diff --git a/crates/oxc_allocator/src/lib.rs b/crates/oxc_allocator/src/lib.rs
index da6f1c14b..33e738d02 100644
--- a/crates/oxc_allocator/src/lib.rs
+++ b/crates/oxc_allocator/src/lib.rs
@@ -1,4 +1,7 @@
-use std::{convert::From, ops::Deref};
+use std::{
+    convert::From,
+    ops::{Deref, DerefMut},
+};
 
 mod arena;
 
@@ -24,6 +27,12 @@ impl Deref for Allocator {
     }
 }
 
+impl DerefMut for Allocator {
+    fn deref_mut(&mut self) -> &mut Self::Target {
+        &mut self.bump
+    }
+}
+
 #[cfg(test)]
 mod test {
     use std::ops::Deref;
diff --git a/tasks/benchmark/benches/lexer.rs b/tasks/benchmark/benches/lexer.rs
index 6d2284767..7a95125f6 100644
--- a/tasks/benchmark/benches/lexer.rs
+++ b/tasks/benchmark/benches/lexer.rs
@@ -27,14 +27,14 @@ fn bench_lexer(criterion: &mut Criterion) {
             BenchmarkId::from_parameter(&file.file_name),
             &file.source_text,
             |b, source_text| {
-                b.iter_with_large_drop(|| {
-                    // Include the allocator drop time to make time measurement consistent.
-                    // Otherwise the allocator will allocate huge memory chunks (by power of two) from the
-                    // system allocator, which makes time measurement unequal during long runs.
-                    let allocator = Allocator::default();
+                // Do not include initializing allocator in benchmark.
+                // User code would likely reuse the same allocator over and over to parse multiple files,
+                // so we do the same here.
+                let mut allocator = Allocator::default();
+                b.iter(|| {
                     let mut lexer = Lexer::new_for_benchmarks(&allocator, source_text, source_type);
                     while lexer.next_token().kind != Kind::Eof {}
-                    allocator
+                    allocator.reset();
                 });
             },
         );
diff --git a/tasks/benchmark/benches/parser.rs b/tasks/benchmark/benches/parser.rs
index db0aaa1f1..05418ee2f 100644
--- a/tasks/benchmark/benches/parser.rs
+++ b/tasks/benchmark/benches/parser.rs
@@ -12,13 +12,13 @@ fn bench_parser(criterion: &mut Criterion) {
             BenchmarkId::from_parameter(&file.file_name),
             &file.source_text,
             |b, source_text| {
-                b.iter_with_large_drop(|| {
-                    // Include the allocator drop time to make time measurement consistent.
-                    // Otherwise the allocator will allocate huge memory chunks (by power of two) from the
-                    // system allocator, which makes time measurement unequal during long runs.
-                    let allocator = Allocator::default();
-                    _ = Parser::new(&allocator, source_text, source_type).parse();
-                    allocator
+                // Do not include initializing allocator in benchmark.
+                // User code would likely reuse the same allocator over and over to parse multiple files,
+                // so we do the same here.
+                let mut allocator = Allocator::default();
+                b.iter(|| {
+                    Parser::new(&allocator, source_text, source_type).parse();
+                    allocator.reset();
                 });
             },
         );