From: Kan Liang <kan.li...@intel.com>
The patch series intends to fix the severe performance issue in
Knights Landing/Mill, when monitoring in heavy load system.
perf top costs a few minutes to show the result, which is
unacceptable.
With the patch series applied, the latency will reduces to
several seconds.
machine__synthesize_threads and perf_top__mmap_read costs most of
the perf top time (> 99%).
Patch 1-4 do the optimization for machine__synthesize_threads.
Patch 5-6 does the optimization for perf_top__mmap_read.
Optimization for machine__synthesize_threads
- Multithreading the whole process.
- The threads number is set to the max online CPU# by default.
User can change the threads number through the new option.
- Introduces hashtable for machine threads to reduce the lock
contention.
- The optimization can also benefit other platforms and other
perf tools, like perf record. But this patch series doesn't
do the optimization for other tools. It can be done later
separately.
- With this optimization applied, there is a 1.56x speedup in
Knights Mill with heavy workload.
Optimization for perf_top__mmap_read
- switch to backward overwrite mode
For non overwrite mode, it tries to read everything in the ring buffer
and does not check the messup. Once there are lots of samples delivered
shortly, the processing time could be very long.
Considering the real time requirement for perf top, it should switch
to backward overwrite mode.
- With this optimization applied, there is a 8.98x speedup in
Knights Mill with heavy workload.
- However, the latency of perf_top__mmap_read is still higher than the
default perf top fresh time (2s) in Knights Mill with heavy workload.
A check is introduced to give some hints to reduce the overhead.
The source code is also available at
https://github.com/kliang2/perf.git perf_top_opt
Here are perf top latency test result on Knights Mill and Skylake server
The heavy workload is to compile Linux kernel as below
"sudo nice make -j$(grep -c '^processor' /proc/cpuinfo)"
Then, "sudo perf top"
The latency period is the time between perf top launched and the first
profiling result shown.
- Latency on Knights Mill (272 CPUs)
Original(s) With patch(s) Speedup
272.68 40.89 6.67x
- Latency on Skylake server (192 CPUs)
Original(s) With patch(s) Speedup
12.28 2.05 5.99x
Changes since V3:
- Switch to backward overwrite mode (jirka)
The backward mode can avoid some problems found in forward,
but the performance drops which compared with V3 for Knights Mill.
Introduce a new patch to check the latency of perf_top__mmap_read.
- Handle the thread_nr = 1 specially (jirka)
Changes since V2:
- patches 1 and 2 for V2 for hashtable and scandir have been merged.
- patches 3, 4 and 7 for V2 are droped. Because the optimization code
doesn't touch those codes. The protection is not needed. (Arnaldo & jirka)
- Using mutex wrappers for multithread only lock. (jirka)
- Move struct synthesize_threads_arg to event.c (jirka)
Changes since V1:
- Patch 1: machine threads and hashtable related renaming (Arnaldo)
- Patch 6: use a smaller locked section for comm_str__put
add a locked wrapper for comm_str__findnew (Arnaldo)
Kan Liang (6):
perf tools: lock to protect namespaces and comm list
perf tools: lock to protect comm_str rb tree
perf top: implement multithreading for perf_event__synthesize_threads
perf top: add option to set the number of thread for event synthesize
perf top: switch to backward overwrite mode
perf top: check the cost of perf_top__mmap_read
tools/perf/Documentation/perf-top.txt | 3 +
tools/perf/builtin-kvm.c | 3 +-
tools/perf/builtin-record.c | 2 +-
tools/perf/builtin-top.c | 43 ++++++---
tools/perf/builtin-trace.c | 2 +-
tools/perf/tests/mmap-thread-lookup.c | 2 +-
tools/perf/ui/browsers/hists.c | 12 ++-
tools/perf/util/comm.c | 18 +++-
tools/perf/util/event.c | 163 +++++++++++++++++++++++++++-------
tools/perf/util/event.h | 3 +-
tools/perf/util/machine.c | 8 +-
tools/perf/util/machine.h | 9 +-
tools/perf/util/thread.c | 53 +++++++++--
tools/perf/util/thread.h | 3 +
tools/perf/util/top.h | 1 +
15 files changed, 262 insertions(+), 63 deletions(-)
--
2.5.5
