Metric Lists

Many of the er_print commands use a list of metric keywords. The syntax of the list is:

metric-keyword-1[:metric-keyword2…]

For dynamic metrics, those based on measured data, a metric keyword consists of three parts: a metric flavor string, a metric visibility string, and a metric name string. These are joined with no spaces, as follows.

flavorvisibilityname

For static metrics, those based on the static properties of the load objects in the experiment (name, address, and size), a metric keyword consists of a metric name, optionally preceded by a metric visibility string, joined with no spaces:

[visibility]name

The metric flavor and metric visibility strings are composed of flavor and visibility characters.

The allowed metric flavor characters are listed in Table 8. A metric keyword that contains more than one flavor character is expanded into a list of metric keywords. For example, ie.user is expanded into i.user:e.user.

Table 8  Metric Flavor Characters

Character Description e Show exclusive metric value i Show inclusive metric value

The allowed metric visibility characters are listed in Table 9. The order of the visibility characters in the visibility string does not affect the order in which the corresponding metrics are displayed. For example, both i%.user and i.%user are interpreted as i.user:i%user.

Metrics that differ only in the visibility are always displayed together in the standard order. If two metric keywords that differ only in the visibility are separated by some other keywords, the metrics appear in the standard order at the position of the first of the two metrics.

Table 9  Metric Visibility Characters

Character Description . Show metric as a time. Applies to timing metrics and hardware counter metrics that measure cycle counts. Interpreted as “+” for other metrics. % Show metric as a percentage of the total program metric. For attributed metrics in the callers-callees list, show metric as a percentage of the inclusive metric for the selected function. + Show metric as an absolute value. For hardware counters, this value is the event count. Interpreted as a “.” for timing metrics. ! Do not show any metric value. Cannot be used in combination with other visibility characters.

When both flavor and visibility strings have more than one character, the flavor is expanded first. Thus, ie.%user is expanded to i.%user:e.%user, which is then interpreted as i.user:i%user:e.user:e%user.

For static metrics, the visibility characters period (.), plus (+), and percent sign (%), are equivalent for the purposes of defining the sort order. Thus, sort i%user, sort i.user, and sort i+user all mean that Performance Analyzer should sort by inclusive user CPU time if it is visible in any form. sort i!user means Performance Analyzer should sort by inclusive user CPU time, regardless of whether it is visible.

You can use the visibility character exclamation point (!) to override the built-in visibility defaults for each flavor of metric.

If the same metric appears multiple times in the metric list, only the first appearance is processed and subsequent appearances are ignored. If the named metric is not on the list, it is appended to the list.

Table 10 lists the available er_print metric name strings for timing metrics, synchronization delay metrics, memory allocation metrics, MPI tracing metrics, and the two common hardware counter metrics. For other hardware counter metrics, the metric name string is the same as the counter name. You can display a list of all the available metric name strings for the loaded experiments with the metric_list command. To list the counter names, issue the collect -h command with no additional arguments. See Hardware Counter Profiling Data for more information on hardware counters.

Table 10  Metric Name Strings

Category String Description Clock profiling metrics total Total Thread Time totalcpu Total CPU Time user User CPU Time system System CPU Time trap Trap CPU Time lock User Lock Time datapfault Data-Page Fault Time textpfault Text-Page Fault Time kernelpfault Kernel Page Fault Time stop Stopped Time wait CPU wait time sleep Sleep Time Hardware counter metrics insts Instructions issued Available on all supported systems. cycles CPU cycles Available on most supported systems. In addition, each processor has its own set of counters. Use collect -h to see the full list of counters for your system. CPI Cycles per instruction, calculated from cycles and insts metrics. Available only if both counters are recorded. IPC Instructions per cycle, calculated from cycles and insts metrics. Available only if both counters are recorded. OpenMP profiling metrics ompwork Time spent doing work either serially or in parallel ompwait Time spent when OpenMP runtime is waiting for synchronization masterthread Master Thread Time is the total time spent in the master thread. It is only available from Oracle Solaris experiments. It corresponds to wall-clock time. Synchronization delay metrics sync Synchronization wait time syncn Synchronization wait count Heap tracing metrics heapalloccnt Number of allocations heapallocbytes Bytes allocated heapleakcnt Number of leaks heapleakbytes Bytes leaked I/O tracing metrics ioreadbytes Bytes Read ioreadcnt Read Count ioreadtime Read Time iowritebytes Bytes Written iowritecnt Write Count iowritetime Write Time ioothrcnt Other IO Count ioothertime Other IO Time ioerrornt IO Error Count ioerrortime IO Error Time Thread Analyzer metrics raccesses Data race accesses deadlocks Deadlocks MPI tracing metrics mpitime Time spent in MPI calls mpisend Number of MPI point-to-point sends started mpibytessent Number of bytes in MPI Sends mpireceive Number of MPI point‐to‐point receives completed mpibytesrecv Number of bytes in MPI Receives mpiother Number of calls to other MPI functions MPI profiling metrics mpiwork Time spent inside the MPI runtime doing work, such as processing requests or messages mpiwait Time spent inside the MPI runtime but waiting for an event, buffer, or message

In addition to the name strings listed in Table 10, two name strings can be used only in default metrics lists. These are hwc, which matches any hardware counter name, and any, which matches any metric name string. Also note that cycles and insts are common to SPARC^® platforms and x86 platforms, but other flavors also exist that are architecture-specific.

To see the metrics available from the experiments you have loaded, issue the metric_list command.