Warehouse-Scale Computers to Exploit Request-Level and Data-Level Parallelism - Computer Architecture: A Quantitative Approach - page 490

c. [Discussion] <6.2> Consider the system with the best JouleSort rating from your ex-
periments above. How would you improve the energy efficiency? For example, try re-
writing the sort code to improve the JouleSort rating.
6.36 [10/10/15] <6.1, 6.2> Figure 6.1 is a listing of outages in an array of servers. When dealing
with the large scale of WSCs, it is important to balance cluster design and software archi-
tectures to achieve the required uptime without incurring significant costs. This question
explores the implications of achieving availability through hardware only.
a. [10] <6.1, 6.2> Assuming that an operator wishes to achieve 95% availability through
server hardware improvements alone, how many events of each type would have to
be reduced? For now, assume that individual server crashes are completely handled
through redundant machines.
b. [10] <6.1, 6.2> How does the answer to part (a) change if the individual server crashes
are handled by redundancy 50% of the time? 20% of the time? None of the time?
c. [15] <6.1, 6.2> Discuss the importance of software redundancy to achieving a high level
of availability. If a WSC operator considered buying machines that were cheaper, but
10% less reliable, what implications would that have on the software architecture?
What are the challenges associated with software redundancy?
6.37 [15] <6.1, 6.8> Look up the current prices of standard DDR3 DRAM versus DDR3 DRAM
that has error-correcting code (ECC). What is the increase in price per bit for achieving the
higher reliability that ECC provides? Using the DRAM prices alone, and the data provided
in Section 6.8 , what is the uptime per dollar of a WSC with non-ECC versus ECC DRAM?
6.38 [5/Discussion] <6.1> WSC Reliability and Manageability Concerns :
a. [5] Consider a cluster of servers costing $2000 each. Assuming an annual failure rate
of 5%, an average of an hour of service time per repair, and replacement parts requir-
ing 10% of the system cost per failure, what is the annual maintenance cost per server?
Assume an hourly rate of $100 per hour for a service technician.
b. [Discussion] Comment on the differences between this manageability model versus
that in a traditional enterprise datacenter with a large number of small or medium-
sized applications each running on its own dedicated hardware infrastructure.
1 T his chapter is based on material from the topic The Datacenter as a Computer: An Introduction to the Design of
Warehouse-Scale Machines , by Luiz André Barroso and Urs Hölzle of Google [2009] ; the blog Perspectives at mvd-
irona.com and the talks “Cloud-Computing Economies of Scale” and “Data Center Networks Are in My Way,” by
James Hamilton of Amazon Web Services [2009 , 2010 ]; and the technical report Above the Clouds: A Berkeley View of
Cloud Computing , by Michael Armbrust et al. [2009] .