caliper.1 (2010 09)

c
caliper(1)
Integrity Systems
caliper(1)
(Requires Optional HP Caliper Software)
Alternatively, you can choose the type of data collected by assigning
arc-stride
or dcache to
the environment variable,
PBO_DATA_TYPE
.
--report
This option also supports these report types: cgprof
, fcount, and fcover.
--system-usage=
[all]|[runstatus][:io][:syscalls]
Controls the collection and reporting of system usage data. Three types of system usage data can be
collected:
runstatus (how much time each process spent running, eligible to run but not running,
and waiting), io (total size and time of logical and physical I/O done by each process/thread),
sys-
calls (the count and time spent in every syscall called by a process), or
all (the default).
If this option is not used, both
runstatus
and syscalls are collected by default. Use the none
option to prevent collection of this data.
This measurement can be made in conjunction with any Caliper measurement. This option is only
available on HP-UX 11i v2 or later.
--user-regions=default
|rum-sum
For runs involving the Itanium PMU, specifies whether the data should be collected for the entire
run (default), or only in regions delimited by the PMU enable/disable instructions (
rum-sum).
For more information, see Limiting PMU Measurements below.
When attaching to a process to perform
cgprof, fcount,orfcover measurements, the dependent
shared libraries of the program must be mapped as private before you can attach to the process. You can
enable private mapping of the shared libraries by using the chatr command with the +dbg enable
option on the program file (see chatr (1)). When attaching to a process for cgprof
, fcount,orfcover
runs, caliper will remain attached to the target process until it exits (see also
--duration=SECONDS).
Stopping
caliper with the use of SIGINT (e.g., Ctrl-C in a terminal window) for cgprof, fcount
,
and
fcover runs will result in all processes being forcibly terminated after Caliper generates a perfor-
mance report or writes data to a database.
Linux
The following option is available only on Linux:
--overflow-block=TRUE
|FALSE
Specifies whether the target application should be blocked when the PMU sampling buffer is full.
The default is TRUE (in other words, the target application will be blocked until caliper has com-
pleted processing all the samples in the buffer). This option is valid only for PMU based per-process
measurements.
CPU Metrics EVENT_SET Description
CPU Metrics measurement type requires HP-UX 11i V2 September 2004 OE (B.11.23.0409) or later. You
can specify the event sets and sampling period with the
--metrics and --sampling-spec
options,
respectively.
The
--cpu-aggregation=COUNT option specifies how many samples will be aggregated into one
aggregated sample. You can measure multiple event sets in the same run. By default, the overview
metrics consisting of the following 8 event sets will be measured: cpi, stall, dispersal, l1icache,
l1dcache, l2cache, tlb, fp.
Example:
caliper cpu -o cpu.txt program
This will run program , measuring and reporting the following metrics by taking one sample every 8 mil-
liseconds:
cpi, stall, dispersal, l1icache, l1dcache, l2cache, tlb, fp. By default 125
low-level samples will be aggregated into one user-reported sample resulting in one aggregated sample
collected per second. The result is saved in the text file, cpu.txt.
You can specify one or more comma separated list of predefined event sets. The following event sets are
available.
overview is the default.
brpath Provides information on the dynamic mix of branch types, branch path distribution,
branch per instruction, etc.
brpred Provides metrics that are useful in assessing the effectiveness of branch prediction.
20 Hewlett-Packard Company − 20 − HP-UX 11i Version 3: September 2010