Oracle-Sun Exadata V2, Meet iDatabase V1
September 21, 2009 Timothy Prickett Morgan
This will be a recurring theme for me in the next couple of years as I watch Oracle and its soon-to-be server, storage, and operating system division (formerly known as Sun Microsystems) preach to us about the benefits of an application development system with a consistent and tightly integrated software stack tuned very closely to iron. The recurring theme will be that I get really annoyed.
As The Four Hundred told you last summer, Hewlett-Packard and Oracle teamed up to create something called the HP Oracle Database Machine, one of Oracle’s several forays into the hardware business, based on a clustered implementation of the Oracle 11g database with Real Application Cluster (RAC) extensions and a clustered storage array and software called Exadata. Fast forward a year to September 2009, and the European Union’s antitrust regulators were supposed to have long-since approved the $7.4 billion Sun acquisition, but have stalled it because of concerns with Oracle getting its hands on the open source MySQL database, which Sun bought nearly two years ago for $1 billion.
With the Sun deal stalled, Oracle and Sun have to give Sun’s customers some hope that it won’t kill off the Sun server business. And to that end the next generation of Oracle’s database machine, now called Exadata V2 by Larry Ellison, Oracle’s co-founder and chief executive officer, excises HP from the name. The product is based on Sun iron, including new flash memory to boost database throughput and tweaks to the Oracle database stack and the Sun storage stack that allow the Exadata box to not only support data warehousing and its sequential reads, but also online transaction processing and its random reads and writes.
If you want to get the play-by-play of the Oracle Exadata V2 launch, check out my coverage over at The Register at this link. I am not going to go through the whole launch event again, because that is not what I am interested in ranting about in The Four Hundred. What I am interested in reminding everyone is that what Oracle says it can do with Exadata V2–data warehousing or OLTP processing on a clustered machine–and what Ellison claimed no other database maker or server maker could do, IBM can, in fact, do with the OS/400 platform. And has been able to do for 13 years.
A brief rundown on the Exadata V2 architect and the performance before I get into what I will call the iDatabase clustered appliance.
The Exadata V2 setup includes Sun’s X4170 servers acting as database servers, running Linux and Oracle 11g plus RAC. These 1U rack servers have two of Intel’s new quad-core Xeon 5500 processors, 72 GB of memory, and a bunch of disks. The storage servers that feed data into these database servers are based on Sun’s X4275 servers, also using two X5500 processors and sporting enough room for a dozen disk drives in a 2U rack chassis. Each storage server is equipped with four flash disk modules, each weighing in at 96 GB and apparently plugging into PCI-Express peripheral slots. (Sun and Oracle tried to make a big deal about this, calling it “FlashFire technology.” But i 6.1 is already flash aware, as you know, and can move data in the single-level store for memory and disk into flash of that will boost performance.) The database and storage servers are linked together using Sun’s 40 Gbit/sec InfiniBand switches. A single rack has eight database servers and 14 storage arrays, for a total of 576 GB of main memory on the database servers and 5 TB of flash memory and 100 TB of disk capacity on the storage servers.
The most important thing is that such a rack of machines, with the fast InfiniBand links and the tweaks for moving data between disk, flash, and main memory, can deliver over 1 million random I/O operations per second, which means it can do OLTP. And because of database compression, a 4 TB database can be held in main memory and a 15 TB database can be held in flash–no need to even go to disk drives. And if customers need to expand their database processing, they can add up to eight racks of these Exadata setups together without having to resort to monster switches. A fully loaded Exdata rack sells for $1.15 million.
Taking aim at IBM’s top-end Power 595 box, Ellison said that two racks of the Exadata V2 boxes would have about the same OLTP performance (presumably on the TPC-C online transaction processing test), but costs one-quarter of the $10 million price tag on the big, bad Power box. And, quite correctly, Ellison pointed out that the Exadata setup (both V1 and V2) had fault tolerance built in, while big SMP servers like the Power 595 do not. He neglected to mention that managing clustered databases under RAC is not as easy as managing databases for an SMP server (which is why most companies are not using RAC).
OK, here’s the rant, and here’s what I want IBM to do. First, I want it to remember that it has DB2 Multisystem and DB2 SMP features for the DB2 for i database integrated into all V4 and V5 releases (and V3R6 and V3R7, if you remember). I wrote this four years ago, when Oracle first started bragging about RAC (which uses clustering technology ripped out of Digital minicomputer clustering that was sold by a desperate Compaq before HP ate it), and it still holds true:
“While DB2 Multisystem allows database tables for applications to be spread across multiple physical machines that have been clustered (using OptiConnect or some other fabric, like InfiniBand, which is due in the future Power6 servers), DB2 SMP allows an SQL query is used for parallelizing queries across the processors in an SMP cluster. DB2 SMP is a tightly coupled, vertical scaling feature, while DB2 Multisystem is a loosely coupled, horizontal scaling technique. DB2 Multisystem is, as far as applications are concerned, utterly transparent. This is important.”
So there is the secret sauce for database clustering. Now, I want IBM to take the iron behind the Smart Analytics System it announced last month based on Power 550 servers, dump a whole bunch of those DS5300 disk arrays and put real and fast flash drives inside some of the Power 550 servers and use them as storage servers, much as Oracle and Sun are doing with the Exadata V2 setup. Forget the Fibre Channel switches to storage and use 40 Gb/sec InfiniBand switches, and why not go with Voltaire, one of IBM’s partners and the one that was dissed by Oracle (which picked Voltaire switches for Exadata V1).
A Power 550 using 5 GHz Power6+ processors can house up to eight cores, and is rated at 37,950 CPWs running OLTP workloads with the i 6.1 operating system. That’s about 375,000 transactions per minute (TPM) per box. (And about half of what an AIX machine actually running the TPC-C test will show because the AIX team does some marginally iffy SQL tuning to radically boost the performance on the TPC-C test.) Oracle is saying it can deliver about 3 million TPM per rack on OLTP workloads, so we need to get eight of the Power 550s into a rack to compete; add another two to cover whatever overhead DB2 Multisystem might take. Or use the Power 560, which crams as many as 16 3.6 GHz Power6 cores into the same 4U chassis and delivers 48,500 CPWs or about 482,600 TPM on the TPC-C test, by my estimates.
Sun has no advantage when it comes to pricing for memory, disks, or flash. Solaris is a hell of a lot less expensive than i 6.1, so IBM would have to significantly reduce the price for i 6.1 and DB2 Multisystem this iDatabase machine I am proposing. Get over it, IBM.
Let’s keep it simple. Use the Power 560 and put six of them in a rack as database servers. Use a bunch of Power 520s as storage arrays. Put in six 450 GB SAS disks and three flash drives of some sort that give the setup roughly the same 384 GB of flash to play with as the Exadata V2 box has for each of its storage servers. Use the 12X I/O drawers to hang lots of disks and flash drives off a relatively small number of Power 520 storage servers, and use the fastest InfiniBand switches from Voltaire to glue the storage and database servers together. And finally, charge $1.15 million for the whole shebang.
And next time, don’t let Larry set the pace two years in a row.