downloaded a year, and the company estimates that there are about 5 million active installations. MySQL runs on 20 different platforms, not just Linux machines, and is often hailed as an alternative to Oracle's eponymous databases, Microsoft's SQL Server, and IBM's DB2. But these databases are more familiar to corporate customers, and they have many features that MySQL is lacking. MaxDB is another open source database that is coded and sold by MySQL, but it is based on the Adabas mainframe database, from German software company Software AG. I think this is a more appropriate database to compare with OS/400's DB2/400, particularly since it can actually run SAP's R/3 and mySAP software suites, which MySQL cannot do. But there is no performance information for MaxDB. Zero. I'm making a wild guess, people, for the sake of comparison, of course.
The reason is this: if you toss in commercial MaxDB licensing costs ($1,490 per CPU) and support costs ($3,000 per CPU for entry support, $9,000 per CPU for premium support), and if you assume that MaxDB performance is a little lower than Oracle 10g on the same Lintel box, then Oracle10g offers about the same price/performance as that solution based on open source software. However, if you want to get the lowest-cost solution in the market and want to learn how to support your own database, spend a few grand for a Lintel box and download MaxDB for free, and you can build a serious midrange machine with lots of functionality.
I'm not suggesting that anyone should do this, of course.
As you can see from the eServer i5 to Lintel server comparison table I built, the two i5 Model 520 Express configurations are a bit pricey when compared with 32-bit Lintel iron. Lintel iron running Oracle10g Standard Edition One (the new Oracle version for two-way servers that costs $4,995 per CPU) and Red Hat Enterprise Server 3 with Standard Edition support ($799 for a license) offer excellent bang for the buck. A Linux machine with MySQL MaxDB 7.5 with entry support costs under $10,000, but if my guess about performance is right (and it may not be), this machine would offer slightly worse value for dollar. The Model 520 Express configurations, which have crimped 5250 performance and geared-down raw performance, do not fair as well in the comparisons with these Linux machines as they would if IBM let the Model 520 run full-out in the Express models.
As I explained last week, the value that logical partitioning (now part of the hardware) and OS/400's workload manager (which allows many different kinds of applications to share a single server) can be easily and correctly used to justify the higher acquisition costs of i5 iron. My comparisons assume that all machines are running OLTP workloads against relational databases with all of their computing capacity dedicated to that task. But in the real world, Windows and Linux servers rarely get much above 25 percent of peak CPU capacity, while OS/400 servers can get into the 60 to 70 percent range. If you assume that you can load up an i5 with twice as much real work, the Model 520 Express configurations are in spitting distance with even cheap Lintel iron, which has very inexpensive open source operating systems and databases that nonetheless need to have support added on.
As you move up from these entry machines through the comparisons I have created with various performance bands (25,000 TPM, 50,000 TPM, and 100,000 TPM), the spread between an i5 server and a Lintel box gets very wide, much wider than with Windows machines. Two-way X86 servers running Linux and either Oracle10g or MaxDB cost roughly between $.50 and $.75 per TPM, depending on the speed of the Xeon processor and server configuration to support it. A two-way i5 running i5/OS V5R3 Standard Edition costs around $2 per TPM, and with V5R3 Enterprise Edition, it costs around $5 to $6 per TPM. That is a very large spread. And I think this spread is an excellent gauge of exactly how unknown Linux is when it comes to supporting OLTP applications.
With that said, however, Oracle and MySQL support their databases on a wide variety of platforms, so it is not really a Linux issue; their aggression on pricing is nothing short of remarkable. Oracle's and MySQL's pricing for commercial products is the result of the pressure that all open source software is having on the market. However, IBM has a large, captive installed base of OS/400 users, and as long as it shows 30 to 40 percent price performance improvement in each product refresh, OS/400 shops will grumble about such comparisons, but they will find ways to justify staying on the OS/400 platform because of its many benefits.
For four-way machines, the Lintel boxes I configured have about a 2:1 price/performance advantage compared with an i5 Model 570 configured with i5/OS Standard Edition, and have a nearly 5:1 advantage compared with i5/OS Enterprise Edition. The cost of Oracle10g Standard Edition (which is for four-way machines and is not to be confused with Oracle10g Standard Edition One) on this box is $15,000 per processor; support for MaxDB 7.5 is $9,000 per processor, plus the $1,490 per processor for the license. The Lintel iron is a little cheaper than Windows boxes, but not all that different. However, if my low-ball estimates on performance turn out to be wrong (and they very well could, and I think they will be over the long haul with the Linux 2.6 kernel), then Linux servers could offer a 15 to 30 percent improvement over Wintel machines in the same power class as the i5s.
You can see now why IBM wants to co-opt Linux and Windows and bring those environments underneath the skins of the OS/400 platform. It can't compete directly with cheap Wintel and Lintel iron, but that cheap iron is used very inefficiently. An efficient OS/400 server with Linux partitions and Windows-based Integrated xSeries Server (IxS) co-processor cards, all running at near peak performance, should offer bang for the buck that is roughly equivalent to a rack of Wintel and Lintel machines with different workloads isolated on different machines running the same kinds of applications.
What IBM needs to do in order to sell more i5s is to actually prove this with benchmarks. The so-called "three-in-one" benchmark IBM, announced last year, was a step in the right direction. But IBM needs to show the value of such server consolidation in terms of upfront acquisition costs as well as ongoing administration and support costs. And it needs to do so before virtual machine partitioning becomes a default item in Windows and Linux servers, thereby negating one of the i5's key advantages.
"The eServer i5 Versus Unix Servers"