Hardware Reference
In-Depth Information
aging temperature. Temperature-aware workload placement is one optimization that has
been proposed to manage temperature to reduce cooling costs. The idea is to identify the
cooling proile of a given room and map the hoter systems to the cooler spots, so that at
the WSC level the requirements for overall cooling are reduced.
a. [5] The coefficient of performance (COP) of a CRAC unit is defined as the ratio of heat
removed (Q) to the amount of work necessary (W) to remove that heat. The COP of a
CRAC unit increases with the temperature of the air the CRAC unit pushes into the
plenum. If air returns to the CRAC unit at 20 degrees Celsius and we remove 10KW of
heat with a COP of 1.9, how much energy do we expend in the CRAC unit? If cooling
the same volume of air, but now returning at 25 degrees Celsius, takes a COP of 3.1,
how much energy do we expend in the CRAC unit now?
b. [5] Assume a workload distribution algorithm is able to match the hot workloads well
with the cool spots to allow the computer room air-conditioning (CRAC) unit to be
run at higher temperature to improve cooling efficiencies like in the exercise above.
What is the power savings between the two cases described above?
c. [Discussion] Given the scale of WSC systems, power management can be a complex,
multifaceted problem. Optimizations to improve energy efficiency can be implemen-
ted in hardware and in software, at the system level, and at the cluster level for the IT
equipment or the cooling equipment, etc. It is important to consider these interactions
when designing an overall energy-efficiency solution for the WSC. Consider a consol-
idation algorithm that looks at server utilization and consolidates different workload
classes on the same server to increase server utilization (this can potentially have the
server operating at higher energy efficiency if the system is not energy proportional).
How would this optimization interact with a concurrent algorithm that tried to use
different power states (see ACPI, Advanced Configuration Power Interface, for some
examples)? What other examples can you think of where multiple optimizations can
potentially conflict with one another in a WSC? How would you solve this problem?
6.29 [5/10/15/20] <6.2> Energy proportionality (sometimes also referred to as energy scale-
down) is the atribute of the system to consume no power when idle, but more importantly
gradually consume more power in proportion to the activity level and work done. In this
exercise, we will examine the sensitivity of energy consumption to different energy pro-
portionality models. In the exercises below, unless otherwise mentioned, use the data in
Figure 6.4
as the default.
a. [5] A simple way to reason about energy proportionality is to assume linearity
between activity and power usage. Using just the peak power and idle power data
from
Figure 6.4
and a linear interpolation, plot the energy-efficiency trends across
varying activities. (Energy eiciency is expressed as performance per wat.) What hap-
pens if idle power (at 0% activity) is half of what is assumed in
Figure 6.4
?
What hap-
pens if idle power is zero?
b. [10] Plot the energy-efficiency trends across varying activities, but use the data from
column 3 of
Figure 6.4
for power variation. Plot the energy efficiency assuming that
the idle power (alone) is half of what is assumed in
Figure 6.4
.
Compare these plots
with the linear model in the previous exercise. What conclusions can you draw about
the consequences of focusing purely on idle power alone?
c. [15] Assume the system utilization mix in column 7 of
Figure 6.4
.
For simplicity, as-
sume a discrete distribution across 1000 servers, with 109 servers at 0% utilization, 80
Search WWH ::
Custom Search