Oracle Takes On Xeons With Sparc S7
June 30, 2016 Timothy Prickett Morgan
It is an accepted principle of modern infrastructure that at a certain scale, customization like that done by Google, Amazon Web Services, Microsoft, or Baidu pays off. While Oracle is building its own public cloud, it does not have the kind of scale that these companies do, but it does have something else that warrants customization and co-design up and down its stack: more than 420,000 customers who generate $38.5 billion in sales.
This, in a nutshell, is why Oracle continues to invest in its Sparc processors even though many of its customers deploy Oracle’s middleware, database, and application software on X86-based systems created by itself and its rivals in the systems racket. And that is why it is explicitly taking on Intel-based platform – including its own – with the new “Sonoma” Sparc S7 processor and the systems that have just debuted employing it.
The Oracle enterprise base is broad and deep, and utterly dwarfs the one that AWS has created in its first decade, and rivals that which Microsoft has created for the Windows Server platform and which is fueling its own Azure cloud. While Oracle has spent more than six years building up its 5,000-strong customer base for its Exadata, Exalogic, and Exalytics engineered systems (which are based on Xeon processors), its software base is orders of magnitude larger and if Oracle can get even a modest attach rate on its Sparc systems in this base, it has an opportunity to radically expand its hardware business. Look at the numbers: Oracle has 310,000 database customers, 120,000 Fusion middleware customers, and 105,000 application customers, and over 75,000 companies that have bought its Sparc and X86 systems. Most of them have Sparc machinery, even though the business for these systems, like all other RISC/Unix iron, has been in decline since reaching a peak in dot-com boom nearly two decades ago.
Say what you will about Oracle co-founder and CTO, Larry Ellison, but he has made good on his word to create a Sparc processor and system roadmap and, more importantly, stick to it and deliver it on time. The “Sonoma” Sparc S7 processor, which was unveiled at last year’s Hot Chips conference, takes the S4 core that was originally used in the Sparc M7 chip that Oracle started shipping in big NUMA systems late last year and creates a chip that is more suitable for scale-out two-socket servers that dominate the datacenters of the world. The S7 was not on the original Sparc roadmap that Oracle put out in 2010 after it acquired Sun Microsystems, but it is logically consistent with two-tiered approach that most server chip makers have had in recent years to cover scale-out and scale-up systems.
The S4 core has a dual-issue, out-of-order execution unit and has dynamic threading that ranges from one to eight threads per core. Each S4 core has 16 KB of L1 data cache and 16 KB of L1 instruction cache. The M7 and S7 chips organize the cores into clusters of four, and to a certain way of thinking the eight-core S7 is a quarter of a 32-core M7 with InfiniBand and some other features integrated onto that could not be crammed into the M7 and that might make it onto the future M8. Each set of four cores on the S7 chip has a 256 KB L2 instruction cache, and each pair of S4 cores block also has a 256 KB writeback data cache. Both of these L2 caches provide up to 500 GB/sec of bandwidth each. The four core block on the S7 die has an 8 MB L3 cache, and sports two DDR4 controllers that can support memory running at 2.13 GHz and 2.4 GHz clock speeds. Using 64 GB memory sticks, a two-socket S7 system can have up to 1 TB of main memory and has a peak memory bandwidth of 77 GB/sec. There are also two PCI-Express 3.0 peripheral controllers on the S7 die as well.
Marshall Choy, vice president of product management for systems at Oracle, says that the memory latency with the S7 is a little bit lower than for the M7 because the memory controllers are now integrated on the processor. With the M7, the processor had a controller but the buffer chips that actually ran the DDR4 protocol were external to the processor and put on the memory cards (just like Intel does with the Xeon E7 and IBM does with the Power8). As IBM will be doing with the scale-out version of the Power9 chip next year and as Intel does with Xeon E5s already, the S7 will use stock DDR4 memory without buffer chips because the memory controller is wholly on the die. Moreover, because the processor is smaller and Taiwan Semiconductor Manufacturing Corp has matured the 20 nanometer process that is used to make the S7, its clock speed, at 4.27 GHz, is a little bit higher for the M7, which cycles at 4.13 GHz. Both have fairly high cycle times compared to Intel Xeons and are in the same ballpark as the IBM Power8. The S7 and M7 cores both support eight threads per core, same as the Power8, and four times what Intel supports per core with the Xeons. So the S7 should, core for core, outperform the M7 and, more importantly, give both the Power8 and the Xeon E5 a run for the money in the scale-out datacenter.
“The goal here is to bring down the economics of the system, whether we are talking about servers or engineered systems,” Choy tells The Next Platform, “and what we have achieved here is commodity X86 cost points but also bringing in added enterprise functionality specifically around software in silicon. This is not just a pricing exercise for us, in terms of getting these X86 price points, but an engineering exercise in reducing our overall costs. Our margin model is preserved across S7 and M7.”
