State of the Art and Future Trends of Datacenter Networks

State of the Art and Future Trends of Datacenter Networks

George Michelogiannakis (Lawrence Berkeley National Laboratory, USA)
Copyright: © 2015 |Pages: 10
DOI: 10.4018/978-1-4666-5888-2.ch181
OnDemand PDF Download:
No Current Special Offers

Chapter Preview



Physical Organization

Datacenters are organized in racks. Each rack typically contains 42 vertical 44.45 milimeter U slots (Nathan Farrington, 2009). Each U slot can currently hold 2- to 4-socket processor motherboards, or parts of a network switch (router) instead. With this configuration, racks are 0.6 meters wide, 1 meter deep and 2 meters high. Racks themselves are typically organized in rows. Each rack has a cold side where air enters for internal cooling and a hot side where hot air exits. The rack’s cooling system must evacuate the heat generated by the processor sockets, DRAM and networking equipment. Racks are placed such as to form cold and hot isles, to assist the critical problem of cooling in datacenters. Cooling systems must be designed to accommodate the worst-case power consumption at 100% utilization. Cold rows are approximately 1.22 meters and allow human access to blades but not the cables, whereas hot rows are 0.9 meters, contain cables, and are the key to the datacenter’s heat extraction strategy. This configuration is illustrated in Figure 1.

Figure 1.

Cold air enters racks from cold row for cooling and exits to the hot rows. The building’s cooling system maintains airflow and supplies cold air.


Cooling is a primary consideration of the entire building the datacenter is housed in. Because cooling and other systems in the building consume power, power usage effectiveness (PUE) was devised as a metric of efficiency and equals the ratio of a datacenter’s total power to the power actually used by the computing equipment. The average datacenter PUE is 2.0, and the most efficient 1.2 (Google Inc.).

Key Terms in this Chapter

Load Balance: Using all available resources in a similar manner and maintaining roughly equal utilization factor for all of them.

Pwh: Peta (10 15 ) watt-hours.

CAGR (Compound Annual Growth Rate): The year-over-year growth rate of an investment or other metric over a specified period of time.

Hop: A channel traversal, including the injection and ejection channels.

Buffer: Storage space at routers to store packets or parts of packets that cannot make progress because they are waiting for resources, typically output ports.

Path Diversity: Having multiple disjoint paths between any source-destination pair.

Locality: Maintaining a short distance between communication pairs by proper task placement or other means.

Flow: A sequence of packets typically from the same application between a given source to a specific destination.

Workload: The communication and computation an application causes (the work to be completed by the datacenter). Applications can be running at the same time in multiple processing cores.

Pin-Limited: A chip is pin-limited if the constraint towards providing more communication bandwidth is the number of pins (connectors to the socket) it can have due to technology constraints.

Complete Chapter List

Search this Book: