To strive for higher outputs

The race for higher processor clock speeds on the RISC platform is decelerating as silicon real estate faces the limits of current technology. 2006 will see the emergence of multithreading on CPUs to shatter the barriers that are slowing down throughput computing. by Anil Patrick R

RISC, the name itself evokes feelings of power and blazing performance rates. For anyone tracking the evolution of these processors, it is evident that the field is narrowing for performance options, Moore’s law notwithstanding.

This is why industry majors are looking at innovations to bolster the high-end enterprise’s RISC server performance quotients and that is why the focus will shift in 2006 from incorporating more cores per CPU to simultaneously processing more threads per core.

Moore’s Law Isn’t Enough

The time has come for approaches that go beyond ramping up clock speed. Unlike x86, the RISC architecture has never relied wholly on clock speed. One bottleneck has been memory performance. Of what use are blazing clock speeds if all the output is just going to wait in a queue while memory plays catch up? This is why CPU manufacturers decided that it was time to explore what more can be achieved within the processor while waiting for the slower memory to catch up.

B Satyanarayan CIO and Head, IT Dimexon Diamonds

With silicon real estate shrinking, the first approach tried was to put multiple cores on a single processor. With this approach, performance improved but it still could not hold a candle to multi-processor machines.

Now the focus has shifted towards running more threads per core to increase outputs. It is clear that this approach if properly implemented will be quite advantageous in bulk data crunching environments. “Most software programs perform each process as you request it. Multithreaded software, however, treats each process as a separate thread. So each request launches a specific thread that runs at the same time as the other threads in the program. In situations like unexpected additional traffic or an increased number of users, single threading processors face problems. This is where multi-threaded processors will prove advantageous,” says B Satyanarayan, CIO and Head, IT, Dimexon Diamonds.

Munish Mittal Vice-president, IT HDFC Bank

Apart from the increased performance, multi-threading processors also possess other advantages like easier deployment, space/power savings and cooler operation. “Deployment of these multithreading processors will be easy. In fact it will be easier than deploying blades which have their own set of challenges such as single vendor dependencies and power/cooling considerations,” says Munish Mittal, VP, IT, HDFC Bank.

The Playing Field

Multi-threading as such is not a new approach. It has been around since 1995 on mainframe platforms in one form or another. However, on the RISC server side, IBM’s Power5 processor was the first to emphasise the need for multi-threading when it was launched in May 2004. As of now, there are two players in the RISC field (HP PA-RISC being almost out of the race) adopting multi-threading. These are IBM’s Power5/ Power5+ and Sun’s UltraSPARC T1 “Niagara” processor released in November 2005.

IBM uses Simultaneous Multi-Threading (SMT) technology while Sun has gone in for Chip Multi-Threading (CMT). The differences are legion. Both methods have their pros and cons. At this stage, IBM is focusing on the large enterprise and HPC segments while Sun is focusing on the Web and application server market.

2006 is going to be the year when multi-threading on RISC will be adopted by large enterprises on a notable scale.

Although it is still premature to predict how the future of multi-threading will pan out, 2006 is still the year when RISC multi-threading will be tested to the core. “I personally believe that at present the vendor focus with multi-threaded processors is to do faster processing at the front end for consolidation, and reduce complexity of deployment, with better ease of management. This will be a positive development if it delivers what it promises,” says Mittal.

While IBM leads the clock speed race, Sun focusses on the threads processed per core. These are early days, and it will take a couple of months before the dust settles and the merits and demerits of each approach can be decided. However, the clear positioning of both chips ought to save evaluators some time when making buying selections.

Power Packed Punch

The IBM Power5+ is a migration of the earlier Power5 from 130 nm to a 90 nm process. Until the entry of Niagara, the Power chips had no competition on the multi-threading space.

The Power chips’ forte has been their higher clock speeds, leading to their supremacy in areas where high floating point processing is required, such as transaction processing and HPC applications. At present, the Power5+ is available as dual core and quad core processors with clock speeds of 1.5, 1.65, and 1.9 GHz.

