Oracle cranks up SuperCluster with Sparc T5 engines
Anything Big Blue can do, Big Red can do cheaper – and better
Software giant and hardware playa Oracle has launched its high-end SuperCluster T5-8 "engineered system" based on its sixteen-core Sparc T5 processor.
It seems that turning in a less-than-stellar quarter ending its fiscal 2013 year inspired Oracle to make as many announcements as possible in the hope of some positive press. Earlier this week was all about Big Red's cloudy partnerships with Microsoft, Salesforce.com, and NetSuite, and now it's the SuperCluster T5-8.
Oracle made its SuperCluster T5-8 announcement in a Thursday webcast, calling the new machines its "fastest engineered system," meaning presumably that it offers more throughput and response time than Xeon-based Exadata database and Exalogic middleware clusters tuned up to do the same work. Oracle did not do any comparisons between the new Sparc-based SuperCluster and its current generation of Exa clusters – which would be interesting, indeed – but as usual it did pick on IBM's Power Systems-AIX combo.
David Lawler, an ex-Sunner who did a stint heading up server marketing at Cisco and who is now back at Oracle as senior vice president of product management and strategy for all of Oracle's hardware, opened up the announcement by saying that Oracle sold over 1,200 engineered systems (of all kinds, not SuperClusters, but all the Exa boxes, too) in the fourth quarter, and that it has sold more than 3,000 systems in its fiscal 2013 ended in May.
Revenues for Oracle's engineered systems were up 45 per cent, year on year, which is great, but overall hardware sales continue to slide as the company pulls back on peddling onesies and twosies of x86 servers and even Sparc servers to a certain extent. In the fourth quarter, Oracle's overall hardware systems business fell 13 per cent to $849m, and support revenues fell 3 per cent to $582m.
But Oracle is hoping to turn the corner in the coming quarters, and the SuperCluster T5-8 announced on Thursday is part of that recovery in system sales.
The Sparc T5 processormade its debut back in March, based on the "S3" core that was used on the Sparc T4 processors. There are two out-of-order integer execution pipelines and one floating point unit on the T5's S3 core, and each core has eight processing threads, for a total of 128 threads per socket. The Sparc T5 chips have twice as many cores as the Sparc T4s, and the clock speed is ramped up by 20 per cent to 3.6GHz.
The T5 chip has 16KB of L1 data cache, 16KB of L1 instruction cache, and 128KB of L2 cache for each of its sixteen cores, plus an 8MB L3 cache shared by the sixteen cores on the die. The T5 has two PCI-Express 3.0 controllers on the die plus four memory controllers that can drive up to sixteen DDR3 memory sticks running at 1.07GHz. There are also cryptographic and encryption accelerators.
The Sparc T5 chips were plunked into a single-socket Sun Blade blade server and in rack-mounted machines that have two, four, or eight sockets. The new Sparc SuperCluster T5-8 is, as its name suggests, based on the eight-socket server with the same name, which also made its debut back in March. The machine comes in two sizes, and in each case it has two T5-8 server nodes.
The Sparc SuperCluster T5-8
The half-rack configuration of the Sparc SuperCluster populates each node with four of its eight Sparc T5 chips plus 1TB of main memory, eight 900GB disk drives spinning at 10K RPM, four 40Gb/sec InfiniBand ports and four 10GB/sec Ethernet ports. The InfiniBand ports are used to link the two server nodes to each other in an Oracle RAC cluster and to the Exadata storage servers that are also in the rack. Across those two nodes, you have 128 cores and 2TB of memory.
The Exadata machine that is a sibling to the SuperCluster gets its name from the Xeon-based storage servers, equipped with disks and flash, along with special algorithms called Smart Scan query offload for pre-processing database queries in the Oracle 11g database while the information is compressed, that were launched with the original Exadata systems three years ago. This compression means that data can move around elements of the SuperCluster and Exadata clusters at much faster speeds than it otherwise would in a normal x86 or Sparc cluster.
