CTO Office

Faster Doesn't Always Mean Better

For years, processor manufacturers have been in a megahertz race. Moore's Law tells us that the number of transistors on a single chip will double every couple of years. Traditionally the additional density has been used to drive increases in processor clock rate; that's why we have gigahertz processors today.

But is a faster processor always better? Memory speeds are increasing at a much slower rate than processor speeds, so memory latency is becoming the primary performance bottleneck for real workloads. In fact, the gap between memory and chip speeds often leave today's processors stalled roughly 75 percent of the time while they wait to retrieve data from memory. Increasing the speed of processor means it spends more time waiting. So what's really important isn't clock speed or megahertz; it's the overall amount of work a processor can get done — it's throughput.

The Need for Throughput Computing

Throughput computing focuses on increasing the overall amount of work done in a given amount of time, rather than the speed of an individual component. Rather than pushing the bleeding edge on clock speed, Sun is designing the new UltraSPARC processors to maximize throughput for network computing loads. These new processors incorporate chip multithreading (CMT), which packs the power of several of today's servers onto a single chip.

How does it work? Today's typical processors execute one thread at a time. When a thread needs to access memory, the processor stalls until the memory request is complete. In contrast, CMT processors can execute several threads at the same time. When one thread is waiting to fetch data, the processor can still execute other threads. With CMT, idle time caused by memory bottlenecks is replaced with useful work. This doesn't make any one thread run faster, but when you look across the whole system it results in a dramatic improvement in processor utilization and overall system performance.

By significantly increasing application throughput, CMT processors can help you do more with less. You'll need fewer systems to handle the same workloads, driving down costs by dramatically reducing space, power, and maintenance requirements. With fewer systems to manage, you can expect improvements in system reliability and availability. Within the next couple of years, we plan to cut the cost and complexity of network computing by delivering blade processors that increase today's blade throughput 15 times. And by mid-decade, we're on track to ship processors that will increase throughput by a factor of 30.

CMT Boosts Normal Workloads, Too

Sun will continue to design processors optimized for single-threaded performance for applications that need it. But we feel that CMT processors are the right answer for the huge number of parallel applications that make up today's business workloads: databases, Web servers, application servers, directory servers, J2EE applications, and more. You can also use CMT processors for server consolidation, where multiple workloads are run on the same system. This allows even single-threaded workloads to benefit from the throughput improvements of CMT. We're working on some exciting improvements to Solaris to make consolidation easier and safer.

CMT processors are a testament to the adage that good things come in small packages. A single blade will be able to do the work of 32 single-processor servers. Power usage will drop dramatically. IT managers will have only one blade system to manage instead of 32 servers. And total cost of ownership could drop significantly.

There's no need to chase after megahertz, because faster isn't always better. Wringing every last hertz out of a microprocessor has diminishing returns. The overall amount of work your system performs is usually more important than the speed of a single transaction. With CMT, you'll get much greater throughput for the same cost. Think about what you could do with the savings. Instead of sinking your IT budget into maintaining operations, you can free up resources to deliver new network services. If that's not an innovation in computing, I don't know what is.