Data collection tools include the collect command, dbx collector command, and er_kernel command. Each of these tools is used to profile programs to collect data and create experiments that can be read by Performance Analyzer or er_print.
The following changes are common to these tools:
New processors are supported for hardware counters and stack unwind: SPARC T5, SPARC M5, SPARC M6, SPARC 64 X, SPARC 64 X+, Intel Ivy Bridge and Haswell.
Clock profiling is enabled by default even if hardware counter profiling is not specified.
Setting archive on is the same as setting archive copy. For collect and er_kernel this means –A on is now the same as –A copy . For dbx collector this means collector archive on is the same as collector archive copy.
The maximum number of threads on Oracle Solaris that can be profiled is now 32768.
The collect utility is a tool you use to profile your application as it runs to collect data and create an experiment that can be read by Performance Analyzer or er_print.
The collect utility is changed in this release as follows:
I/O tracing is supported with a new –i flag.
Java profiling is enabled by default whenever the target is a JVM. You no longer need to specify –j on.
–P option for collecting data from a running process is now supported on Linux systems. Note that this only works for single-threaded applications.
The –c option for collecting count data is now supported on Linux.
Hardware counter processing now supports multiple –h arguments and a default counter set. You can set the environment variable SP_COLLECTOR_HWC_DEFAULT to enable hardware counter profiling by default.
Hardware counter-based memoryspace profiling is enabled by default for precise counters on SPARC and x86 systems. The plus sign (+) is no longer required to capture memory addresses. See Memoryspace Profiling Improvements for more information.
collect -F =expr will no longer match the expression against the process lineage.
When using –P to attach to a Java program, be sure to specify –j on.
The dbx collector is a subcommand of the dbx debugger that you can use for performance data collection. See the collector(1) man page for more information.
In addition to the changes common to all data collection tools the dbx collector command is changed in this release as follows:
Another experiment can be started after a detach or experiment termination.
Attach is supported on Linux system for single-threaded native applications. For more information about limitations with attaching, see dbx attach Profiling (collect -P) in Oracle Solaris Studio 12.4: Release Notes .
dbx collector supports the following new commands:
collector iotrace on - Specifies to turn on I/O tracing.
collector duration - Specifies a time range for running an experiment.
collector java - Specifies whether to collect Java profile data. Default is off. You should use the command whenever profiling Java, whether by attaching or launching the JVM.
collector pausesig - Specifies a signal to be used for pausing or resuming data collection.
collector samplesig - Specifies a signal to be used to record a sample.
collector hwprofile addcounter - Specifies additional counters for hardware counter overflow profiling.
The er_kernel command profiles the Oracle Solaris kernel and generates an experiment that you can examine in Performance Analyzer or er_print.
In addition to the changes common to all data collection tools, the er_kernel utility is changed as follows:
Clock-profiling metrics in user sub-experiments are recorded as in collect experiments, but only User CPU Time and System CPU Time are recorded; wait times are not recorded.
The names of hardware counter metrics reported in the kernel founder experiment are prefaced by K_ . Metrics in user sub-experiments use the same names as metrics in collect experiments.
Kernel dataspace profiling is supported on Solaris SPARC, for precise counters, and for systems with version 1.8 of DTrace, or later.
Clock-profiling for the founder experiment is now recorded as in collect experiments, except they have only a Kernel CPU Time metric. Hardware counter profiling is not enabled by default.
Improved recording of kernel call stacks.