Oracle Gets Systems Design, and Starts Proving It

September 27, 2010 Timothy Prickett Morgan

Oracle is ratcheting up its so-called “engineered systems” to run its operating system, middleware, and application stack, and is beginning to give people the impression that it is dead serious about being in the hardware business. Oracle did not come easily to hardware, being a software vendor, but thinks it can get hardware gross margins up to 60 percent and double system revenues to around $2 billion a quarter over the next couple of years. And completely engineered systems, from chip to app, are the key.

Last week at its OpenWorld extravaganza in San Francisco, the top brass at the company, including new co-president Mark Hurd, formerly president, chief executive officer and chairman of Hewlett-Packard, introduced a long-awaited upgrade to the Sparc-based Sun Fire midrange server lineup as well as two new clusters for running databases and middleware. Perhaps more telling was the fact that these new clusters will be highly tuned to run a forthcoming Solaris 11 Unix release and what is arguably a fork of the Red Hat Enterprise Linux (RHEL) variant of Linux. Oracle is no longer content with just using RHEL’s kernel and selling cheaper support contracts for a rebranded version of RHEL, which the company calls Oracle Enterprise Linux. Now, Oracle is going to start crafting its own Linux kernel, called the Oracle Unbreakable Linux Kernel, of course, to give its middleware, databases, and applications performance advantages that will not be available on RHEL or other Linuxes. And which will presumably also be worked into Solaris 11.

This is what you do when you have control of the complete stack, as IBM knows full well from its experience with mainframes and AS/400s–a lesson it has yet to fully apply with its AIX Unix and Linux partnerships with Red Hat and Novell.

Ed Screven, who has the vague title of chief corporate architect at Oracle and who is in charge of all of the company’s open source efforts, said that in the first iteration of the Oracle Linux kernel, which is based on Red Hat’s current RHEL 5.5 kernel but no doubt has lots of performance enhancements that are due later this year in RHEL 6, the performance gains were substantial. Running an Oracle software stack on an eight-socket server based on Intel‘s eight-core Xeon 7500 processors, OEL using the RHEL kernel was able to do 197,000 cache reads per second (with 8 KB file sizes) and deliver 4 GB/sec of I/O bandwidth using solid state disks. With the Oracle Linux kernel, Oracle’s engineers were able to push up to 1 million IOPs and 9.5 GB/sec of bandwidth on the same flash storage, which is a quintupling of IOPs and more than a factor of two better oomph in terms of raw I/O bandwidth.

The InfiniBand drivers borrowed from the Open Fabrics Alliance and underpinning Oracle’s Real Application Clustering do a lot better, too. The InfiniBand OEFD drivers allow for direct node-to-node connection in a cluster through the Remote Direct Memory Access (RDMA) protocol that is embedded in InfiniBand, and they support Reliable Data Sockets (RDS) communication between database nodes in the cluster. Both RHEL and OEL support this software, but Oracle’s own kernel now does it better, according to Screven. With the current RHEL kernel, a single InfiniBand card in the eight-socket box that Oracle tested could process 89,000 messages per second, compared to 273,000 with the Oracle kernel. That’s a factor of three improvement.

“So what?” you say. Subsystem performance is not necessarily application performance. But it sure leads to it. Oracle’s tests show that running an unspecified application, on the same eight-socket box, just switching to the Oracle Unbreakable Enterprise Kernel boosted online transaction performance to 3.2 million transactions per minute (TPM) compared to 1.8 million TPM with the RHEL 5.5 kernel. That’s a 75 percent improvement in performance, all through software.

Contrast this with what IBM has done with OS/400 and its successor platforms over the years. The IBM i operating system is more expensive than the AIX-Oracle combination on the same exact Power Systems iron and yields poorer performance, lower thread scalability, and lower overall scalability. As a very funny quote from an IT manager put it when Big Blue was going bust in the early 1990s, in a line I will never forget (even if I get repeated blows to the noggin): You can find better, but you can’t pay more.

So it is an embarrassment to me and to you when IBM controls the entire IBM i stack and charges such a premium for it without giving the kind of performance and ease of use advantages that the AS/400 did indeed have compared to other platforms in the late 1980s and early 1990s. I want to kick someone in the shins really hard when I see IBM i 6.1.1 and 7.1 get short-sheeted on the thread counts, as I explained to you would be the case way back in February before either i 7.1 or AIX 7.1 got here.

Oracle has only 5,000 Linux customers and 35,000 Solaris customers running its database, middleware, and application software today across its customer base of 345,000 users. With those kinds of performance advantages–at the application level–how hard do you think it will be to talk customers into an all-Oracle stack?

