lkml.org 
[lkml]   [2010]   [Mar]   [3]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
From
Subject[PATCH 0/5] Add multiple event support to perf report V2
Date
Currently perf report is unable to handle multiple events even though
perf record is capable of capturing such a stream. For instance if we run
'perf record -e dtlb-misses -e cache-misses sleep 5'
today and then ask report to show the results we get:

# Samples: 18866
#
# Overhead Command Shared Object Symbol
# ........ ....... ................. ......
#
47.52% sleep [kernel.kallsyms] [k] page_fault
32.93% sleep [kernel.kallsyms] [k] locks_remove_posix
8.52% sleep [kernel.kallsyms] [k] acl_permission_check
3.46% sleep [kernel.kallsyms] [k] __wake_up_bit
3.31% sleep [kernel.kallsyms] [k] vfs_read
3.23% sleep [kernel.kallsyms] [k] kmem_cache_alloc
0.56% sleep [kernel.kallsyms] [k] get_page_from_freelist
0.15% sleep [kernel.kallsyms] [k] pipe_read
0.13% sleep [kernel.kallsyms] [k] pipe_lock_nested
0.13% sleep [kernel.kallsyms] [k] mm_alloc
0.06% sleep [kernel.kallsyms] [k] mm_init
#
# (For a higher level overview, try: perf report --sort comm,dso)
#
Worse than only reporting on one of the events, perf report has
summed the two events. i.e. There is no distinguishing between
dtlb-misses and cache-misses.

This set of patches addresses this problem and now when perf report
is run with the same event stream that produced the above report, the
output is this:

# Samples: 8620 cache-misses
#
# Overhead Command Shared Object Symbol
# ........ ....... ................. ......
#
72.06% sleep [kernel.kallsyms] [k] locks_remove_posix
18.64% sleep [kernel.kallsyms] [k] acl_permission_check
7.56% sleep [kernel.kallsyms] [k] __wake_up_bit
1.22% sleep [kernel.kallsyms] [k] get_page_from_freelist
0.34% sleep [kernel.kallsyms] [k] pipe_read
0.13% sleep [kernel.kallsyms] [k] mm_init
0.02% sleep [kernel.kallsyms] [k] pipe_lock_nested
0.01% sleep [kernel.kallsyms] [k] mm_alloc
0.01% sleep [kernel.kallsyms] [k] kmem_cache_alloc

# Samples: 10246 dTLB-load-misses
#
# Overhead Command Shared Object Symbol
# ........ ....... ................. ......
#
87.51% sleep [kernel.kallsyms] [k] page_fault
6.10% sleep [kernel.kallsyms] [k] vfs_read
5.94% sleep [kernel.kallsyms] [k] kmem_cache_alloc
0.22% sleep [kernel.kallsyms] [k] mm_alloc
0.22% sleep [kernel.kallsyms] [k] pipe_lock_nested

#
# (For a higher level overview, try: perf report --sort comm,dso)
#

Changlog:
From V1:
-Removed STREAM_ID from recorded attributes
-Only record SAMPLE_ID if more than one event is requested

Eric B Munson (5):
Add ID and to recorded event data when recording multiple events
Change add_hist_entry to take the tree root instead of session
Add storage for seperating event types in report
Change perf_session post processing functions to take histogram tree
Add multiple event support to perf report

tools/perf/builtin-annotate.c | 6 +-
tools/perf/builtin-diff.c | 12 +++--
tools/perf/builtin-record.c | 3 +
tools/perf/builtin-report.c | 111 ++++++++++++++++++++++++++++++++++++-----
tools/perf/util/event.h | 9 +++
tools/perf/util/hist.c | 45 +++++++++--------
tools/perf/util/hist.h | 12 +++--
tools/perf/util/session.c | 1 +
tools/perf/util/session.h | 1 +
9 files changed, 153 insertions(+), 47 deletions(-)


\
 
 \ /
  Last update: 2010-03-03 16:41    [from the cache]
©2003-2011 Jasper Spaans