Hardware Reference
In-Depth Information
27.
For each topology shown in Fig. 8-37, determine the degree of fault tolerance each one
has, defined as the maximum number of links that can be lost without partitioning the
network in two.
28.
Consider the double-torus topology of Fig. 8-37(f) but expanded to a size of
k
×
k
.
What is the diameter of the network? (
Hint
: Consider odd
k
and even
k
separately).
29.
An interconnection network is in the form of an 8
8 cube. Each link has a full-
duplex bandwidth of 1 GB/sec. What is the bisection bandwidth of the network?
30.
Amdahl's law limits the potential speed-up achievable on a parallel computer. Com-
pute, as a function of
f
, the maximum possible speed-up as the number of CPUs ap-
proaches infinity. What are the implications of this limit for
f
×
8
×
0. 1?
31.
Figure 8-51 shows how scaling fails with a bus but succeeds with a grid. Assuming
that each bus or link has a bandwidth
b
, compute the average bandwidth per CPU for
each of the four cases. Then scale each system to 64 CPUs and repeat the calculations.
What is the limit as the number of CPUs goes to infinity?
=
32.
In the text, three variations of
send
were discussed: synchronous, blocking, and
nonblocking. Give a fourth method that is similar to a blocking
send
but has slightly
different properties. Give an advantage and a disadvantage of your method as com-
pared to blocking
send
.
33.
Consider a multicomputer running on a network with hardware broadcasting, such as
Ethernet. Why does the ratio of read operations (those not updating internal state vari-
ables) to write operations (those updating internal state variables) matter?
