Spawn several threads when we fail to enqueue work in the blocki… (#181)

* Rebase onto master

* Switch to unbounded channels

src/task/spawn_blocking.rs

@@ -2,7 +2,7 @@ use std::sync::atomic::{AtomicU64, Ordering};
 use std::thread;
 use std::time::Duration;
 
-use crossbeam_channel::{bounded, Receiver, Sender};
+use crossbeam_channel::{unbounded, Receiver, Sender};
 use once_cell::sync::Lazy;
 
 use crate::task::{JoinHandle, Task};
@@ -79,7 +79,7 @@ static POOL: Lazy<Pool> = Lazy::new(|| {
     // before being acted on by a core. This helps keep
     // latency snappy in the overall async system by
     // reducing bufferbloat.
-    let (sender, receiver) = bounded(0);
+    let (sender, receiver) = unbounded();
     Pool { sender, receiver }
 });
 
@@ -95,6 +95,9 @@ fn maybe_create_another_blocking_thread() {
         return;
     }
 
+    let n_to_spawn = std::cmp::min(2 + (workers / 10), 10);
+
+    for _ in 0..n_to_spawn {
         // We want to avoid having all threads terminate at
         // exactly the same time, causing thundering herd
         // effects. We want to stagger their destruction over
@@ -117,6 +120,7 @@ fn maybe_create_another_blocking_thread() {
             })
             .expect("cannot start a dynamic thread driving blocking tasks");
     }
+}
 
 // Enqueues work, attempting to send to the threadpool in a
 // nonblocking way and spinning up another worker thread if