Going Green — and Gaining Performance

As more than one energy expert has noted, efficiency plays an enormous part in power consumption, regardless of the context.

SunÙs three-family server strategy fits into this thinking quite well. Based on three separate chip sets designed for particular workloads, use of each member of the family for particular situations can go a long way in reducing power consumption while boosting computational performance.

Sun Fire CoolThreads servers based on UltraSPARC T1 chip multithreading (CMT) processors are ideally suited for massive transactional throughput — making them a perfect fit for managing large numbers of Web-based transactions. In areas that require the industry-standard x86 chip and demand both multithreaded and high floating point performance, AMD Opteron processor-based Sun Fire x64 servers might prove a better choice. And for heavy-duty database applications where single-threaded performance trumps throughput, UltraSPARC IV+-based Sun Enterprise servers deliver the necessary horsepower.

New UltraSPARC T2 CPU doubles threads from 32 to 64

As more than one energy expert has noted, efficiency plays an enormous part in power consumption, regardless of the context.

SunÙs three-family server strategy fits into this thinking quite well. Based on three separate chip sets designed for particular workloads, use of each member of the family for particular situations can go a long way in reducing power consumption while boosting computational performance.

Sun Fire CoolThreads servers based on UltraSPARC T1 chip multithreading (CMT) processors are ideally suited for massive transactional throughput — making them a perfect fit for managing large numbers of Web-based transactions. In areas that require the industry-standard x86 chip and demand both multithreaded and high floating point performance, AMD Opteron processor-based Sun Fire x64 servers might prove a better choice. And for heavy-duty database applications where single-threaded performance trumps throughput, UltraSPARC IV+-based Sun Enterprise servers deliver the necessary horsepower.

Optimizing each member of the Sun server family for energy efficiency and computational performance is a constant process that has resulted in numerous innovations, with more to follow. A case in point can be found in preparations for rolling out the next-generation UltraSPARC T2 for CoolThreads servers later this year.

This processor will add more functionality to a server line noted for handling heavy traffic, and it will raise the bar for processor efficiency. A brief look at what to expect in UltraSPARC T2 — and how it builds off the success of UltraSPARC T1 — shows how Sun is providing solutions for markets in which performance and power consumption are of prime importance.

Thread Count Doubling Becomes Par for the Course
Just 16 months ago, Sun rolled out the UltraSPARC T1 chip in the new CoolThreads T1000 and T2000 server lines used for Web application tiers. These environments, with their increasing number of transactions, somewhat mirror the way population growth strains scarce resources. But in computing, at least, scaling to meet the increasing demand without requiring more space and power resources provides part of the answer. With eight cores per processor and four threads per core, UltraSPARC T1 acts like 32 distinct servers working simultaneously.

For example, in the transaction-heavy telecommunications industry, where CoolThreads T2000 servers are widely deployed, some customers report reducing datacenter operating costs by up to 60 percent — and much of these savings come from the reduction in power consumption. Plus, by replacing as many as five older-generation servers with single CMT-based servers such as the CoolThreads T2000, telcos can reduce the need to continually expand or build datacenters — each of which has an environmental impact beyond server power consumption.

One of the reasons for this dramatic power savings comes from how each of the 32 separate threads in a CMT-based server interacts with a database simultaneously. Should one thread stall, the others simply keep working. This makes the chip ideal for use in telecommunications and other environments characterized by heavy online transaction processing (OLTP).

SunÙs CMT roadmap continues to answer OLTP challenges by doubling thread counts over a relatively short period time. UltraSPARC T1 has 32 nonextendible threads, which is sufficient for most high-transaction environments now, but to increase computing power while decreasing power consumption, UltraSPARC T2 will come with 64 nonextendible threads.

Sun foresees the need for extending thread count beyond 64 separate instances of a computer, which is why the chip that will follow UltraSPARC T2 (currently code-named Victoria Falls) is being designed to have 128 threads. These threads, however, will be fully extendible. This will make it possible to link two instances of Victoria Falls — both sharing common memory through a hub chip — for a grand total of 256 threads.

So how do these chips compare with other chips on the market? The current thread count on UltraSPARC T1 is eight times greater than the number of threads found on AMD Opteron and many Intel cores. Thus, the power savings are not equal. An UltraSPARC T1 or UltraSPARC T2 core consumes about 4 watts, while these competitive cores may consume up to 25 watts. And in OLTP throughput performance, both the UltraSPARC T1 and T2 chips have performance equivalent to or better than an Intel core given the right workload.

