diff --git a/Cargo.toml b/Cargo.toml
index b5d5dfc621..0793ca3b13 100644
--- a/Cargo.toml
+++ b/Cargo.toml
@@ -40,9 +40,9 @@ mimalloc-sys = { version = "0.1.6", optional = true }
 mmtk-macros = { version = "0.20.0", path = "macros/" }
 num_cpus = "1.8"
 num-traits = "0.2"
-pfm = { version = "0.1.0-beta.3", optional = true }
-probe = "0.5"
+pfm = { version = "0.1.1", optional = true }
 portable-atomic = "1.4.3"
+probe = "0.5"
 regex = "1.7.0"
 spin = "0.9.5"
 static_assertions = "1.1.0"
diff --git a/src/scheduler/stat.rs b/src/scheduler/stat.rs
index f40857182c..027d3c490c 100644
--- a/src/scheduler/stat.rs
+++ b/src/scheduler/stat.rs
@@ -240,7 +240,12 @@ impl<VM: VMBinding> WorkerLocalStat<VM> {
         let mut counters: Vec<Box<dyn WorkCounter>> = vec![Box::new(WorkDuration::new())];
         #[cfg(feature = "perf_counter")]
         for e in &mmtk.options.work_perf_events.events {
-            counters.push(Box::new(WorkPerfEvent::new(&e.0, e.1, e.2)));
+            counters.push(Box::new(WorkPerfEvent::new(
+                &e.0,
+                e.1,
+                e.2,
+                *mmtk.options.perf_exclude_kernel,
+            )));
         }
         counters
     }
diff --git a/src/scheduler/work_counter.rs b/src/scheduler/work_counter.rs
index 1e1d8b5770..505105b042 100644
--- a/src/scheduler/work_counter.rs
+++ b/src/scheduler/work_counter.rs
@@ -162,9 +162,10 @@ mod perf_event {
         /// 0, -1 measures the calling thread on all CPUs
         /// -1, 0 measures all threads on CPU 0
         /// -1, -1 is invalid
-        pub fn new(name: &str, pid: pid_t, cpu: c_int) -> WorkPerfEvent {
+        pub fn new(name: &str, pid: pid_t, cpu: c_int, exclude_kernel: bool) -> WorkPerfEvent {
             let mut pe = PerfEvent::new(name, false)
                 .unwrap_or_else(|_| panic!("Failed to create perf event {}", name));
+            pe.set_exclude_kernel(exclude_kernel as u64);
             pe.open(pid, cpu)
                 .unwrap_or_else(|_| panic!("Failed to open perf event {}", name));
             WorkPerfEvent {
diff --git a/src/util/options.rs b/src/util/options.rs
index bbe408c00f..3b0bff5249 100644
--- a/src/util/options.rs
+++ b/src/util/options.rs
@@ -733,6 +733,9 @@ options! {
     // Measuring perf events for GC and mutators
     // TODO: Ideally this option should only be included when the features 'perf_counter' are enabled. The current macro does not allow us to do this.
     phase_perf_events:      PerfEventOptions     [env_var: true, command_line: true] [|_| cfg!(feature = "perf_counter")] = PerfEventOptions {events: vec![]},
+    // Should we exclude perf events occurring in kernel space. By default we include the kernel.
+    // Only set this option if you know the implications of excluding the kernel!
+    perf_exclude_kernel:    bool                  [env_var: true, command_line: true] [|_| cfg!(feature = "perf_counter")] = false,
     // Set how to bind affinity to the GC Workers. Default thread affinity delegates to the OS
     // scheduler. If a list of cores are specified, cores are allocated to threads in a round-robin
     // fashion. The core ids should match the ones reported by /proc/cpuinfo. Core ids are
@@ -749,7 +752,7 @@ options! {
     thread_affinity:        AffinityKind         [env_var: true, command_line: true] [|v: &AffinityKind| v.validate()] = AffinityKind::OsDefault,
     // Set the GC trigger. This defines the heap size and how MMTk triggers a GC.
     // Default to a fixed heap size of 0.5x physical memory.
-    gc_trigger     :        GCTriggerSelector    [env_var: true, command_line: true] [|v: &GCTriggerSelector| v.validate()] = GCTriggerSelector::FixedHeapSize((crate::util::memory::get_system_total_memory() as f64 * 0.5f64) as usize),
+    gc_trigger:             GCTriggerSelector    [env_var: true, command_line: true] [|v: &GCTriggerSelector| v.validate()] = GCTriggerSelector::FixedHeapSize((crate::util::memory::get_system_total_memory() as f64 * 0.5f64) as usize),
     // Enable transparent hugepage support via madvise (only Linux is supported)
     transparent_hugepages: bool                  [env_var: true, command_line: true]  [|v: &bool| !v || cfg!(target_os = "linux")] = false
 }
diff --git a/src/util/statistics/counter/perf_event.rs b/src/util/statistics/counter/perf_event.rs
index a203f41257..3b0d616d26 100644
--- a/src/util/statistics/counter/perf_event.rs
+++ b/src/util/statistics/counter/perf_event.rs
@@ -9,9 +9,10 @@ pub struct PerfEventDiffable {
 }
 
 impl PerfEventDiffable {
-    pub fn new(name: &str) -> Self {
+    pub fn new(name: &str, exclude_kernel: bool) -> Self {
         let mut pe = PerfEvent::new(name, true)
             .unwrap_or_else(|_| panic!("Failed to create perf event {}", name));
+        pe.set_exclude_kernel(exclude_kernel as u64);
         // measures the calling thread (and all child threads) on all CPUs
         pe.open(0, -1)
             .unwrap_or_else(|_| panic!("Failed to open perf event {}", name));
diff --git a/src/util/statistics/stats.rs b/src/util/statistics/stats.rs
index 61a2a05172..6e754b1384 100644
--- a/src/util/statistics/stats.rs
+++ b/src/util/statistics/stats.rs
@@ -91,7 +91,7 @@ impl Stats {
                 shared.clone(),
                 true,
                 false,
-                PerfEventDiffable::new(&e.0),
+                PerfEventDiffable::new(&e.0, *options.perf_exclude_kernel),
             ))));
         }
         Stats {