.\"
.\" mono manual page.
-.\" (C) 2003 Ximian, Inc.
-.\" (C) 2004-2005 Novell, Inc.
+.\" Copyright 2003 Ximian, Inc.
+.\" Copyright 2004-2009 Novell, Inc.
.\" Author:
.\" Miguel de Icaza (miguel@gnu.org)
.\"
.PP
The
.I logging profiler
-will eventually replace the default profiler as it is more complete
-and encompasses the functionality of all the other profilers for Mono.
-It supports the following execution modes:
+is a general purpose profiler that can track many different kinds of
+events and logs those into a file as the program executes. This is
+different than previous profilers in Mono that kept the information in
+memory and rendered a summary of the results when the program shut
+down.
+.PP
+Using the logging profiler means that useful information can be
+gathered for long-running applications, applications that terminate
+abormally (crashes, segfaults, complicated tear down processes) or
+when no data about the shutdown is required.
+.PP
+The data collected by the running threads is kept independent of each
+other to minimize the runtime overhead and the information dumped into
+the log at regular intervals.
+.PP
+A sample use is very simple:
+.nf
+ $ mono --profile=logging program.exe
+
+ $ mprof-decoder program.mprof
+.fi
+.PP
+In the above example the logging profiler is used in its default mode
+that merely records GC statistics for the execution of program.exe.
+The profiling data collected is stored in the file program.mprof. The
+mprof-decoder tool is then used to analyze the data.
+.PP
+You can instruct the logging profiler to record different one or more
+sets of events. These are the modes supported:
.IP
-.I Statistical:
-the program instruction pointer is periodically sampled (it works also with
-unmanaged functions). If call chains are requested, for each sample the
-profiler gets a partial stack trace (up to a desired depth) so that
+.I Statistical Profiling (stat)
+the program instruction pointer is periodically sampled to determine
+where the program is spending most of its time. Statistical
+profiling has a very low impact on a running application and it is
+very useful to get a general picture of where time is being spent on a
+program.
+.IP
+If call chains are requested, for each sample the profiler gets a
+partial stack trace (limited by the specified depth) so that
caller-callee information is available.
.IP
.I Instrumenting:
method, load and unload for assemblies, modules and and individual classes,
and appdomain and thread creation and destruction.
.PP
-Instead of reporting the collected
-information at the end of the execution of the program, this profiler logs
-all the events periodically into a file during program execution.
-To minimize the performance impact with multi-threaded applications,
-the logging uses per-thread buffers that are routinely saved to disk.
-.PP
-The output file contains compressed events, to process the data you should
-use tools like the "Mono.Profiler" tool provided on the Mono SVN
-repository.
-.PP
This profiler is activated passing the \fB--profile=logging\fR option to
the mono runtime, and is controlled attaching further options, like
-\fB--profile=logging:stat\fR for doing statistical profiling (multiple
+\fB--profile=logging:statistical\fR for doing statistical profiling (multiple
options are separated by commas).
.PP
As a quick primer, here are a few examples of the most common usage modes:
.PP
-To write the resulting data to "mydata.mprof" (defaults to statistical
-profiling):
+To perform statistical profiling:
.nf
- mono --profile=logging:o=mydata.mprof program.exe
+ mono --profile=logging:stat program.exe
.fi
.PP
To perform statistical profiling, inspecting call chains up to depth 8:
.nf
- mono --profile=logging:s=8 program.exe
+ mono --profile=logging:stat=8 program.exe
.fi
.PP
-To profile allocations with caller method attribution:
+To profile allocations (by default the call stack will be analized for
+each allocation, producing detailed caller method attribution infornation):
.nf
- mono --profile=logging:a,ts program.exe
+ mono --profile=logging:allocations program.exe
.fi
.PP
-To profile garbage collection activity (collection time and objects freed
-at each collection):
+To profile garbage collection activity at a high level (collection time and objects freed
+at each collection for each class are reported, but heap snapshots are not saved to disk):
.nf
- mono --profile=logging:g,as program.exe
+ mono --profile=logging:allocations-summary program.exe
+.fi
+.PP
+To perform heap profiling taking heap snapshots:
+.nf
+
+ mono --profile=logging:heap=all program.exe
+.fi
+.PP
+To write the resulting data to a different file:
+.nf
+
+ mono --profile=logging:output=mydata.mprof program.exe
.fi
.PP
Then you would need to invoke the decoder \fImprof-decoder(1)\fR
-on the output file to see the profiling results.