That on-chip acceleration includes accelerators for security, encryption, database processing, and analytics that were first etched into transistors with the S4 cores in the M7 processors for high-end Sparc M7 systems. One important feature in the software-in-silicon stack from Oracle is the ability to transmit fully encrypted virtual machines in Solaris when live migrations occur between two physical machines, and another is the Data Analytics Acceleration, which is an offload engine built onto the chip that does certain kinds of SQL processing common in databases (from Oracle and others) and speeds up the SQL queries by a factor of 10X.
“We have seen on the X86 side for the past couple of generations now per core performance has stagnated a little bit, but with Sparc we continue to increase per core performance,” brags Choy. “Depending on the workload, we will provide 50 to 100 percent improvement in per core performance compared to the latest generation Broadwell Xeon cores. Obviously we have a great story around efficiency, with the offload, and much more bandwidth.”
(We will be drilling down into the performance of the S7 systems versus Xeon E5 and Power8 machines separately.)
The one thing that made the S7 chip particularly interesting was the integrated InfiniBand host controller that provides two x4 lanes that run at 56 Gb/sec speeds and that provide a total of 28 GB/sec of bandwidth out of each Sonoma chip. In a two-socket configuration, both InfiniBand ports can be active, providing dual rail configurations with redundancy and load balancing to get around traffic snarls in the network. For reasons that Oracle is being a little bit vague about, these InfiniBand ports are not activated on the S7 chip, which could mean all of the quirks might not be worked out of them yet. As we have previously reported, Oracle is working on its own 100 Gb/sec EDR InfiniBand leaf and fabric switches, based on its own silicon, not that of partner Mellanox Technologies. These are expected to ship around now, but have not as far as we know. Oracle has a roadmap to develop its own 200 Gb/sec HDR InfiniBand switch ASICs, too. InfiniBand is at the heart of the engineered systems that Oracle sells, and the company no doubt wants to use InfiniBand to build its public cloud and the clusters that its customers put on site.
It is not clear when Oracle will fire up those InfiniBand ports on the S7, but it may be concerned about an impedance mismatch with 56 Gb/sec ports on the S7 server chips and 100 Gb/sec ports on its impending switches. Perhaps Oracle is taking a bit of time to crank up the clocks on the ports for a next-rev on the chips?
The S7 Systems
The S7 processors, like other Sparc chips, run Oracle’s Solaris Unix variant, which is well-regarded but nowhere as mainstream as Linux or Windows Server is these days – with Linux dominating a lot of the back-end and front-end at many large enterprises where Oracle derives the bulk of its revenues. The S7 machines support Solaris 11.3 to be specific. Solaris is bundled with the system, as is the case with all Sparc machines from Oracle, and like Linux is prebundled on its X86 iron. The systems can support Oracle 11g Release 2 or Oracle 12c in the Enterprise Editions of the company’s database.
There are two rack-based machines that employ the S7 processor, plus a preconfigured MiniCluster system that falls short of a full-on engineered system but is more like a converged system complete with servers, storage, and switching.
The Sparc S7-2 server is a 1U pizza box machine that has two processors, and customers can buy them with either one or two CPUs installed. The base system comes with a four-port 10 Gb/sec Ethernet controller, like other Sparc and Xeon servers from Oracle, and has room for three 2.5-inch drives, which can be 600 GB or 1.2 TB SAS disks, 400 GB SAS flash drives, or 3.2 TB NVM-Express flash drives. The Sparc S7-2L server comes in a 2U form factor with two processors configured in the box from the getgo and room for 24 2.5-inch drives or 15 3.5-inch drives. (Oracle supports 8 TB SAS-3 disk drives in the 3.5-inch bays.) The 24 bay setup can have up to a dozen NVM-Express devices.
The MiniCluster S2-7 is just what the name suggests, it is a baby cluster built from two S7-2 servers and an Oracle disk array. It has a total of 32 cores and 1 TB of memory as configured (using 32 GB memory sticks and fully populating the machine, so be careful) plus 16.8 TB of flash storage and 48 TB of disk storage.
The S7 systems are available now; pricing for configured machines was not available at press time. Oracle is also selling dedicated instances based on S7 nodes on its Oracle Cloud public cloud.
The real question, then, is when does Oracle switch its Exadata platform from Xeon to Sparc? Choy says that Oracle will continue to sell X86-based engineered systems, as it has been doing and is mum about any specific future plans for the S7 systems. But clearly if Oracle really believes that Sparc machines are better, then there should be a variant of Exadata, Exalogic, and Exalytics – you do the naming, and all you gotta do is add an S in front – based totally on Sparc and that shows these machines outrunning Xeons.