In a Power5 or Power5+ chip, two threads are run per core using SMT. With the SMT approach, multiple threads are able to execute instructions at the same time on the free executors in the CPU. For example, consider a core with multiple executors, with one of the executors empty. This executor will go in for the next waiting thread and process it. This means that the processor is able to execute more instructions while waiting for the memory to respond. IBM has had the edge on this front due to its faster clock speeds.

According to IBM, its present focus is not on the number of threads that can be processed per core. “Even when we first came out with a dual core processor, it was technically possible to run more than two threads. However, there is no significant performance boost if six threads are run on a single core,” says Rajeev Sreekantan, Product Manager, pSeries servers, IBM India.

The logic is that yield does not grow linearly; therefore, there is no need to increase the number of threads. As per IBM, the performance boost achieved with anything more than two threads per chip is not sufficient to justify the effort. “With four threads per chip there is an effective performance boost of 40 percent depending on the application and the way it is threaded. For the next one to two years we don’t need to increase the number of threads per core since performance/process core is sufficient today in IBM’s microprocessors,” says Sreekantan.

As the Niagara Flows

While IBM believes in blazing speeds, Sun’s focus is more on increasing the number of cores per processor and the threads that can be processed per core. Hence, the Niagara is an eight-core processor that can run 32 threads (four threads per core) simultaneously, while operating at a clock speed of 1.2 GHz.

According to Sun, Niagara is not meant to be a fast processor; the focus is on higher throughputs. “A distinction has to be made between how fast threads are created vis-à-vis how many of them can be processed at the same time. A higher number of threads being simultaneously processed means that a higher number of users can be connected at the same time,” says Anil Valluri, Country Director, Client Solutions, Sun Microsystems. This is one reason why Sun has targeted the Niagara at the Web/database segments. “This is not a processor for floating point calculations. It is more for applications such as OLTP, web caching, e-mail, firewalls, application servers, and databases,” says Valluri.

Instead of competing with IBM on the higher clock speed race, Sun decided to attack the memory latencies to improve performance. This was meant to decrease the gap between Moore’s law and memory performance. The Niagara has memory controllers built in on the chip along with major chip space saving designs. “We knocked out unwanted elements and processes that are used for the traditional 5 to 10 percent performance boosts that other vendors use. The same real estate is used for going up to eight cores saving close to 80 percent of it,” says Valluri.

Two of the most interesting features of the Niagara are its low power consumption levels and cooler operating temperatures. Niagara runs at 70 W per processor compared to the Power chips that consume around 350 W and 250 W, a desirable feature in the data centre environments that the Niagara targets.

Sun hopes to cash in on the multi-threaded nature of Solaris in a big way with the Niagara. Existing Solaris software runs on it. “Solaris 10 has been threaded for around six years. Linux has no kernel code for running multiple threads and Windows has no threading capability,” says Valluri. Sun is also trying to sort out the licensing issues associated with eight cores at present. For example, Oracle has already announced that every Niagara chip will be licensed at the rate of two cores.

Multi-threading will eventually change the way servers are sized. Sizing documents and tools will have to be changed.

CIOs Need Convincing

Although multi-threaded RISC CPUs are going to ship in quantity in 2006, CIOs have qualms about the maturity of this technology.

Satyanarayan stresses the need for careful evaluation. “Multi-threading technology is already coming in, but there is a need for careful evaluation by organisations before selection. The biggest concern is whether the technology is mature,” he says.

Yet another concern is that enterprise software applications will have to be rewritten to take full advantage of multi-threaded hardware. That said, there will be performance boosts even if the software hasn’t been optimised for multi-threaded CP Us. “Existing software will gain significant performance boosts even if it has not been written for optimal operation on a multi-threaded processor,” says Satyanarayan.

There’s hope yet as mission-critical applications are prime targets for this technology—once it is bulletproof, that is.

“At HDFC bank we are more interested in applications such as databases, data warehousing and BI. Applications will benefit once multi-threading processors move in this direction,” says Mittal.

anilpatrick@networkmagazineindia.com