+on the output file to see the profiling results, or to examine heap
+snapshots and allocations in detail \fImprof-heap-viewer(1)\fR.
+.PP
+The operating modes described above are the default ones, and are sufficient
+to use the profiler.
.PP
-These are all the available oprions, organized by category:
+To further customize the profiler behavior there are more options, described
+below.
+.PP
+These options can be individually enabled and disabled prefixing them
+with an (optional) '+' character or a '-' character. For instance,
+the "allocations" option by default records also the full call stack
+at each allocation. If only the caller is wanted, one should use
+"allocations,-save-allocation-stack", or to disable call tracking
+completely (making the profiler less intrusive)
+"allocations,-save-allocation-caller,-save-allocation-stack". In
+practice the "allocation" option by default behaves like
+"allocations,save-allocation-caller,save-allocation-stack", but the
+user can tune this to his needs.
+.PP
+These are all the available options, organized by category:
.PP
\fBExecution profiling modes\fR
.RS
\fIjit\fR, \fIj\fR
Collect information about time spent by the JIT engine compiling
methods.
-.TP
-\fItrack-stack\fR, \fIts\fR
-Track the execution stack. By itself this option does nothing, but it
-enables more detailed reporting in other options (because the profiler
-will use the stack data).
.ne
.RE
.PP
.TP
\fIallocations\fR, \fIalloc\fR or \fIa\fR
Collect information about each allocation (object class and size).
-If combined with the "ts" option, for each allocation the profiler will
-also show the responsible method (and also correctly dintinguish allocations
-performed during JIT time).
-These information are also available when combining the "a" and "c" options,
-but since "c" is much more intrusive "ts" should be preferred.
+By default this also implies "+save-allocation-caller" and
+"+save-allocation-stack".
+.TP
+\fIsave-allocation-caller\fR, \fIsac\fR
+Save the direct caller of each allocation. The profiler filters out wrapper
+methods, and also recognizes if the allocation has been performed by the
+runtime while jitting a method.
.TP
\fIsave-allocation-stack\fR, \fIsas\fR
-Save the execution stack at each allocation. While the "ts" option saves the
-direct caller, this one records the whole stack trace. Note that for technical
-reasons the decoder will correctly ignore wrapper methods only using the data
-saved by the "ts" option (the stack traces still contain all the wrappers).
-Anyway the "ts" and "sas" options can be combined, and the decoder will
+Save the full managed execution stack at each allocation.
+While the "sac" option saves the direct caller, this one records the whole
+stack trace.
+Note that in the call stack the wrapper methods are not filtered out.
+Anyway the "sac" and "sas" options can be combined, and the decoder will
attribute the allocation to the correct method even if the wrapper is at the
top of the stack trace.
.TP
\fIgc\fR or \fIg\fR
Measure the time spent in each collection, and also trace heap resizes.
.TP
-\fIheap-shot\fR, \fIheap\fR or \fIh\fR
+\fIheap-shot[=ARG]\fR, \fIheap[=ARG]\fR or \fIh[=ARH]\fR
Performs full heap profiling. In this case on each
collection a full heap snapshot is recorded to disk.
Inside the snapshots, each object reference is still represented so
that it's possible to investigate who is responsible for keeping objects
alive.
.PP
-The actual production of heap snapshots could produce large log files,
-so it can be controlled in three ways:
+If the value of ARG is
+.B all,
+a heap snapshot is taken at each collection.
+.PP
+If the value is signal name or a number (one of SIGUSR1, SIGUSR2 and
+SIGPROF, remembering that SIGPROF is not compatible with the
+statistical profiler), every time the application receives this signal
+a garbage collection is forced and a snapshot is taken. The natural way
+to take a heap snapshot is then using the kill command to deliver the
+appropriate signal to the application.
+.PP
+Heap profiling also enables full allocation profiling (with call
+stacks), and each allocation can be related to its corresponding
+object in the snapshots, enabling investigations like "find all
+objects of a given class allocated by a given method and still live at
+a given collection, and then find all objects referencing them".
