Server road map: Beyond quad-core
- 27 February, 2007 11:28
In 1973, Pete Townshend and The Who wrote and sang about Quadrophenia. And although it took another 34 years for quad-core servers to be counted as a commercial success, by all accounts, multicore server evolution is just beginning.
As the decade draws to a close, x86-based servers will have eight or even 16 cores in a single chip, said Nathan Brookwood, an analyst at Insight 64. The reason: Adding more cores is the fastest way to performance gains.
Improving memory technology can add 5 percent to 10 percent to system performance, and an updated processor architecture might provide an additional 10 percent boost, Brookwood said. But doubling core density within a processor can instantly add 50 percent or more in performance.
"Compare the level of performance gain we are seeing with quad-core processors to what Intel was able to provide in the move from Pentium 3 to Pentium 4," Brockwood said. Even though the Pentium 4 was a whole new microarchitecture, the move boosted performance by only around 20 percent, he explained. Intel's first quad-core Xeons, by contrast, are promising a 40 percent or greater increase.
There seems to be no point in the foreseeable future at which doubling cores every two years for mainstream servers will reach diminishing returns. Eight-core designs in 2009, 16 cores in 2011 and 32 cores in 2013 will be the route to processor performance enhancement just about indefinitely, most observers agree.
"There is always more work to be done," said Martin Reynolds, a Gartner Inc. analyst. "With more cores, you can get more work done."
For his part, Brookwood said, the multicore era is at its earliest stages. "We are not running into walls there."
That said, though, there's no definite word on how the industry will get there. Intel Corp. and Advanced Micro Devices Inc. have taken different paths to their quad-core designs. Some analysts believe, though, that ultimately AMD might have to take a more Intel-like approach to really catch up, and then pass Intel, in the multicore market. (See related story: "")
Microprocessor makers turned to multicore designs to solve some fundamental problems. Semiconductor technology continues along the path defined by Intel co-founder Gordon Moore in 1965. Moore's Law says that the number of transistors on a given chip will double roughly every two years. But the heat generated by packing so much in one tiny space has demanded a new approach to achieving incremental performance gains.
The biannual Moore's Law increase comes at the same time the width of transistor lines within a chip shrinks. This allows more transistors to be placed inside a given chip area. Today, leading semiconductor vendors are producing chips at either 90- or 65-nanometer line widths, and the move to 45 nm will begin by some vendors later this year.
But while the transistor budgets continue to increase, microprocessor designs began to hit a wall a few years ago in their ability to continue to accelerate the clock frequencies of those chips while keeping the heat produced at a manageable level. Digital Power Group, a Washington-based energy research firm, estimates that computers now consume about 10 percent of all the electricity generated in the U.S., a figure that could double by 2015. Legislation is being considered to force businesses and technology providers to reduce energy consumption.
By moving to multiple cores inside a single chip, processor manufacturers can reduce or maintain clock speeds and at the same time contain the associated heat generated. Overall performance can be dramatically boosted by doubling the available processing engines inside the same silicon real estate while maintaining stable power levels.
"It's really providing amazing new performance levels," said David Tuhy, a general manager at Intel's Business Client Group. "We're offering 50 percent more performance than our best dual-core processors, and it's four and half times the performance of our original single core Xeon. And the power didn't go up."
There seems to be no upper limit to the core escalation for the foreseeable future. Intel recently announced it has created a research chip with 80 cores, which is expected to dissipate less energy than its current quad-core design. That chip is probably five to eight years away from commercialization, but other vendors are already hitting the market with "massively parallel" processor offerings.
Sun Microsystems in late 2005 introduced its first Sparc processors with multiple cores, code-named Niagara. That chip has eight cores, and each core operates with four independent threads, providing a total of 32 computing elements on a single chip. By midyear, Sun plans to introduce Niagara 2, which will remain eight cores but will have eight threads per core for a total of 64 execution threads. Sun also plans to introduce in mid-2008 its Rock processor, another Sparc-based design which will have 16 cores.
On the megacore front is Azul Systems, which has been offering servers based on its 24-core Vega processor since 2005. In December, Azul introduced new servers that use its latest-generation Vega 2 processor, which has 48 cores.
Early multicore customers
CitiStreet, a benefits outsourcer, is one of the first businesses to deploy the Vega 2-based Azul Compute Appliance servers. CitiStreet has seven of the servers, each with two 48-core processors. The systems are used across all production, disaster recovery, acceptance and test environments.
Barry Strasnick, CIO of CitiStreet, says the servers allow his company to quickly scale infrastructure to meet high growth demands while providing a 100 percent performance boost over the dual-core Xeon-based servers it had used previously.
"Cost-effectively managing the growth we are experiencing requires scalability and performance [beyond] what traditional servers alone can deliver," Strasnick said.
