By John Soat
At Oracle OpenWorld 2015 in San Francisco, John Fowler, Oracle’s executive vice president for systems, introduced new servers based on the company’s latest microprocessor, SPARC M7. The SPARC M7 is the sixth SPARC microprocessor Oracle has released since it acquired Sun Microsystems in 2010. The new processor is something of an engineering watershed for Oracle because it features security and performance functions directly hard-wired onto the chip using a technique known as “software in silicon.”
Fowler sat down to talk about the engineering effort that led to the SPARC M7, and the implications of the new processor in the technology market.
John Fowler introduces software for the M7.
Q: Where does the new SPARC M7 processor fit into Oracle’s overall technology strategy?
Fowler: Oracle acquired Sun some five-and-a-half years ago. At the time people didn’t understand what we were doing. But now, after generations of engineered systems and innovations, they can see that we’re really about providing very highly integrated and co-engineered products that include hardware and software.
The exciting thing about SPARC M7 is that this is where we take that whole concept down to the silicon level. We’re able to exploit the opportunity of combining software and hardware engineers and do some things in the microprocessor itself to directly help enterprise applications.
We’re obviously very excited about this. These kinds of development efforts are extraordinarily expensive and long-lived. Oracle has both the economic capability plus the intellectual property capability to undertake this kind of project.
We don’t think of the SPARC M7 as a product. We think of it as an expression of our strategy.
Q: How did the technique known as “software in silicon” come about?
Fowler: We took hundreds of ideas that the hardware and software co-design teams had and we distilled those into a subset of ideas that was the most interesting. Then we wrote simulation models and then design models for the most promising ideas. We coded those in simulated processors and then ran real benchmarks against them to see if they worked properly, the way we thought they would.
So what you see coming out in SPARC M7 is not the product of just a few ideas, it’s the product of a very disciplined, empirical, and engineering-driven activity that starts as an idea, goes all the way through simulation modeling and then into product. This has been quite an exciting journey for us.
Q: Is this an innovation in terms of microprocessor development?
Fowler: There’s an interesting aspect here if you look at the history of microprocessors. About 20 years ago 64-bit processors became available and they enabled a whole new generation of software. Ten years ago multicore and multithreaded processors became available. That really changed the economics of performance. We’ve been climbing up the core and thread curve ever since.
Unfortunately, you get an element of diminishing returns by just constantly adding more cores and capabilities. A more thoughtful view is to take a look at what functions you can put on the chip and therefore enable better computing.
I believe the next decade is going to be about doing more in terms of embedding software functions on the chips. The SPARC M7 is the first in this “capability generation.” We’re very happy to be first out of the gate.
Q: Where do you see SPARC M7 making the most difference to enterprise computing?
Fowler: There are three basic areas that we worked on. The first area is around security, where we tried to tackle two important elements.
The first element is very high-speed encryption. We’ve actually had a history of working on this in chips and we improved it again in SPARC M7. By incorporating very high performance encryption in the chip we’re able to not only do it quickly but also leave a substantial amount of the processor’s resources to do other projects. That’s very important. A software-only encryption scheme typically will consume all of the processing resources of the chip.
Encryption is a foundational element of enterprise computing, and we believe we should have done this a long time ago. Everyone needs to run their data centers fully encrypted—nothing should be done “in the clear.” That’s the future of computing. Absolutely everything is encrypted, whether it’s stored on a disk, on a wire, on your laptop, in the back-end data center, or on a tape drive. This is the first processor that enables that.
Q: What other security area was important?
Fowler: The second thing we did on security was to enable memory protection "in a feature known as Silicon Secured Memory". The software team drove this effort initially, because they were developing large in-memory data stores and wanted a way to protect them from corruption.
We later figured out that this was a more general feature that, if added to the processor and enabled for all applications, would eliminate a pretty broad class of compromises. Take Heartbleed for example. Heartbleed became a brand-name vulnerability because of a programming error, and our hardware memory protects against this kind of error.
Q: What goes into the decision-making of what to try to put on a chip?
Fowler: The database team focused in on, how do we improve not just the performance but also the efficiency of the database? This is driven by our vision that everything moves to in-memory computing. I know in the world of hardware today people talk a lot about flash. Frankly, I don’t think it’s that interesting. It’s an intermediate point to putting everything in memory.
In the case of database integration with the processor, we did two specific things that are closely related. First, we took two portions of SQL processing—the part that scans for particular strings across a large amount of memory and the part that helps you filter and join rows—which are very low levels of how the database operates, and we encapsulated those in co-processors in silicon.
That means, first of all, if you want to tackle a very, very large problem, we have extraordinary performance. Also, since these are operations that normally use a lot of processor time when executed only in software, we’re giving back a huge amount of efficiency to the customer. That’s because we’re making the database more effective on our processors.
So that’s an area that took a lot of study. What would we put in a chip that would be straightforward enough to encapsulate in a chip design, but still have a huge benefit to the customer? And since we own the database and the chip, we’re able to deconstruct and find out which functions of the database go best on the chip.
Q: What other embedded feature contributes significantly to performance?
Fowler: That query acceleration is paired with another glorious feature: memory decompression. The idea here is—and maybe it’s a little counter-intuitive—we’ve connected up a silicon accelerator for decompressing data together with the SQL stuff that I talked about a moment ago. It lets a customer take a database that’s fairly large in size, say two terabytes, and have the whole thing reside in memory at say, one-eighth or one-tenth of the size that it normally is on disk, and still operate on it with full performance.
So these two features go together. We offload database processing from the cores, which makes them more efficient and faster. Then we’re also able to put the entire database into a lot less memory, which lets users have either less expensive systems or tackle larger problems.
Those are the first two pillars of SPARC M7—security and the database. And they’re both very strong upgrades to what you can get in processors today.
Q: And the third?
Fowler: The last thing, and it continues to be our hallmark in processor engineering, is that we wanted to make sure that we had the world’s fastest commercial microprocessor. So whether you want to tackle a large problem, or simplify your environment by having a small number of machines—whether you’re running Oracle applications or not—we wanted to have the world’s fastest commercial microprocessor.
The SPARC M7 is the first server chip in the world with 32 cores and 256 threads. By any measure of memory bandwidth, overall performance, and benchmarking, it’s the fastest commercial processor in the world.
For us, that’s the bottom line to the bigger SPARC M7 story: Add in security, do some significant upgrades for the database, and then, for good measure, make sure we’re faster than anybody else—even without those tricks.Q: Now that servers using SPARC M7 are available, what are you most excited about?
Fowler: One of the things engineers get excited about is witnessing the ways customers actually use these things. It will be very interesting to see how different customers take advantage of these—Silicon Secured Memory and the new encryption technology, the database acceleration, the memory compression—because they are very rich features.