Parallelism - Scalability and Amdhal's Law
Scalability of a system, from a performance engineering
point of view, is the ability of a system to use additional resources available
to it in a judicious manner and maintain the performance parameters within
acceptable limits. Load testing can be used to determine whether a system is
able to progressively use additional hardware available to it and maintain
Non-Functional Requirement (NFR) metrics at a constant level under increased
user load. A performance test engineer can determine whether his software is
scaling well by looking at how three parameters of the system behave, namely,
Resource Utilization, Throughput and Response Time.
Ideally, with an increase in user load, the system should be
able to progressively increase Resource Utilization. A graph which plots
Resource Utilization vs. User Load should have a positive slope. Likewise, for
a scalable application, Throughput should also increase with an increase in
user load. On the other hand, the response time should remain more or less
constant; adhering to the NFR. This graph will ideally have a slope of zero.
But in a real world scenario, a deviation of upto 15 percentage is considered
acceptable. Here is an excellent article on Scalability from MSDN.
Today's software is radically distributed and the need for
these to be intrinsically scalable is important. Design of software for
parallel performance, i.e. scalability is determined by the percentage of code
that can be parallelized. No software is fully scalable, i.e., there will
always be a small amount of code which cannot be parallelized. Amdahl's Law
states that
Speedup = 1 / (s + p / N)
where N is the number of processors, s is the amount
of time spent (by a serial processor) on serial parts of a program and p
is the amount of time spent (by a serial processor) on parts of the program
that can be done in parallel.
