HP UX Developer Tools Getting Performance Advice, Specifying cpu Event Sets, Choosing CPU Events

Getting Performance Advice

$ caliper ecount myapp $ caliper fprof myapp $ caliper advise

collect and save initial data; generate advice on collected data and possible next steps

Specifying cpu Event Sets

(for cpu measurement; not available on Linux)

-m overview	broad overview of CPU (includes
	event sets: cpi, stall,
	dispersal, l1icache,
-m overview:kernel	l1dcache, l2cache, tlb, fp)
	restrict overview to execution in
-m brpath	kernel space
	branch types, path distributions,
-m brpred	branches per instructions, …
	branch prediction effectiveness
-m c2c	cache coherence activity
-m cpi	cycles per instruction
-m cpubus	central electronics complex
-m cspec	(CEC) traffic and CPU interaction
	control speculation effectiveness
-m dispersal	parallelism at instruction
-m dspec	dispersal
	data speculation effectiveness
-m fp	floating-point operation density,
	execution rate, and flush/trap
-m l1dcache	events density
	L1 data cache miss rates
-m l1icache	L1 instruction cache miss and
-m l2cache	prefetch information
	L2 unified cache miss rates
-m l2dcache	L2 data cache miss rates
-m l2icache	(dual-core Itanium 2 only)
	L2 instruction cache miss rates
-m l3cache	(dual-core Itanium 2 only)
	L3 unified cache miss rates
-m queues	bus request queue information
-m replay	cycles lost due to instruction
	replays (Intel® Itanium® 9500
-m stall	processors only)
	CPU performance limiters
-m sysbus	system bus utilization
-m threadswitch	impact of HyperThreading
-m tlb	(dual-core Itanium 2 only)
	translation lookaside buffer
	(TLB) misses

Specifying CPU Event Privilege Levels

(--scope process, the default, assumed)

-m CPU_CYCLES:user	count CPU_CYCLES events
-m CC:k	that occur in user space
	count CPU_CYCLES that occur
-m cpi:all	in kernel space
	record CPI for both user and
	kernel space (cpu
-s ,,CC:all	measurement)
	sample on CPU_CYCLES in
	both user and kernel space
--event-defaults kernel
--ev k	set default privilege level to
	kernel for all events

Choosing CPU Events

(summed across run for ecount, fprof, pmu_trace, {di}cache, {di}tlb, branch)

-m	event	count event events
-m	event,event2,…	count multiple events

Controlling Amount of Data Reported

--summary-cutoff 2

--su 2	in Function Summary section,
	only report functions responsible
--su 1,80	for 2 percent or more of data
	only report functions responsible
	for 1% or more of data; stop
	reporting functions when the
--detail-cutoff 10	cumulative percentage is at 80%
--de 10	in Function Details section,
	only report functions responsible
	for 10 percent or more of data
--report-details none
-r none	do not include Function Details
-r all	section in report
	show source statements and
	disassembly in Function Details
--context-lines 1,3
--co 1,3	report 1 extra source line before
	and after source lines with data;
	report 3 extra instructions before/
--process-cutoff 5	after each instruction with data
--process-cut 5	in Process Summary section,
	only report processes respon-
	sible for >=5 percent of data
--db-save	save module information so
	reports can be generated when
	modules are not present

--dcache-data-profile causes the dcache

measurement to create a data summary report that reports global variables and process regions associated with dcache misses

--report-hotpathsWhen true, adds hot call paths section to call graph reports

Finding Source Files, Libraries, Executables

(for use in symbol, source, and disassembly reporting)

--module-search-path dir1:dir2 search for libraries and executables in dir1, then dir2

--module-s dir1:dir2

search dir1, dir2 (short form)

--source-path-map dir1:dir2

--soudir1:dir2 search for source files in dir1, then dir2

--sou /work,/net/work

replace /work with /net/work in source file paths when searching for source files

Selecting Measurement Duration

-esecs collect data for secs seconds (default: collect until all measured processes exit or SIGINT)

Reporting Threads

--threads all report per-thread data; (default) --threadssum-allno per- thread data --threads merge merges data for threads having

same thread starter function name

Specifying Location of Kernel

(for use in symbol and disassembly reporting)

--kernel-path /tmp/vmlinux

--k /tmp/vmlinux kernel path is /tmp/vmlinux (unzipped copy of Linux kernel)

Collecting Memory Usage Data

(for any measurement; not available on Linux) --memory-usage begin:end

--mem b:e		measure at start and end of run
		measure at start/end (short form)
--mem timed:10s		measure every 10 seconds
--mem	timed:1m	measure every minute
--mem	all	measure once a second and at

start and end of run

Specifying HP Caliper Options in a Text File

-foptionsFile read HP Caliper options from the text file called optionsFile