CVPERF(1) CVPERF(1)
cvperf - WorkShop Performance View
cvperf {<experiment-record> | <pixie-counts-file>}
cvperf presents a graphical display of the performance data gathered in a
WorkShop Performance experiment whose experiment directory is given by
<experiment-record>.
cvperf can also present the display of performance information from a
pixie counts file given by <pixie-counts-file>. To use this feature, you
will need to install a patch with a later version of pixie than is on the
current released system. Contact Customer Support to obtain the patch.
The WorkShop Performance tools can record a number of different
experiments, each of which provides one or more metrics of performance.
Each has its strengths and weaknesses; the experiment types are
described in the CASEVision/WorkShop User's Guide, and are briefly
described below.
A ``Bottleneck'' experiment measures the performance of the program by
sampling the active process' callstack every 100 milliseconds, and using
the callstack data to attribute exclusive total time to the function at
the bottom of each callstack (i.e., the function being executed at the
time of the sample), and to attribute inclusive total time to all the
functions above the one currently being executed. In addition, this
experiment takes a caliper-sample every second, so that the user can
examine the data for any interval of execution.
A ``Total Time'' experiment uses the same statistical callstack sampling
to compute performance metrics, but does not take caliper-samples, except
as the user specifies them.
A ``CPU Time'' experiment instruments the program to count function calls
(both direct, and through function pointers), as well as do statistical
PC sampling on the resultant instrumented executable. The PC sample data
is used to compute exclusive CPU time for each function, and the call
information is used to propagate the exclusive time to the callers of
each routine.
An ``Ideal Time'' experiment generates the same data as a a pixie-fied
executable. It instruments the program to maintain precise basic block
counts, and uses a machine model to convert the block execution counts
into an idealized exclusive time. This time will be less than the real
time that any run would take, as it excludes any pipeline delays across
basic blocks, and does not take any cache or memory latencies into
account. Inclusive time is propagated to callers based on the call
information, just as it is for ``CPU Time'' experiments.
Page 1
CVPERF(1) CVPERF(1)
An ``I/O Trace'' experiment traces each of the I/O calls made by the
executable, and computes inclusive and exclusive read and write counts,
and can show a time-based event chart for all calls with a given filedescriptor.
A ``System Call Trace'' experiment traces all calls, and can show a
time-based event chart for all calls, as well as compute read and write
counts. This experiment is a super-set of the ``I/O Trace'' experiment.
A ``Page Fault Trace'' experiment computes inclusive and exclusive page
fault counts, and also shows an event line with the time-based sequence
of page faults.
A ``Find Memory Leaks'' experiment traces all calls to malloc, free,
etc., and computes inclusive and exclusive memory allocations and leaks,
and will also show any errors in calling these routines. For more
information, see the malloc_cv(3) man page for more information.
A ``Floating Point Exception'' trace experiment traces all such
exceptions, and computes inclusive and exclusive counts of these
exceptions, as well as an event line showing the time sequence of the
exceptions.
A ``PC Sampling Time'' experiment uses a statistical counting of where
the process PC is as a function of time, and computes exclusive time from
that data. There is no information in the data that allows computing of
inclusive times.
A ``Custom Task'' can be run by choosing the particular instrumentation
and data desired. Its use is discouraged, because not all combinations
will give meaningful results. If there is a custom task that you desire,
please tell us, and we will endeavor to make a standard one for it.
The Performance Analyzer main window has a Task component, a Function
List component, a per thread User Time/System Time State Chart component,
and a TimeLine component.
The Task Component- simply states the task or objective that the user
selected for this experiment.
The Function List Component- is an annotated list of all the functions in
the program. The annotations available depend on the experiment task;
annotations can be enabled or disabled by bringing up the Preferences
Dialog from the Config menu.
The State Chart Component- allows the user to quickly guess at the
bottleneck resources in the different phases of the run, and also to see
thread balance information. For some experiments, the State Chart is
replaced by an event line for the primary events being traced. For
``Find Memory Leaks'' experiment, it is replaced by a plot of process
address space and resident size as a function of time.
Page 2
CVPERF(1) CVPERF(1)
The TimeLine Component- displays the experiment as a set of events over
time, and provides calipers to allow the user to specify an interval of
interest.
There are also a number additional views, some of which are applicable to
all experiments, and some of which are meaningful only for some
experiment types. The additional views are:
Usage View (Graphs)- presents a graphical display of the process resource
usage data in the form of stripcharts and event charts.
Usage View (Numerical)- presents a textual display of the process and
system wide resource usage data.
IO View- presents a per file descriptor chart of the number of bytes
transferred. This view is available only for I/O tracing or System call
tracing experiments.
Call Graph View- presents the target program as nodes and arcs.
Butterfly View- shows the caller and callee relationships for any
selected function.
Leak View- presents a list of all of the memory leaks in the target
program. A leak is any region that is malloc'd, but not free'd. The
leaks are aggregated by common callstack. This view is available only
for Memory Leak experiments.
Malloc View- presents a list of all of the memory allocation in the
target program, aggregated by common callstack. This view is available
only for Memory Leak experiments.
Malloc Error View- presents a list of all of the memory allocation errors
in the target program, aggregated by common callstack. This view is
available only for Memory Leak experiments.
Heap View- presents a graphical display of use of the heap by the target
program. This view is available only for Memory Leak experiments.
Call Stack View- presents the callstack at the event currently selected
in the TimeLine component, or in any event chart for tracing experiments.
Working Set View- presents a list of all of the DSO's in the program,
with information on the efficiency of use of the text (instruction)
pages. This view is available only for Ideal time experiments.
Source View- presents an annotated Source View, and is brought up by
double clicking on any function in the function list, on any node in the
Call Graph View, or on any frame in the Call Stack View. If cvperf is
being run in a session with the debugger, cvd, also running, the
annotations will appear in the debugger's main source window. If there
is no debugger window running, a separate Source View will be used.
Page 3
CVPERF(1) CVPERF(1)
Disassembly View- presents an annotated Disassembled View of the
currently selected function.
cvd(1), cvmeter(1), on-line help for cvperf.
PPPPaaaaggggeeee 4444 [ Back ]
|