The trick of course, will be to sustain the performance advantages. There are two ways to do that, and call me a cynic, but I bet that Oracle will revive an old saying from the Lotus spreadsheet wars two decades ago: “The software ain’t done until WebSphere and DB2 and Windows and SQL Server don’t run.”

So I think Oracle will have real performance advantages in its stack, and also that the hooks to new features won’t necessarily be quickly available to those who sell other platforms. With Oracle not dumping out its Solaris code until after the code is released in production, other operating system providers will be playing catch up to study whatever features Oracle comes up with; ditto for its own Linux kernel tweaks. Red Hat, which has taken a very conservative approach to adding new features to Linux, is going to be in a footrace with Oracle, which, by the way, will only have to certify its own stacks of middleware and apps on its bleeding-edge kernels. Red Hat has to certify thousands of applications and coordinate with other development teams outside of its own walls, all of them who probably support more than one platform and have their own development headaches.

Who is going to move faster and with more sure footing?

Sometimes, you have to hand it to Larry Ellison, Oracle’s co-founder and CEO, as much as I hate to admit it. He got lucky in partnering with Bob Miner, who actually created the Oracle database and commercialized the idea invented by IBM ahead of IBM. And Ellison has made many smart moves in the three decades he has been at this. But the smartest one yet might be to not go with his initial reaction to Sun, which was gut it for Solaris and Java and sell off the iron. Someone somewhere made Ellison see that with the whole hardware-software stack, he could upset the whole applecart of IT and answer to no one. (Cut to sea spray splashing on Captain Ellison, cannons roaring and spitting fire. . . .)

So the Exadata X2-8 and Exalogic clustered platforms announced last week got all the headlines, but the complete picture has a lot higher resolution than many people are giving Oracle credit for.

These clustered machines are exactly the kind of iron I have been saying since 1995 that IBM should be bringing to market for SMB and enterprise shops. Clusters are cheaper and can be rigged to be as reliable as two-node clusters of big SMP systems.

The Exadata X2-8 is a version of the Exadata V2 data warehousing and OLTP appliance that Oracle announced with then-independent Sun Microsystems last September at the OpenWorld extravaganza. But this one uses the eight-core Xeon 7500s in a pair of eight-socket servers clustered together and fed by flash-backed storage nodes running Oracle’s Exadata compression and query optimization software. Specifically, the database portion of the Exadata X2-8 is based on two Sun Fire X4800 servers, each with eight 2.26 GHz Xeon X7560 processors and 1 TB of main memory and eight InfiniBand host bus adapter cards. The rack has three of the Sun Switch 36 InfiniBand switches, which have 36 ports and which run at 40 Gb/sec speeds. The storage side of the rack has 14 rack-based, two-socket servers–the Sun Fire X4270 M2–which are each equipped with two six-core, low-voltage Xeon L5640 processors running at the same 2.26 GHz speed. It is not clear how much memory these storage nodes have, but they have 5.3 TB of flash storage (based on Sun’s own PCI-Express flash module, with four 96 GB flash units in each server) and the option of putting in a dozen 600 GB 15K RPM or a dozen 2 TB 10K RPM disks per server.

With the skinnier and faster SAS drives and all of that flash, Oracle can deliver 1 million IOPS and 50 GB/sec of uncompressed bandwidth out of the flash into the database servers and the disks can do 50,000 IOPS and deliver 25 GB/sec of bandwidth. With these skinny disks, you are talking 100 TB of raw disk capacity for OLTP or data warehouse applications and a load rate of 5 TB per hour. If you need more than this, the Switch 36 modules in up to eight racks can be daisy chained together.

The Exalogic Elastic Cloud is a funky name for a cluster of machines designed to run a highly tuned version of Oracle’s WebLogic application server and its JRockit and HotSpot Java virtual machine. The Exalogic machine is comprised of 30 two-socket Sun Fire servers, for a total of 360 cores. 2.8 TB of main memory, 960 GB of flash storage, and 40 TB of disk capacity per rack. The whole shebang is glued together using InfiniBand switches. One rack, says Oracle, can process over 1 million Web requests per second, a factor of 12 higher than standard, untuned, and unintegrated application servers. As for Java, Oracle is boasting a rack can deliver 1.8 million messages per second, which is 4.5 times better than the competition can deliver in the same rack running the same Fusion middleware–and presumably without the benefit of flash storage.

Obviously, the new Exadata X2-8 and Exalogic clustered appliances are running the new Oracle homegrown kernel, and they will also, Oracle said, run the development release of the future Solaris 11, which is called Solaris 11 Express Edition. This developer version is due to be released at the end of the year, with Solaris 11 going into production in 2011, probably in the second half of the year but maybe earlier.

As for pricing on all of these, Oracle did not say. As is its way.

This Issue Sponsored By

Table of Contents

Content archive

Recent Posts

Subscribe

Pages

Search