Power profiling of data center equipment

Power-aware and thermal-aware techniques such as power-throttling and workload manipulation have been developed to counter the increasing power density in the current data centers. An Abstract Heat Flow Model (AHFM) has been established at the IMPACT Lab to predict the thermal profile of the data center. This enables on-line decision-making for thermal-aware job placement in the data cneter equipment (server/chassis). The basis for any such power-aware and/or thermalaware technique including the ones using AHFM, however, depends heavily on the equipment power consumption model assumed.

The goal of this project is to perform power profiling for computational equipments in the data centers based on experimental measurements, so that appropriate power-consumption model can be established. In this regard, power profiling of Dell PowerEdge 1855 and 1955 blade servers has been performed at the Arizona State University (ASU) data center for High Performance Computing Initiative (HPCI).

Gamut (Generic Application eMUlaTor) benchmark is used to vary the utilization of different resources (such as CPU utilization & Disk I/O). Additionally, four different compute intensive applications - matrix multiplication, convolution, digital image filtering, and non-linear physics application- have been used.

Power measurements of Dell PowerEdge 1855 and 1955 blade systems have been performed using the DUALCOM power meter from CyberSwitching Inc. The blade servers are powered from the chassis. The power meter is connected between the chassis and its power supply to measure the current drawn (in Amps) by the chassis (as shown in Figure 2). We used the SNMP based CSTools utilities supplied by CyberSwitching to retrieve the information from the power meter. The product of the current drawn with the supply voltage (208 V) of the ASU data center gives the power measurements (in Watts).

The power measurement is performed in three steps
1. empty chassis power measurement to observe power requirements to run the chassis unit.
2. chassis power measurement with only one server to observe power requirements of a single server.
3. full chassis power measurement to observe the aggregate power requirements of a fully loaded chassis.

We observe that the empty chassis power consumption for PowerEdge 1855 is 820W, whereas for PowerEdge 1955 it is 490W. When the server is idle (i.e. utilization is 0%), the total power consumption of the chassis is 600W and 940W for PowerEdge 1955 and 1855 respectively. The difference of their respective empty chassis power consumption from these values give the idle server power consumption as I(1955) = 600W - 490W = 110W, and I(1855) = 940W - 820W = 120W (I(1955) and I(1855) denote the idle server power consumption for PowerEdge 1955 and 1855 respectively). Similarly, at 100% CPU utilization for single server in the chassis, the total power consumption is T(1955) = 690W - 490W = 200W, and T(1855) = 990W - 820W = 170W.

It is further observed that Disk and Memory I/O does not affect the server power consumption, whereas the CPU utilization linearly affects the power consumption. If u denotes the utilization, then the power consumption can be approximated as P(1855, u) = 120 + 50u for PowerEdge 1855, and P(1955, u) = 110 + 90u for PowerEdge 1955. For example, at CPU utilization 50%, the chassis power consumption with a single server is for PowerEdge 1955 is 110 + 90*0.5 + I(1955) = 645W, which verifies the experimental results.

• T. Mukherjee, G. Varasamopoulos, S. K. S. Gupta, and Sanjay Rungta, ''Measurement based Power Profiling of Data Center Equipment", In the First International Worshop of Green Computing (in conjunction with CLUSTER 2007), Austin, USA, Sept, 2007. [PDF]