Sign up for our newsletter and get the latest HPC news and analysis.
Send me information from insideHPC:

Open Computing Drives Innovation and Efficiency

The Open Compute Project Foundation was created to design the most efficient server, storage and related designs for the next generation of data centers in an open and collaborative development model.  By sharing designs that maximize density, minimize power consumption and deliver expected performance, completely new computing environments can be developed, free from the limitations of legacy thinking.  Organizations can now innovate new designs that push maximized computing densities for the most cost-effective and power-efficient performance.

This the second article in a series, from the insideHPC Guide to Open Computing.

            Increased Densities

HPC servers typically consist of two individual sockets, associated memory, power supplies, PCI buses, and other components. These servers would usually be designed to fit into a one-rack unit space measuring 19 inches (482.6 millimeters) wide, 1.75 inches (44.45 millimeters) high and up to 29 inches (736 millimeters) deep.  The rack unit was the standard to which every rack-based server had to conform. Whether a particular server actually required an entire rack unit of space was often not taken into account. The maximum density of compute power (in terms of sockets per cubic inch) for a 42U high-standard rack was 84 sockets (width x depth x 42*1U).  While this was historically impressive and translated to 1,344 cores in a single maximized rack (84 sockets * 16 cores/socket), there was still a large quantity of wasted space measured against the optimal potential for this volume.  One reason for this suboptimal density was that servers typically contained additional hardware for various essential tasks that enable HPC.

            Rack Hardware

A computer rack is a fairly simple item.  Its basic function is to hold a cluster of computers and related equipment, while allowing for these components to be cooled.  While this arrangement works, a closer inspection reveals numerous inefficiencies with racks that are designed to accommodate different sizes and types of equipment. Rack hardware has evolved to be standard in some respects, while vendor-specific in others.  Although the standard width of a rack today is 19 inches (482.6 millimeters), there is no compelling reason as to why it must remain the standard. The “U” that is commonly referred to as rack height is 1.75 inches (44.45 millimeters). Thus, computer equipment vendors have had to design products as a multiple “U,” whether or not that is the ideal height for computer servers. The space utilization for an Electronic Industries Association (EIA) rack is about 75 percent; that is, 75 percent of the available volume can be used for computer and related equipment. With 25 percent wasted space, there is great need for more efficient designs.

            Power Efficiency

A full-sized rack full of computer servers and associated hardware only works properly when the correct electrical current flows to all components. The power delivered to the servers follows a complex path — from delivery, to the rack, to power distribution to each server. Typically, each server must be designed with a power supply that equals the maximum power draw with all components active (CPU, memory, storage, network). If all the servers and associated equipment require 20 kilowatt-hour (kwh) to 25 kwh, then power distribution units installed within the rack must be able to accept this amount of power. The complete system will only work if the data center can supply this level of power to each rack. If that is not possible, either some servers will be removed, or they will be replaced with slower, less powerful units.

Next week’s article will explore Open Compute Solutions. If you prefer, you can download the complete ‘insideHPC Guide to Open Computing’  from the insideHPC White Paper Library courtesy of Penguin Computing.

Resource Links: