--- a/src/bin.rs
+++ b/src/bin.rs
@@ -121,13 +121,15 @@
 
     // get_physical() appears to be more accurate: hyperthreading doesn't
     // help us much
+    let cpus = num_cpus::get_physical();
+    rayon::ThreadPoolBuilder::new().num_threads(cpus).build_global().unwrap();
 
     if trials_per_key < 2u64.pow(32) {
         let raw_rate = measure_rate();
         println!(
             "one core runs at {}, CPU cores available: {}",
             format_rate(raw_rate),
-            num_cpus::get_physical(),
+            cpus,
         );
         let total_rate = raw_rate * (num_cpus::get_physical() as f64) / (trials_per_key as f64);
         let seconds_per_key = 1.0 / total_rate;
@@ -140,9 +142,7 @@
 
     println!("hit Ctrl-C to stop");
 
-    // 1M trials takes about 10s on my laptop, so let it run for 1000s
-    (0..100_000_000)
-        .into_par_iter()
+    rayon::iter::repeat(0)
         .map(|_| search_for_prefix(&prefix, 0, end))
         .try_for_each(print)?;
     Ok(())