+.PP
+This kind of heap snapshot analysis is performed using the mprof-heap-viewer(1)
+application.
+.PP
+The number of heap snapshots taken (and the moment in which they are taken)
+can be further customized with the following options:
.TP
\fIgc-dumps=N\fR, \fIgc-d=N\fR, \fIgcd=N\fR
states the number of snapshots that must be dumped (since the application
-starts). Zero means no dumps at all, -1 means dump at all collections.
-.TP
-\fIgc-signal=<signal>\fR, \fIgc-s\fR or \fIgcs\fR
-(where <signal> is one of "SIGUSR1", "SIGUSR2", or "SIGPROF")
-specifies a signal that will immediately trigger a collection and a dump.
+starts). Zero means no dumps at all, -1 means dump at all collections.
.TP
\fIgc-commands=FILE\fR, \fIgc-c=FILE\fR or \fIgcc=FILE\fR
specify a "command file". The file must contain an integer value in ASCII
onwards.
If the file is present at application startup it takes precedence over an
eventual \fIgcd=N\fR option.
+.TP
+These options exist because it can happen that the user wants to investigate
+what happens during collections but without forcing a collection using a
+signal, because forcing a collection alters the program behavior.
+Of course it is possible to simply take a snapshot at every collection, but
+in some workloads this is could not be feasible (too much data).
+So we have this "garbage collection dumps" counter to control how many
+snapshots to take.
.ne
.RE
.PP
.ne
.RE
.PP
-In its current state, this profiler can also perform heap analysis (like
-heap-shot), and the decoder is already able to read the data, however
-the user interface for this feature is experimental (the
-\fImprof-heap-viewer\fR tool in the mono-tools module).
+In its current state, this profiler can also perform heap analysis
+like the HeapShot profiler, but there is no UI to process this
+information.
.PP
Another known issue is that if the timer is not strictly monotonic (like
rtdsc), differences between times can underflow (they are handled as
loses track of the execution stack and misattributes the caller for a few
allocations (and method execution time).
.PP
+The output file contains compressed events, to process the data you should
+use tools like the "Mono.Profiler" tool provided on the Mono SVN
+repository.
+.PP
More explanations are provided here: "http://www.mono-project.com/LoggingProfiler".
.SH EXTERNAL PROFILERS
There are a number of external profilers that have been developed for
and the application might crash or terminate at any given point
afterwards.
.PP
-The \fB--debug=casts\FR option can be used to get more detailed
+The \fB--debug=casts\fR option can be used to get more detailed
information for Invalid Cast operations, it will provide information
about the types involved.
.PP
\fBno-gdb-backtrace\fR
This option will disable the GDB backtrace emitted by the runtime
after a SIGSEGV or SIGABRT in unmanaged code.
+.TP
+\fBsuspend-on-sigsegv\fR
+
+This option will suspend the program when a native SIGSEGV is received.
+This is useful for debugging crashes which do not happen under gdb,
+since a live process contains more information than a core file.
.ne
.RE
.TP
above locations. If you don't want the mapping to be performed you can set this
variable in your environment before starting the application and no action will
be taken.
+.TP
+\fBMONO_MESSAGING_PROVIDER\fR
+Mono supports a plugin model for its implementation of System.Messaging making
+it possible to support a variety of messaging implementations (e.g. AMQP, ActiveMQ).
+To specify which messaging implementation is to be used the evironement variable
+needs to be set to the full class name for the provider. E.g. to use the RabbitMQ based
+AMQP implementation the variable should be set to:
+
+.nf
+Mono.Messaging.RabbitMQ.RabbitMQMessagingProvider,Mono.Messaging.RabbitMQ
.SH ENVIRONMENT VARIABLES FOR DEBUGGING
.TP
\fBMONO_ASPNET_NODELETE\fR
http://www.mono-project.com
.SH SEE ALSO
.PP
-certmgr(1), csharp(1), mcs(1), mdb(1), monocov(1), monodis(1), mono-config(5), mozroots(1), xsp(1).
+certmgr(1), csharp(1), mcs(1), mdb(1), monocov(1), monodis(1),
+mono-config(5), mozroots(1), pdb2mdb(1), xsp(1).
.PP
For more information on AOT:
http://www.mono-project.com/AOT
projects / mono.git / blobdiff