Warehouse-Scale Computers to Exploit Request-Level and Data-Level Parallelism - Computer Architecture: A Quantitative Approach - page 481

b. [10] Discuss some software-based optimizations that can utilize the new level of the
memory hierarchy.
c. [25] Repeat part (a), instead assuming that each node has a 32 GB PCIe card that is able
to cache 50% of all disk accesses.
d. [15] As discussed in “Fallacies and Pitfalls” ( Section 6.8 ) , replacing all disks with
SSDs is not necessarily a cost-effective strategy. Consider a WSC operator that uses
it to provide cloud services. Discuss some scenarios where using SSDs or other Flash
memory would make sense.
6.18 [20/20/Discussion] <6.3> Memory Hierarchy : Caching is heavily used in some WSC
designs to reduce latency, and there are multiple caching options to satisfy varying access
patterns and requirements.
a. [20] Let's consider the design options for streaming rich media from the Web (e.g.,
Netflix First we need to estimate the number of movies, number of encode formats
per movie, and concurrent viewing users. In 2010, Netflix had 12,000 titles for online
streaming, each title having at least four encode formats (at 500, 1000, 1600, and 2200
kbps). Let's assume that there are 100,000 concurrent viewers for the entire site, and an
average movie is one hour long. Estimate the total storage capacity, I/O and network
bandwidths, and video-streaming-related computation requirements.
b. [20] What are the access paterns and reference locality characteristics per user, per
movie, and across all movies? ( Hint : Random versus sequential, good versus poor
temporal and spatial locality, relatively small versus large working set size.)
c. [Discussion] What movie storage options exist by using DRAM, SSD, and hard drives?
Compare them in performance and TCO.
6.19 [10/20/20/Discussion/Discussion] <6.3> Consider a social networking Web site with 100
million active users posting updates about themselves (in text and pictures) as well as
browsing and interacting with updates in their social networks. To provide low latency,
Facebook and many other Web sites use memcached as a caching layer before the backend
storage/database tiers.
a. [10] Estimate the data generation and request rates per user and across the entire site.
b. [20] For the social networking Web site discussed here, how much DRAM is needed
to host its working set? Using servers each having 96 GB DRAM, estimate how many
local versus remote memory accesses are needed to generate a user's home page?
c. [20] Now consider two candidate memcached server designs, one using conventional
Xeon processors and the other using smaller cores, such as Atom processors. Given
that memcached requires large physical memory but has low CPU utilization, what
are the pros and cons of these two designs?
d. [Discussion] Today's tight coupling between memory modules and processors often
requires an increase in CPU socket count in order to provide large memory support.
List other designs to provide large physical memory without proportionally increas-
ing the number of sockets in a server. Compare them based on performance, power,
costs, and reliability.
e. [Discussion] The same user's information can be stored in both the memcached and
storage servers, and such servers can be physically hosted in different ways. Discuss
the pros and cons of the following server layout in the WSC: (1) memcached collocated
on the same storage server, (2) memcached and storage server on separate nodes in the