try to decrease the variance in the memory usage for benchmark

src/uu/du/benches/du_bench.rs

@@ -77,9 +77,17 @@ fn du_all_wide_tree(bencher: Bencher, (total_files, total_dirs): (usize, usize))
 /// Benchmark du on deep directory structures
 #[divan::bench(args = [(100, 3)])]
 fn du_deep_tree(bencher: Bencher, (depth, files_per_level): (usize, usize)) {
-    let temp_dir = TempDir::new().unwrap();
-    fs_tree::create_deep_tree(temp_dir.path(), depth, files_per_level);
-    bench_du_with_args(bencher, &temp_dir, &[]);
+    bencher
+        .with_inputs(|| {
+            let temp_dir = TempDir::new().unwrap();
+            fs_tree::create_deep_tree(temp_dir.path(), depth, files_per_level);
+            temp_dir
+        })
+        .bench_values(|temp_dir| {
+            let temp_path_str = temp_dir.path().to_str().unwrap();
+            let args = vec![temp_path_str];
+            black_box(run_util_function(uumain, &args));
+        });
 }
 
 /// Benchmark du -s (summarize) on balanced tree

src/uu/numfmt/benches/numfmt_bench.rs

@@ -63,8 +63,12 @@ fn numfmt_from_si(bencher: Bencher, count: usize) {
 /// Benchmark large numbers with SI formatting
 #[divan::bench(args = [10_000])]
 fn numfmt_large_numbers_si(bencher: Bencher, count: usize) {
-    // Generate larger numbers (millions to billions range)
-    let numbers: Vec<String> = (1..=count).map(|n| (n * 1_000_000).to_string()).collect();
+    // Generate numbers that all produce uniform SI output lengths (all in 1-9M range)
+    // This avoids variance from variable output string lengths
+    let numbers: Vec<String> = (1..=count)
+        .map(|n| ((n % 9) + 1) * 1_000_000)
+        .map(|n| n.to_string())
+        .collect();
     let mut args = vec!["--to=si"];
     let number_refs: Vec<&str> = numbers.iter().map(|s| s.as_str()).collect();
     args.extend(number_refs);