Power Consumption

This requirement, while restrictive, is quite achievable. Let's take a look at some numbers. The four motherboard Norbert cluster consumes 250 Watts of power when idle as measured by placing a Kill-a-Watt meter on the power cord. This number, while low is not going to be acceptable to the personal user, where the system may be powered on most or all of the day and the extra processing power may only be used for a small portion of the this time. Part of the project will work to reduce this to less than 100 Watts (possibly less than 50 Watts) when idle. These numbers are achievable by employing the new Core Duo C-states and the new tickless idle feature in the kernel. (See the GearHeads? column in the September Issue.)

When it is fully loaded running the HPL benchmark Norbert draws between 371-483 Watts. The number varied as the benchmark progressed which I assumed is an indication of how modern processors are managing power consumption based on load. Even with the maximum power draw, the power performance ratio is a comfortable 11 Watts per GFLOPS.

Heat Issues

When one thinks of HPC they often envision racks of servers sucking in cold air and pushing out lots of hot air. In an office or desktop environment this is not going to be acceptable. While the CPU is certainly a source of heat, another large source is the power supply. A typical power supply is about 65% efficient (this number depends on the load as well). In terms of heat, that means that 35% of the electricity going into the power supply comes out as heat. New power supplies are aiming for an 80% or better efficiently and thus a reduction in heat. Therefore, in order to minimize the heat generated by the cluster, highly effective power supplies will be required. One aspect of the project will be to find ways to increase the efficiency of the power supply and possibly use only one power supply for the whole cluster.

B<Noise Issues>

In addition to heat, anyone who spends time around servers knows that noise can be an issue as well. To move the cool air through a small server chassis, banks of small fans are normally used. A modern 1U server may employ 12-15 small fans or a lesser amount of blowers to move air through the case. Small fast fans equals big noise. Fortunately in a desktop office environment, larger fans can be employed and the noise level made tolerable. Such is the goal for Limulus. Currently the Norbert cluster used cases with a 3.5 inch fan and a single CPU cooler. The noise level is quite low and certainly tolerable in an office environment. Part of the project will also address more dense packaging schemes that are just as quiet.

B<Price-to-Performance>

One defining feature of a Limulus will be the sub $100 per GFLOPS performance ratio. This number clearly puts Limulus in the low cost realm and ensures that as hardware performance increases that the system cost remains reasonable. It should be noted that Norbert has achieved $55 per GFLOPS.