Scalable Shared-Memory Multiprocessing

116 ^

System Performance Issues

protocol reduces

the

latency

of

communication

and

increases

the

utilization

of

the consum-

ing processors since they

do not

incur remote read misses.

Globally serializing sections

of

code

is a

very common occurrence

in

parallel programs,

especially when parallelism

is

exploited across loop iterations. After execution

of a

parallel

loop,

synchronization

is

achieved

by a

barrier, before continuing

on to the

next section

of

code

(see

Section 2.1

for an

example of barrier usage). Barriers consist

of

two phases.

In the

first phase, aU

of

the processes enter

the

barrier

and

wait

for the

last process

to

arrive.

In th ...