An object of type CpuTimeLimit may be used to "limit" the CPU time
taken by a computation: if the computation takes too long then an
exception (of type CoCoA::TimeoutException) is thrown.
When creating a CpuTimeLimit object you must specify a time limit
in seconds as a positive double: e.g. CpuTimeLimit CheckTime(10).
You must tell the CheckTime object explicitly when it should check whether
the time limit has been reached by calling CheckTime(), i.e. by calling
its member function operator(). If the time limit has been reached,
it throws an exception of type CoCoA::TimeoutException (derived from
CoCoA::ErrorInfo); otherwise the call does nothing (other than the check).
The typical use is with a potentially long loop. Just before the loop
one creates the CpuTimeLimit object, then at the start of each
iteration inside the loop one calls operator().
IMPORTANT CoCoALib checks for timeout only when the member
function CpuTimeLimit::operator() is called; so CoCoALib
will not notice that time-out has occurred between successive calls
to operator().
It is possible to use a single CpuTimeLimit object for several
loops, but then it is best to call myPrepareForNewLoop just before
entering each new loop; the variability of the iterations of that loop
can be specified.
An optional second parameter to the ctor specifies the "variability" of
time between successive iterations: e.g. if the checks are in a loop
where each iteration takes more or less the same time then the variability
is low (IterationVariability::low); if the iterations can vary greatly in
computation time then the variability is high (IterationVariability::high);
by default the variability is medium (IterationVariability::medium).
There is one real constructor, and one pseudo-constructor:
CpuTimeLimit(seconds) where seconds is a positive double; the measurement of CPU use begins immediately; there is an upper limit of one million seconds.
CpuTimeLimit(seconds, variability) as above, but specify how variable time between successive checks might be
NoCpuTimeLimit() returns a CpuTimeLimit object which has infinite timeout
Variability should be: IterationVariability::low if successive iterations take more or less the same time;
IterationVariability::high if successive iterations can take widely differing amounts of time. The default
is IterationVariability::medium which indicates some sort of compromise.
Let CheckForTimeout be an object of type CpuTimeLimit.
There are two operations:
CheckForTimeout.IamTimedOut() -- returns true if time-out has occurred, otherwise false
CheckForTimeout(context) -- does nothing unless timeout has occurred, when it throws a TimeoutException object; context is a string literal which is copied into the "context" field of the exception
CheckForTimeout.myPrepareForNewLoop() -- if the same CpuTimeLimit object is to be used inside more than one loop, then call this before every loop (except the first one)
CheckForTimeout.myPrepareForNewLoop(v) -- like myPrepareForNewLoop but also specify a "variability" for the new loop
IsUnlimited(CheckForTimeout) -- return true iff CheckForTimeout was created by NoCpuTimeLimit
There is one class for exceptions:
TimeoutException(context) -- the arg context is a string literal indicating where the time-out was detected (usually it is a good idea to use the name of the function which was interrupted)
The class TimeoutException derives from ErrorInfo.
This is the fourth design. The first was based on SIGVTALRM, but it was not
clear how portable that would be. The second was based on CheckForInterrupt,
but the calls to CpuTime were too costly (and it depended on a global
variable). The third design was based on ProgressReporter: it assumed
that the times between successive clock checks do not vary too much.
This new fourth design revises the third, and lets the caller specify
the "variability" of time between successive checks.
The idea is to check the actual cpu time only occasionally, and not every
time operator() is called. It uses a similar strategy to that of
ProgressReporter; based on the variability, a "proper check" is
performed only every few times operator() is called. The number of
"ignored" calls uses an estimate of how many iters can safely be
performed before the next "proper check" -- this assumes that calls
to operator() occur at fairly regular intervals.
The private data field myInterval has a special role if its value is
negative: it means that the CpuTimeLimit object has infinite time-out,
so should never check cpu usage.
Perhaps offer a version which uses only elapsed time? This should be easy to implement!
Inconvenient having to pass CpuTimeLimit as explicit parameters;
but how else to do this in a threadsafe way?
A normal call to CpuTime() may not work as desired in a multithreaded
context. It is not clear how to solve this portably.
2019