Information Technology Reference
In-Depth Information
increase in the energy consumption (from
E
IDLE
to
E
MAX
) when using an
increasing number of CPU cores. However, the model is only focused on the
energy consumption of the processor; it does not take into account the energy
consumption due to memory utilization and I/O devices. Thus, in this work we
aim at validating the energy consumption model and estimate deviations from
the previous model due to other energy consuming components.
In order to evaluate the model, three basic tests were executed using a server
from our HPC infrastructure at Universidad de la Republica. The server is an
HP Proliant DL385 G7 server with two AMD Opteron 6172 processors with 12
cores running at 2.1 GH, and 24 GB of RAM memory.
For the energy evaluation, a specialized Power Distribution Unit (PDU) was
used: CyberPower PDU20SWHVIEC8FNET. We connected only the server run-
ning the tests to the PDU, as it lacks the capability of per outlet measurement. A
specific application was developed to poll and log the energy consumption data,
due to a limitation on the granularity of the logging capabilities of the PDU,
which is only able to save log data at a rate of one measurement per minute.
The logging application was executed in a separated computer also connected
to the PDU, in order to avoid adding its own energy consumption to the mea-
surements. Using the logging application, we were able to log a minimum of four
and a mean of six instant energy measurements per second during each test.
The tests consist in executing an increasing number of applications in order
to use different number of cores, from a single core up to twenty four cores. The
applications used in the tests range from a simple mathematical operation to a
complex transformation, in order to evaluate different scenarios:
1.
Single loop
. The first test consists on running a simple C++ loop performing
a multiplication a huge number of times, this way ensuring a fully CPU-
bound test using only one CPU.
2.
LINPACK
. The second test is based on an open source sequential implemen-
tation of the LINPACK benchmark [6]. We adjusted the LINPACK parame-
ters to have an acceptable execution time while not using too much memory,
to reduce the race for cache and RAM memories when running 24 instances.
3.
Fast Fourier Transform
. This test is similar to the previous one, but based
on an open source implementation of the Fast Fourier Transform [4]. In this
case, the evaluation was made using only up to 23 instances of the test,
because the parameter setting resulted in an execution time for twenty-four
instances that doubled the twenty-three one, due to race for RAM memory.
In the tests, the energy consumption was estimated from the logs obtained
using the PDU by applying an interpolation of the instant power measurements.
The graphics in Figure 2 shows the energy consumption when using an increasing
number of cores for the three applications in the test (loop, LINPACK, and FFT,
respectively). The fourth graphic in Figure 2 is an example of the instant power
usage as function of time for the loop test case, where the execution of the tests
using an increasing number of cores were performed one after the other.
