diff --git a/src/benchmark.cc b/src/benchmark.cc
index fc36fedb1..19129e4ca 100644
--- a/src/benchmark.cc
+++ b/src/benchmark.cc
@@ -141,6 +141,9 @@ BM_DEFINE_bool(benchmark_counters_tabular, false);
 // information about libpfm: https://man7.org/linux/man-pages/man3/libpfm.3.html
 BM_DEFINE_string(benchmark_perf_counters, "");
 
+// Setting this to false allows measuring only the main benchmark thread.
+BM_DEFINE_bool(benchmark_perf_counters_inherit, true);
+
 // Extra context to include in the output formatted as comma-separated key-value
 // pairs. Kept internal as it's only used for parsing from env/command line.
 BM_DEFINE_kvpairs(benchmark_context, {});
@@ -424,7 +427,8 @@ void RunBenchmarks(const std::vector<BenchmarkInstance>& benchmarks,
     // This perfcounters object needs to be created before the runners vector
     // below so it outlasts their lifetime.
     PerfCountersMeasurement perfcounters(
-        StrSplit(FLAGS_benchmark_perf_counters, ','));
+        StrSplit(FLAGS_benchmark_perf_counters, ','),
+        FLAGS_benchmark_perf_counters_inherit);
 
     // Vector of benchmarks to run
     std::vector<internal::BenchmarkRunner> runners;
@@ -453,7 +457,8 @@ void RunBenchmarks(const std::vector<BenchmarkInstance>& benchmarks,
 
     // The use of performance counters with threads would be unintuitive for
     // the average user so we need to warn them about this case
-    if ((benchmarks_with_threads > 0) && (perfcounters.num_counters() > 0)) {
+    if ((benchmarks_with_threads > 0) && (perfcounters.num_counters() > 0) && 
+    FLAGS_benchmark_perf_counters_inherit) {
       GetErrorLogInstance()
           << "***WARNING*** There are " << benchmarks_with_threads
           << " benchmarks with threads and " << perfcounters.num_counters()
@@ -770,6 +775,8 @@ void ParseCommandLineFlags(int* argc, char** argv) {
                       &FLAGS_benchmark_report_aggregates_only) ||
         ParseBoolFlag(argv[i], "benchmark_display_aggregates_only",
                       &FLAGS_benchmark_display_aggregates_only) ||
+        ParseBoolFlag(argv[i], "benchmark_perf_counters_inherit", 
+                      &FLAGS_benchmark_perf_counters_inherit) ||
         ParseStringFlag(argv[i], "benchmark_format", &FLAGS_benchmark_format) ||
         ParseStringFlag(argv[i], "benchmark_out", &FLAGS_benchmark_out) ||
         ParseStringFlag(argv[i], "benchmark_out_format",
@@ -891,6 +898,7 @@ void PrintDefaultHelp() {
           "          [--benchmark_counters_tabular={true|false}]\n"
 #if defined HAVE_LIBPFM
           "          [--benchmark_perf_counters=<counter>,...]\n"
+          "          [--benchmark_perf_counters_inherit={true|false}]\n" // <--- Add this
 #endif
           "          [--benchmark_context=<key>=<value>,...]\n"
           "          [--benchmark_time_unit={ns|us|ms|s}]\n"
diff --git a/src/perf_counters.cc b/src/perf_counters.cc
index e6f220921..fe21b67ed 100644
--- a/src/perf_counters.cc
+++ b/src/perf_counters.cc
@@ -135,7 +135,7 @@ static std::vector<uint64_t> GetPMUTypesForEvent(const perf_event_attr& attr) {
 }
 
 PerfCounters PerfCounters::Create(
-    const std::vector<std::string>& counter_names) {
+    const std::vector<std::string>& counter_names, bool inherit) {
   if (!counter_names.empty()) {
     Initialize();
   }
@@ -202,8 +202,9 @@ PerfCounters PerfCounters::Create(
     // Note: the man page for perf_event_create suggests inherit = true and
     // read_format = PERF_FORMAT_GROUP don't work together, but that's not the
     // case.
+    
     attr.disabled = is_first;
-    attr.inherit = true;
+    attr.inherit = inherit;
     attr.pinned = is_first;
     attr.exclude_kernel = true;
     attr.exclude_user = false;
@@ -311,10 +312,11 @@ bool PerfCounters::Initialize() { return false; }
 bool PerfCounters::IsCounterSupported(const std::string&) { return false; }
 
 PerfCounters PerfCounters::Create(
-    const std::vector<std::string>& counter_names) {
+    const std::vector<std::string>& counter_names, bool inherit) {
   if (!counter_names.empty()) {
     GetErrorLogInstance() << "Performance counters not supported.\n";
   }
+  (void)inherit; // This just tells the compiler to ignore the variable
   return NoCounters();
 }
 
@@ -322,9 +324,9 @@ void PerfCounters::CloseCounters() const {}
 #endif  // defined HAVE_LIBPFM
 
 PerfCountersMeasurement::PerfCountersMeasurement(
-    const std::vector<std::string>& counter_names)
+    const std::vector<std::string>& counter_names, bool inherit)
     : start_values_(counter_names.size()), end_values_(counter_names.size()) {
-  counters_ = PerfCounters::Create(counter_names);
+  counters_ = PerfCounters::Create(counter_names, inherit);
 }
 
 PerfCounters& PerfCounters::operator=(PerfCounters&& other) noexcept {
diff --git a/src/perf_counters.h b/src/perf_counters.h
index 4e4534431..2d3f09827 100644
--- a/src/perf_counters.h
+++ b/src/perf_counters.h
@@ -113,7 +113,7 @@ class BENCHMARK_EXPORT PerfCounters final {
   // implementation and OS specific.
   // In case of failure, this method will in the worst case return an
   // empty object whose state will still be valid.
-  static PerfCounters Create(const std::vector<std::string>& counter_names);
+  static PerfCounters Create(const std::vector<std::string>& counter_names, bool inherit = true);
 
   // Take a snapshot of the current value of the counters into the provided
   // valid PerfCounterValues storage. The values are populated such that:
@@ -148,7 +148,7 @@ class BENCHMARK_EXPORT PerfCounters final {
 // Typical usage of the above primitives.
 class BENCHMARK_EXPORT PerfCountersMeasurement final {
  public:
-  PerfCountersMeasurement(const std::vector<std::string>& counter_names);
+  PerfCountersMeasurement(const std::vector<std::string>& counter_names, bool inherit);
 
   size_t num_counters() const { return counters_.num_counters(); }