A Commonsense Approach Reduces Power Consumption
Other developments with UltraSPARC T2 and Victoria Falls will allow these CMT chips to extend their efficiencies to markets beyond Web infrastructure and database serving. While UltraSPARC T1 currently has a single floating point unit shared by all processors, UltraSPARC T2 will have eight separate instances of floating point units. This will make the power-saving chip a very viable option for other markets, such as bioinformatics, which is largely defined by high floating point content.

Conservation is part of any energy efficiency plan, and the Sun CMT roadmap takes a commonsense approach to rationing power in memory functions. Memory controllers in UltraSPARC T1 hold power across the chip to about 65 watts, while clock-eating shuts the clock down when it is not needed. An instruction-issue technique further saves energy by metering the rate at which instructions are issued, helping to keep temperatures down.

Energy consumption at the system level rises slightly with UltraSPARC T2 and Victoria Falls due to increased power in the core and fully buffered dual in-line memory (DIMM) in these chips. The addition of fully buffered DIMM will make the two chips ideal for direct-attached controllers and processors that require enormous amounts of memory capacity at high bandwidth.

Yet fully buffered DIMM can also lead to both increased latency and power consumption. This amplification of energy use is offset by the efficiencies of more threads, of course, but several techniques will be used to meet the challenges of new features in UltraSPARC T2 and Victoria Falls. Just as energy-conscious people turn off lights when they leave a room, the clock memory in UltraSPARC T2 and Victoria Falls detects idle periods in memory and issues commands that turn off clocks until the processor requires access, which reduces power burdens.

A second power-saving feature can be found in memory metering. This helps to cap the peak power of the memory subsystem. Users define the number of memory activations they wish to allow during a specific period. The controller then limits memory activation, which aids in capping peak power consumption.

In Sun labs, the power savings have been significant. Preliminary tests show that by running the Solaris 10 OS with moderate use of memory, power consumption can be reduced by up to 50 percent.

Areas that would seem unlikely to keep the electrical meter from advancing are also being used to reduce power consumption in UltraSPARC T2 and Victoria Falls. In SSL and other authorization techniques, processing often stalls because data gets sent to a separate SSL processor. But by putting dedicated authentication processing on the chip itself, the SSL process is accelerated by having one or two threads handle authentication — while other threads continue processing other tasks.

Sun Servers Provide Real Environmental and Enterprise Choices UltraSPARC T2 will be a hard act to follow for many other processor and server manufacturers. Add the Sun Fire x64 and Sun Fire Enterprise lines, and all the members of the Sun server family represent a complete portfolio for datacenter solutions with specific kinds of workloads that require particular kinds of platforms.

Applications that run only on an x86/x64 instruction set or Windows will be able to take advantage of AMD Opteron-processor based Sun Fire x64 servers — and get more even more performance when Sun doubles the core count later in the 2007. This core, it should be noted, is identical to the core used in many Energy Star-compliant desktops.

All of this ties into Sun-wide initiatives to reduce energy consumption in the datacenter through close examination of how power is actually used. For example, under lightly loaded conditions, systems with 16 and 32 cores can migrate workload down to fewer and fewer cores. While these cores canÙt be powered off completely, they can be temporarily disabled in the same manner that UltraSPARC T2 suspends clock functions for the lowest level of power consumption.

This can be a huge help during periods of relative inactivity. In situations characterized by peaks and valleys, where load can range between 100 percent and 15 percent, users can now consolidate increasing amounts of load down to fewer and fewer servers when appropriate.

Such a scenario usually does not characterize database environments that rely on UltraSPARC IV+ Sun Fire Enterprise servers. However, these platforms do have periods of downtime. More important for enterprises running applications that require single-thread performance — such as SAP — Sun Fire Enterprise server socket architecture is designed to minimize power inefficiency.

Picking the right tool for the right job wonÙt answer every power consumption issue, but it will create efficiencies so energy problems can be avoided. In addition to picking the right tool, Sun is also examining how these tools work with other IT components. For example, Sun will soon ensure that more efficient power supplies will work with all rackable Sun servers, such as the x64 line. This will ensure at least 80 percent power supply efficiency, regardless of how large the load may be on a particular server.

Weaning the world from carbon-based power is likely to be a long process — and will require more than one way of approaching the problem. In the IT microcosm that mirrors the rest of the world, SunÙs leadership in servers and processors is providing many of the answers — now and for the foreseeable future.

Read the original article: http://www.sun.com/emrkt/innercircle/newsletter/0407feature.html?cid=916979