In the half-rack configuration of the SuperCluster T5-8, there are four of these Exadata storage servers, each configured with two six-core Xeon processors, four 400GB SmartFlash solid state cache cards, and either a dozen 600GB disks spinning at 15K RPM for high performance or a dozen 3TB 7.2K RPM disks for high capacity.
This half-rack has 28.8TB of raw disk storage (using the skinny disks) with 13TB of it usable by the Oracle 11g database. It's able to push 14,400 database disk I/O operations per second, thanks in large part to the 29GB/sec of bandwidth across the flash and 7GB/sec across the disks. Those figures are for data that is uncompressed, and using the fast rather than the fat drives. With fat drives, you have 144TB of raw capacity with 64TB of it usable, but the database disk IOPS drops down to 8,000.
For the full rack configuration of the Sparc SuperCluster T5-8, you populate all eight sockets in each of the two nodes, and double up all of the components in the cluster excepting the raw disks in the server chassis. (Those are already maxxed out.) That gives you 256 cores and 4TB of main memory, and a machine that can do 28,800 database disk IOPS using skinny disks and 16,000 using fat disks.
Both machines also have a ZFS 7320 storage server with 60TB of capacity with dual controllers that is used for storing various system files and binaries.
The SuperCluster T5-8 can run general-purpose applications on Solaris 10 update 1/13 containers, but requires Solaris 11.1 to run the 11g database and the "elastic cloud" automagic scaling software Oracle has cooked up.
How Oracle stacks up the SuperCluster T5-8 to IBM's Power machines
John Fowler, Oracle's executive vice-president of hardware, once again took aim at IBM's latest Power Systems, saying that a pair of Power 780+ machines using eight-core Power7+ processors and backed by flash-infused DS8850 disk arrays would do about the same amount of work as the half-rack SuperCluster T5-8. But the Oracle setup would do it for about a tenth the cost.
"This is not done by the brute force of the system," explained Fowler after breathing in deeply, "but by the efficiency of the SuperCluster and our ability to integrate the various components and do the system software to get a step-function capability to achieve the same performance as a conventional architecture."
Well, sure. But you can also add flash cards and database compression and other tricks to the Power Systems lineup to help boost the performance of AIX-based machines running Big Blue's DB2 database. To be fair, Oracle has done a much better job of integrating and productizing its engineered systems, and IBM has been adding equivalent functions in a piecemeal fashion with a story that has not hung together as well.
If IBM has some spare time tomorrow, it might want to gin up a more appropriate configuration of Power Systems, AIX, and DB2 that more closely matches the SuperCluster, with integrated flash, maybe DB2 PureScale clustering, and DB2 BLU Acceleration compression – you can't let a rival's claims go unchallenged, and complaining to some advertising-policing agency with no teeth months after the fact is just plain silly.
Fowler said that the new SuperCluster T5-8 had about 2.5 times the throughput performance (presumably measured in database transactions), about 33 per cent more storage capacity, and about 3 times better bang for the buck compared to the SuperCluster T4-4 machine that it replaces.
He added that based on customer surveys, Oracle customers moving to SuperClusters were able to deploy systems about five times faster than with conventional standalone servers, storage, and networking stacks, and that it cost about a third less to administer the SuperCluster than the typical Unix cluster with its disparate components and multiple vendors. The SuperCluster is built, managed, patched as a single system by Oracle, which means customers don't have to fuss with it too much.
Oracle has not yet updated its hardware price list, but a full rack of the Sparc SuperCluster T4-4 cost $1,165,000. If the new SuperCluster T5-8 has 2.5X the oomph and a factor of 3X better bang for the buck, then that means the newer machine should cost somewhere on the order of $970,800 per rack.
Oracle did not provide a delivery date for the new SuperCluster, but all of the components in the engineered system are available today, so it should be shipping. ®