lkml.org 
[lkml]   [2012]   [Aug]   [8]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
From
SubjectRe: [PATCH 11/11] perf scripts python: Add event_analyzing_sample.py as a sample for general event handling
Date
Hi Arnaldo and Ingo,

On Wed, 8 Aug 2012 14:13:48 -0300, Arnaldo Carvalho de Melo wrote:
> From: Feng Tang <feng.tang@intel.com>
>
> Currently only trace point events are supported in perf/python script,
> the first 3 patches of this serie add the support for all types of
> events. This script is just a simple sample to show how to gather the
> basic information of the events and analyze them.
>
> This script will create one object for each event sample and insert them
> into a table in a database, then leverage the simple SQL commands to
> sort/group them. User can modify or write their brand new functions
> according to their specific requirment.
>
> Here is the sample of how to use the script:
>
> $ perf record -a tree
> $ perf script -s process_event.py

Please edit the script name to event_analyzing_sample.py at least to
prevent future confusion. For other issues, please see my review
comments on Feng's original posts. (They can be incremental.)

>
> There is 100 records in gen_events table
> Statistics about the general events grouped by thread/symbol/dso:
>
> comm number histgram
> ==========================================
> swapper 56 ######
> tree 20 #####
> perf 10 ####
> sshd 8 ####
> kworker/7:2 4 ###
> ksoftirqd/7 1 #
> plugin-containe 1 #
>
> symbol number histgram
> ==========================================================
> native_write_msr_safe 40 ######
> __lock_acquire 8 ####
> ftrace_graph_caller 4 ###
> prepare_ftrace_return 4 ###
> intel_idle 3 ##
> native_sched_clock 3 ##
> Unknown_symbol 2 ##
> do_softirq 2 ##
> lock_release 2 ##
> lock_release_holdtime 2 ##
> trace_graph_entry 2 ##
> _IO_putc 1 #
> __d_lookup_rcu 1 #
> __do_fault 1 #
> __schedule 1 #
> _raw_spin_lock 1 #
> delay_tsc 1 #
> generic_exec_single 1 #
> generic_fillattr 1 #
>
> dso number histgram
> ==================================================================
> [kernel.kallsyms] 95 #######
> /lib/libc-2.12.1.so 5 ###
>
> Signed-off-by: Feng Tang <feng.tang@intel.com>
> Cc: Andi Kleen <andi@firstfloor.org>
> Cc: David Ahern <dsahern@gmail.com>
> Cc: Ingo Molnar <mingo@elte.hu>
> Cc: Peter Zijlstra <peterz@infradead.org>
> Cc: Robert Richter <robert.richter@amd.com>
> Cc: Stephane Eranian <eranian@google.com>
> Link: http://lkml.kernel.org/r/1344419875-21665-6-git-send-email-feng.tang@intel.com
> Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
> ---
> .../perf/scripts/python/event_analyzing_sample.py | 193 ++++++++++++++++++++
> 1 file changed, 193 insertions(+)
> create mode 100644 tools/perf/scripts/python/event_analyzing_sample.py
>
> diff --git a/tools/perf/scripts/python/event_analyzing_sample.py b/tools/perf/scripts/python/event_analyzing_sample.py
> new file mode 100644
> index 0000000..46f05aa
> --- /dev/null
> +++ b/tools/perf/scripts/python/event_analyzing_sample.py
> @@ -0,0 +1,193 @@
> +# process_event.py: general event handler in python

Hopefully here also.

