Ellison rides SPARC T4 SuperCluster into data centers
Four star general purpose, sir!
Just putting out four new entry and midrange servers based on its new eight-core SPARC T4 processor is not sufficient to get SPARC/Solaris customers fired up about buying gear from and paying system maintenance to Oracle. Every CIO wants to know – in fact needs to know – that there is headroom in their systems in case their workloads grow.
Oracle resells Fujitsu's SPARC64-VII+ machines, badged as the SPARC Enterprise M machines, for customers who have big jobs that require a shared memory system. But over the past two decades, Oracle co-founder and CEO Larry Ellison has made no secret of the fact that he believes that computing in future will be parallel, spreading data and database crunching across multiple compute nodes, instead of trying to create ever-larger shared memory systems to hold databases.
This is by no means an exhaustive list, but Oracle Parallel Server was tuned for IBM's RS/6000 PowerParallel machines back in the late 1990s, and then Oracle was best buddies with Dell, pushing the initial Real Application Server (RAC) clusters.
Three years ago, Oracle partnered with Hewlett-Packard to create the first generation of Exadata parallel database appliance – sometimes called the HP Oracle Database Machine – which ran Oracle 11g. It has now put out x86-based Exadata database appliances – with columnar hybrid data compression on Exadata storage arrays – that actually pre-chew SQL queries on compressed and encrypted data on the parallel file system. After this, the appliances pass partial answers up to the database engines (in compressed form over a 20Gb/sec InfiniBand network, so it is very fast). The combination of compression and parallel processing on both the storage and database nodes made the Exadata machines pretty fast.
The SPARC SuperCluster T4
Then Oracle shelled out $7.4bn to buy Sun Microsystems, and kicked HP to the curb. The Exadata V2 database appliances, announced in September 2009, were retooled with Sun Fire x86 servers and peppered with flash storage to radically improve the performance of the storage subsystems while the whole shebang was beefed up with a speedier 40Gb/sec InfiniBand network. Last September, Oracle updated the Exadata line with fatter Xeon 7500 nodes and also gave the SPARC architecture some love with the original SPARC SuperCluster T3-2, which launched in December 2010. Now, SPARC-based systems were invited to the parallel party. Sun also put out x86-based clusters, called Exalogic, for running parallelized application servers.
Which brings us to this year's SPARC SuperCluster T4, announced yesterday evening in San Francisco by Ellison and John Fowler, executive vice president of hardware engineering at Oracle. Well, almost.
Ellison lectured considerably on the benefits of parallelism and data compression for database processing, and talked quite a bit about the Exadata machines, of which Oracle has sold 1,000 machines thus far – Oracle's "most successful product ever," he claimed – and plans to sell an additional 3,000 machines before the end of the year. (It is not clear if Oracle meant calendar or fiscal year there.)
"We're a lot faster than IBM's biggest pSeries machine," Ellison proclaimed, comparing a cluster of x86 servers running the 11g database and the Exadata storage software on an InfiniBand backbone to a wonking 256-core Power 795 SMP server. Here's how he stacked the two machines up, fully loaded:
Eight Exadata X2-2 racks versus one IBM Power 795 and four DS8700 arrays
The comparison is rigged in a bunch of ways, but it is representative of the two different types of computing that go on inside of data centers for database processing. Thanks to the flash and data compression, the Exadata has a factor of 10 times the storage I/O operations per second running the Oracle 11g database than the IBM machine, according to Ellison, and that is something that Oracle absolutely controls. Do you think Oracle wants to run 11g better on IBM hardware? Do you think IBM's or HP's or Dell's flash storage will get unique tuning ahead of Oracle's own flash storage? Or even within six months?
This is the "engineered systems" game that Oracle will be playing. Ellison said that two racks of Exadata could do queries anywhere from 10 to 50 times faster than the Power 795/DS8700 combo, with 4 to 10 times the OLTP throughput and with 10 times the amount of storage (with compression turned on) – and do so for a cost of $3.3m, compared to $18.86m for the IBM hardware. "The Exadata system costs way less than a memory upgrade on the IBM pSeries, and you have to be willing to run a lot faster," Ellison quipped. "The P795 is one big, expensive single point of failure," he added, pointing out that Oracle RAC was inherently fault-tolerant.
Ellison didn't just take shots at IBM. He took one at data warehousing giant Teradata too, pointing out that Japanese telco and communications company Softbank replaced their 36 racks of Teradata machines with three racks of Exadata and ran their queries eight times faster.
"Teradata really appreciated that, so we mention this on the front page of the Wall Street Journal every week," Ellison said, getting the audience to chuckle.