Professional Documents
Culture Documents
This article was the first publica- The first characteristic of interest be strongly dependent on sweep-
tion by Gene Amdahl on what is the fraction of the computation- ing through the array along differ-
became known as Amdahl's Law. al load which is associated with ent axes on succeeding passes, etc.
Interestingly, it has no equations data management housekeeping. The effect of each of these compli-
and only a single figure. For this This fraction has been very nearly cations is very severe on any com-
issue of the SSCS News, Dr. constant for about ten years, and puter organization based on geo-
Amdahl agreed to redraw the fig- accounts for 40% of the executed metrically related processors in a
ure. In the available hard copy it instructions in production runs. In paralleled processing system. Even
was illegible. We print this his- an entirely dedicated special pur- the existence of regular rectangular
toric paper to enable members to pose environment this might be boundaries has the interesting
read the original source from reduced by a factor of two, but it is property that for spatial dimension
some 40 years ago. highly improbably that it could be of N there are 3N different point
The Editors reduced by a factor of three. The geometries to be dealt with in a
nature of this overhead appears to nearest neighbor computation. If
For over a decade prophets have be sequential so that it is unlikely the second nearest neighbor were
voiced the contention that the to be amenable to parallel process- also involved, there would be 5N
organization of a single computer ing techniques. Overhead alone different point geometries to con-
has reached its limits and that truly would then place an upper limit tend with. An irregular boundary
significant advances can be made on throughput of five to seven compounds this problem as does
only by interconnection of a multi- times the sequential processing an inhomogeneous interior. Com-
plicity of computers in such a rate, even if the housekeeping putations which are dependent on
manner as to permit cooperative were done in a separate processor. the states of variables would
solution. Variously the proper The non-housekeeping part of the require the processing at each
direction has been pointed out as problem could exploit at most a point to consume approximately
general purpose computers with a processor of performance three to the same computational time as
generalized interconnection of four times the performance of the the sum of computations of all
memories, or as specialized com- housekeeping processor. A fairly physical effects within a large
puters with geometrically related obvious conclusion which can be region. Differences or changes in
memory interconnections and con- drawn at this point is that the effort propagation rates may affect the
trolled by one or more instruction expended on achieving high paral- mesh point relationships.
streams. lel processing rates is wasted Ideally the computation of the
Demonstration is made of the unless it is accompanied by action of the neighboring points
continued validity of the single achievements in sequential pro- upon the point under considera-
processor approach and of the cessing rates of very nearly the tion involves their values at a pre-
weaknesses of the multiple proces- same magnitude. vious time proportional to the
sor approach in terms of applica- Data management housekeep- mesh spacing and inversely pro-
tion to real problems and their ing is not the only problem to portional to the propagation rate.
attendant irregularities. plague oversimplified approaches Since the time step is normally
The arguments presented are to high speed computation. The kept constant, a faster propagation
based on statistical characteristics physical problems which are of rate for some effects would imply
of computation on computers over practical interest tend to have interactions with more distant
the last decade and upon the oper- rather significant complications. points. Finally, the fairly common
ational requirements within prob- Examples of these complications practice of sweeping through the
lems of physical interest. An addi- are as follows: Boundaries are like- mesh along different axes on suc-
tional reference will be one of the ly to be irregular; interiors are like- ceeding passes poses problems of
most thorough analyses of relative ly to be inhomogeneous; computa- data management which affects all
computer capabilities currently tions required may be dependent processors; however, it affects geo-
published- "Changes in Computer on the states of the variables at metrically related processors more
Performance," Datamation, Sep- each point; propagation rates of severely by requiring transposing
tember 1966, Professor Kenneth E. different physical effects may be all points in storage in addition to
Knight, Stanford School of Busi- quite different; the rate of conver- the revised input-output schedul-
ness Administration. gence, or convergence at all, may ing. A realistic assessment of the
The diagram above illustrating “Amdahl’s Law” shows that a highly parallel machine has a harder time delivering a fair
fraction of its peak performance due to the sequential component of the given computation and the overhead of coor-
dination (e.g. synchronization) between the processors. Assuming a fixed sized problem, Amdahl speculated that most
programs would require at least 25% of the computation to be sequential (only one instruction executing at a time),
with overhead due to interprocessor coordination averaging 10%. The curves show that the more you depend on par-
allelism for performance, the slower the system is likely to be in the probable case, 65%. The lowest curve (A) repre-
sents the 32-wide SIMD processor, and the top curve (C) is for the modified vector processor. Scaled problems reduce
the sequential component and the coordination overhead to a negligible level, making large numbers of processors very
efficient in those cases. Justin Rattner, Intel Senior Fellow, justin.rattner@intel.com, July 2007.