Thanks,
Namhyung


> +#
> +# Current perf report is alreay very powerful with the anotation integrated,
> +# and this script is not trying to be as powerful as perf report, but
> +# providing end user/developer a flexible way to analyze the events other
> +# than trace points.
> +#
> +# The 2 database related functions in this script just show how to gather
> +# the basic information, and users can modify and write their own functions
> +# according to their specific requirment.
> +#
> +# The first sample "show_general_events" just does a baisc grouping for all
> +# generic events with the help of sqlite, and the 2nd one "show_pebs_ll" is
> +# for a x86 HW PMU event: PEBS with load latency data.
> +#
> +
> +import os
> +import sys
> +import math
> +import struct
> +import sqlite3
> +
> +sys.path.append(os.environ['PERF_EXEC_PATH'] + \
> + '/scripts/python/Perf-Trace-Util/lib/Perf/Trace')
> +
> +from perf_trace_context import *
> +from EventClass import *
> +
> +#
> +# If the perf.data has a big number of samples, then the insert operation
> +# will be very time consuming (about 10+ minutes for 10000 samples) if the
> +# .db database is on disk. Move the .db file to RAM based FS to speedup
> +# the handling, which will cut the time down to several seconds.
> +#
> +con = sqlite3.connect("/dev/shm/perf.db")
> +con.isolation_level = None
> +
> +def trace_begin():
> + print "In trace_begin:\n"
> +
> + #
> + # Will create several tables at the start, pebs_ll is for PEBS data with
> + # load latency info, while gen_events is for general event.
> + #
> + con.execute("""
> + create table if not exists gen_events (
> + name text,
> + symbol text,
> + comm text,
> + dso text
> + );""")
> + con.execute("""
> + create table if not exists pebs_ll (
> + name text,
> + symbol text,
> + comm text,
> + dso text,
> + flags integer,
> + ip integer,
> + status integer,
> + dse integer,
> + dla integer,
> + lat integer
> + );""")
> +
> +#
> +# Create and insert event object to a database so that user could
> +# do more analysis with simple database commands.
> +#
> +def process_event(param_dict):
> + event_attr = param_dict["attr"]
> + sample = param_dict["sample"]
> + raw_buf = param_dict["raw_buf"]
> + comm = param_dict["comm"]
> + name = param_dict["ev_name"]
> +
> + # Symbol and dso info are not always resolved
> + if (param_dict.has_key("dso")):
> + dso = param_dict["dso"]
> + else:
> + dso = "Unknown_dso"
> +
> + if (param_dict.has_key("symbol")):
> + symbol = param_dict["symbol"]
> + else:
> + symbol = "Unknown_symbol"
> +
> + # Creat the event object and insert it to the right table in database
> + event = create_event(name, comm, dso, symbol, raw_buf)
> + insert_db(event)
> +
> +def insert_db(event):
> + if event.ev_type == EVTYPE_GENERIC:
> + con.execute("insert into gen_events values(?, ?, ?, ?)",
> + (event.name, event.symbol, event.comm, event.dso))
> + elif event.ev_type == EVTYPE_PEBS_LL:
> + event.ip &= 0x7fffffffffffffff
> + event.dla &= 0x7fffffffffffffff
> + con.execute("insert into pebs_ll values (?, ?, ?, ?, ?, ?, ?, ?, ?, ?)",
> + (event.name, event.symbol, event.comm, event.dso, event.flags,
> + event.ip, event.status, event.dse, event.dla, event.lat))
> +
> +def trace_end():
> + print "In trace_end:\n"
> + # We show the basic info for the 2 type of event classes
> + show_general_events()
> + show_pebs_ll()
> + con.close()
> +
> +#
> +# As the event number may be very big, so we can't use linear way
> +# to show the histgram in real number, but use a log2 algorithm.
> +#
> +
> +def num2sym(num):
> + # Each number will have at least one '#'
> + snum = '#' * (int)(math.log(num, 2) + 1)
> + return snum
> +
> +def show_general_events():
> +
> + # Check the total record number in the table
> + count = con.execute("select count(*) from gen_events")
> + for t in count:
> + print "There is %d records in gen_events table" % t[0]
> + if t[0] == 0:
> + return
> +
> + print "Statistics about the general events grouped by thread/symbol/dso: \n"
> +
> + # Group by thread
> + commq = con.execute("select comm, count(comm) from gen_events group by comm order by -count(comm)")
> + print "\n%16s %8s %16s\n%s" % ("comm", "number", "histgram", "="*42)
> + for row in commq:
> + print "%16s %8d %s" % (row[0], row[1], num2sym(row[1]))
> +
> + # Group by symbol
> + print "\n%32s %8s %16s\n%s" % ("symbol", "number", "histgram", "="*58)
> + symbolq = con.execute("select symbol, count(symbol) from gen_events group by symbol order by -count(symbol)")
> + for row in symbolq:
> + print "%32s %8d %s" % (row[0], row[1], num2sym(row[1]))
> +
> + # Group by dso
> + print "\n%40s %8s %16s\n%s" % ("dso", "number", "histgram", "="*74)
> + dsoq = con.execute("select dso, count(dso) from gen_events group by dso order by -count(dso)")
> + for row in dsoq:
> + print "%40s %8d %s" % (row[0], row[1], num2sym(row[1]))
> +
> +#
> +# This function just shows the basic info, and we could do more with the
> +# data in the tables, like checking the function parameters when some
> +# big latency events happen.
> +#
> +def show_pebs_ll():
> +
> + count = con.execute("select count(*) from pebs_ll")
> + for t in count:
> + print "There is %d records in pebs_ll table" % t[0]
> + if t[0] == 0:
> + return
> +
> + print "Statistics about the PEBS Load Latency events grouped by thread/symbol/dse/latency: \n"
> +
> + # Group by thread
> + commq = con.execute("select comm, count(comm) from pebs_ll group by comm order by -count(comm)")
> + print "\n%16s %8s %16s\n%s" % ("comm", "number", "histgram", "="*42)
> + for row in commq:
> + print "%16s %8d %s" % (row[0], row[1], num2sym(row[1]))
> +
> + # Group by symbol
> + print "\n%32s %8s %16s\n%s" % ("symbol", "number", "histgram", "="*58)
> + symbolq = con.execute("select symbol, count(symbol) from pebs_ll group by symbol order by -count(symbol)")
> + for row in symbolq:
> + print "%32s %8d %s" % (row[0], row[1], num2sym(row[1]))
> +
> + # Group by dse
> + dseq = con.execute("select dse, count(dse) from pebs_ll group by dse order by -count(dse)")
> + print "\n%32s %8s %16s\n%s" % ("dse", "number", "histgram", "="*58)
> + for row in dseq:
> + print "%32s %8d %s" % (row[0], row[1], num2sym(row[1]))
> +
> + # Group by latency
> + latq = con.execute("select lat, count(lat) from pebs_ll group by lat order by lat")
> + print "\n%32s %8s %16s\n%s" % ("latency", "number", "histgram", "="*58)
> + for row in latq:
> + print "%32s %8d %s" % (row[0], row[1], num2sym(row[1]))
> +
> +def trace_unhandled(event_name, context, event_fields_dict):
> + print ' '.join(['%s=%s'%(k,str(v))for k,v in sorted(event_fields_dict.items())])
> +
> +def print_header(event_name, cpu, secs, nsecs, pid, comm):
> + print "%-20s %5u %05u.%09u %8u %-20s " % \
> + (event_name, cpu, secs, nsecs, pid, comm),


\
 
 \ /
  Last update: 2012-08-09 05:41    [W:0.114 / U:0.036 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site