Web and e-mail hosting provider Concentric Systems Inc. made a switch from older single-core Sparc-based servers to Sun's Niagara-based servers late last year. The company has been able to replace as many as eight of the older systems with each new server, said Barbara Branaman, Concentric's president.
"We are always looking for ways to handle more volume on fewer boxes, which of course can help us reduce energy consumption," Branaman said. "Being able to grow capacity within the same physical footprint and power envelope is a huge advantage."
To date, Concentric has deployed nine Sun Fire T2000 and T1000 Niagara-based servers. The company has plans to add five systems and is looking forward to further performance increases expected by the upcoming Niagara 2-based servers, she said.
Geoff Shorter, IT infrastructure manager at The Charlotte Observer, is expecting significant improvement in virtualization density when the newspaper begins implementing quad-core servers based on Xeon processors later this year.
The newspaper has already started migrating some of its most-critical applications to a virtualized environment on dual-core Xeon servers where Shorter has been able to run seven to 12 virtual servers per processor. He believes he will be able to get 15 to 30 virtual severs per processor on quad-core systems.
"If you can get 10 virtual servers on one hardware node, that may cost you about US$12,000, as compared to US$50,000 for 10 hardware-based servers," he said.
Bandwidth, software and other issues
Simply migrating to multicore systems does not guarantee efficiency, however, cautions Gartner's Reynolds. IT managers must do some upfront planning to ensure that they have enough network bandwidth to keep the additional processor cores fed, and they must make sure their applications are optimized to take full advantage of multicore environments.
"IT managers don't get fired because the electricity bill is too high," Reynolds said. "They get fired because they can't deliver the computing requirements of the organization."
Businesses also need to thoroughly evaluate the effects of software licensing as they move to servers with multicore processors, he explained. While Microsoft has already promised that it will continue to base its licenses on sockets and not the number of processing cores, the licensing path for other applications is not as clear.
"Each business needs to ensure they are not going to get hit with a big license upgrade fee as they move from two-core to four-core systems," Reynolds said.
In response, the chip makers say that licensing costs are not going to be a huge deal. "The big hump was in going from single to dual core, but now we have a pretty solid understanding of most licensing strategies in the market," said Pat Patla, director of Opteron marketing at AMD.
Stori Waugh, senior manager at Dell's server product group, said the company has been working closely with all major application and operating system vendors to advance the "license by socket, not by core" strategy. As many as 90 percent of software vendors will adopt a per-socket licensing model, she said.
Another question is how efficient software can be on multicore processors when the applications were designed for earlier generations of hardware. The processor manufacturers maintain that the bulk of the application optimization for multicore environments is done; they say it was completed during the transition from single- to dual-core systems.
Brookwood agreed that the "heavy lifting" for migrating software to multicore environments was completed in the dual-core transition, but he noted that generational fine-tuning will be needed to get optimum performance out of the new processors.
"It will always be dependent on the specific software package," Brookwood said. Virtualization is one example of the ongoing work between third-party software vendors and the chip makers. AMD and Intel have rolled out dual-core, x86 processors with hardware-assisted virtualization features within the past year, he says. Companies such as VMware Inc. and Microsoft Corp. continue to work to optimize their virtualization software to make the most of latest processors.
"The ultimate test is always whether it works for IT professional and makes their applications better," Brookwood said.
Markus Levy, an analyst who serves as president of the Multicore Association and the Embedded Microprocessor Benchmark Consortium, said an increasing number of applications will require higher-level optimization efforts. In other words, he said, it won't be enough to simply run existing software on next-generation processors with larger available core density.
"Even when Intel goes to 16 cores, there will be a need for additional acceleration technologies," Levy said. "As we add more and more cores, we'll also see that general purpose processors can only do so much for some tasks, and the need for specialized acceleration technology will increase."
Multicore chips aplenty
A quick rundown of the multicore options available now and those that will soon be on the market.
-- Intel in late 2006 introduced its first quad-core Xeon processors. The company projects it will sell more than a million quad-core processors by midyear. The processors are manufactured by packaging two Intel dual-core processors in a single chip.
-- AMD promises its first quad-core Opteron processors by midyear. They will be manufactured using a "native" design that will place four independent cores on a single chip.
-- IBM has been offering quad-core Power processors since 2005. Similar to Intel, the chips are manufactured using a multichip module.
-- Sun introduced the Sparc-based Niagara in late 2005. The chip has eight cores, each operating with four independent threads. By midyear, Sun promises Niagara 2, which will have eight cores, each with eight threads. Also in the wings, the 16-core, Sparc-based Rock processor is expected by mid-2008.
-- Azul Systems introduced its 24-core Vega chip in 2005. In December, the company announced Vega 2, a 48-core processor.