github-repos/benchmark
1
0
Fork 0
mirror of https://github.com/google/benchmark.git synced 2025-02-26 11:20:28 +08:00
Code Issues Packages Projects Releases Wiki Activity

Allow for collection of more PMUs than are physically available.

Browse Source 
Normally we can only collect as many PMUs as we have physical PMU
counters because they are managed as a single pinned group and extra
requested PMUs will not be collected. With this patch we instead collect
each PMU as a separate group once this limit is hit using PMU
multiplexing.
This commit is contained in:
Jamison Collins 2024-01-12 19:10:42 +00:00
parent ea71a14891
commit 2d609d4fa1
2 changed files with 23 additions and 15 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

34
src/perf_counters.cc
View File

@ -86,6 +86,12 @@ PerfCounters PerfCounters::Create(
Initialize(); Initialize();
} }
// Initially try to create a single group which holds all PMUs since that
// has lower overhead. Group multiplexing doesn't work on all platforms so if
// that fails try again without using groups.
bool group_pmus = true;
retry_without_groups:
// Valid counters will populate these arrays but we start empty // Valid counters will populate these arrays but we start empty
std::vector<std::string> valid_names; std::vector<std::string> valid_names;
std::vector<int> counter_ids; std::vector<int> counter_ids;
@ -96,11 +102,6 @@ PerfCounters PerfCounters::Create(
counter_ids.reserve(counter_names.size()); counter_ids.reserve(counter_names.size());
const int kCounterMode = PFM_PLM3; // user mode only const int kCounterMode = PFM_PLM3; // user mode only
// Group leads will be assigned on demand. The idea is that once we cannot
// create a counter descriptor, the reason is that this group has maxed out
// so we set the group_id again to -1 and retry - giving the algorithm a
// chance to create a new group leader to hold the next set of counters.
int group_id = -1; int group_id = -1;
// Loop through all performance counters // Loop through all performance counters
@ -150,7 +151,7 @@ PerfCounters PerfCounters::Create(
// case. // case.
attr.disabled = is_first; attr.disabled = is_first;
attr.inherit = true; attr.inherit = true;
attr.pinned = is_first; attr.pinned = group_pmus;
attr.exclude_kernel = true; attr.exclude_kernel = true;
attr.exclude_user = false; attr.exclude_user = false;
attr.exclude_hv = true; attr.exclude_hv = true;
@ -171,10 +172,14 @@ PerfCounters PerfCounters::Create(
} }
if (id < 0) { if (id < 0) {
// If the file descriptor is negative we might have reached a limit // If the file descriptor is negative we might have reached a limit
// in the current group. Set the group_id to -1 and retry // in the current group. Let's try again without groups.
if (group_id >= 0) { if (group_id >= 0 && group_pmus) {
// Create a new group // Close all performance counters
group_id = -1; for (int close_id : counter_ids) {
::close(close_id);
}
group_pmus = false;
goto retry_without_groups;
} else { } else {
// At this point we have already retried to set a new group id and // At this point we have already retried to set a new group id and
// failed. We then give up. // failed. We then give up.
@ -197,8 +202,10 @@ PerfCounters PerfCounters::Create(
if (group_id < 0) { if (group_id < 0) {
// This is a leader, store and assign it to the current file descriptor // This is a leader, store and assign it to the current file descriptor
leader_ids.push_back(id); leader_ids.push_back(id);
if (group_pmus) {
group_id = id; group_id = id;
} }
}
// This is a valid counter, add it to our descriptor's list // This is a valid counter, add it to our descriptor's list
counter_ids.push_back(id); counter_ids.push_back(id);
valid_names.push_back(name); valid_names.push_back(name);
@ -214,9 +221,9 @@ PerfCounters PerfCounters::Create(
// This should never happen but if it does, we give up on the // This should never happen but if it does, we give up on the
// entire batch as recovery would be a mess. // entire batch as recovery would be a mess.
GetErrorLogInstance() << "***WARNING*** Failed to start counters. " GetErrorLogInstance() << "***WARNING*** Failed to start counters. "
"Claring out all counters.\n"; "Clearing out all counters.\n";
// Close all peformance counters // Close all performance counters
for (int id : counter_ids) { for (int id : counter_ids) {
::close(id); ::close(id);
} }
@ -254,7 +261,7 @@ bool PerfCounters::IsCounterSupported(const std::string&) { return false; }
PerfCounters PerfCounters::Create( PerfCounters PerfCounters::Create(
const std::vector<std::string>& counter_names) { const std::vector<std::string>& counter_names) {
if (!counter_names.empty()) { if (!counter_names.empty()) {
GetErrorLogInstance() << "Performance counters not supported.\n"; GetErrorLogInstance() << "Performance counters not supported.";
} }
return NoCounters(); return NoCounters();
} }
@ -280,3 +287,4 @@ PerfCounters& PerfCounters::operator=(PerfCounters&& other) noexcept {
} }
} // namespace internal } // namespace internal
} // namespace benchmark } // namespace benchmark

2
src/perf_counters.h
View File

@ -71,7 +71,7 @@ class BENCHMARK_EXPORT PerfCounterValues {
} }
// This reading is complex and as the goal of this class is to // This reading is complex and as the goal of this class is to
// abstract away the intrincacies of the reading process, this is // abstract away the intricacies of the reading process, this is
// a better place for it // a better place for it
size_t Read(const std::vector<int>& leaders); size_t Read(const std::vector<int>& leaders);