Warehouse-Scale Computers to Exploit Request-Level and Data-Level Parallelism - Computer Architecture: A Quantitative Approach - page 489

6.34 [15/Discussion] <6.4> Some recent studies have defined a metric called TPUE, which
stands for “true PUE” or “total PUE.” TPUE is defined as PUE * SPUE. PUE, the power
utilization effectiveness, is defined in Section 6.4 as the ratio of the total facility power over
the total IT equipment power. SPUE, or server PUE, is a new metric analogous to PUE, but
instead applied to computing equipment, and is defined as the ratio of total server input
power to its useful power, where useful power is defined as the power consumed by the
electronic components directly involved in the computation: motherboard, disks, CPUs,
DRAM, I/O cards, and so on. In other words, the SPUE metric captures inefficiencies asso-
ciated with the power supplies, voltage regulators, and fans housed on a server.
a. [15] <6.4> Consider a design that uses a higher supply temperature for the CRAC
units. The efficiency of the CRAC unit is approximately a quadratic function of the
temperature, and this design therefore improves the overall PUE, let's assume by 7%.
(Assume baseline PUE of 1.7.) However, the higher temperature at the server level
triggers the on-board fan controller to operate the fan at much higher speeds. The fan
power is a cubic function of speed, and the increased fan speed leads to a degradation
of SPUE. Assume a fan power model:
where ns is the normalized fan speed = fan speed in rpm/18,000
and a baseline server power of 350 W. Compute the SPUE if the fan speed increases
from (1) 10,000 rpm to 12,500 rpm and (2) 10,000 rpm to 18,000 rpm. Compare the PUE
and TPUE in both these cases. (For simplicity, ignore the inefficiencies with power de-
livery in the SPUE model.)
b. [Discussion] Part (a) illustrates that, while PUE is an excellent metric to capture the
overhead of the facility, it does not capture the inefficiencies within the IT equipment
itself. Can you identify another design where TPUE is potentially lower than PUE?
( Hint : See Exercise 6.26.)
6.35 [Discussion/30/Discussion] <6.2> Two recently released benchmarks provide a good
starting point for energy-efficiency accounting in servers—the SPECpower_ssj2008 bench-
mark (available at htp://www.spec.org/power_ssj2008/ ) and the JouleSort metric (available at
http://sortbenchmark.org/ ).
a. [Discussion] <6.2> Look up the descriptions of the two benchmarks. How are they sim-
ilar? How are they different? What would you do to improve these benchmarks to bet-
ter address the goal of improving WSC energy eiciency?
b. [30] <6.2> JouleSort measures the total system energy to perform an out-of-core sort
and atempts to derive a metric that enables the comparison of systems ranging from
embedded devices to supercomputers. Look up the description of the JouleSort metric
at htp://sortbenchmark.org . Download a publicly available version of the sort algorithm
and run it on different classes of machines—a laptop, a PC, a mobile phone, etc.—or
with different configurations. What can you learn from the JouleSort ratings for difer-
ent setups?