diff --git a/src/bin/sccache-dist/main.rs b/src/bin/sccache-dist/main.rs
index ea702b71e..9c116568b 100644
--- a/src/bin/sccache-dist/main.rs
+++ b/src/bin/sccache-dist/main.rs
@@ -241,6 +241,7 @@ fn run(command: Command) -> Result<i32> {
             scheduler_url,
             scheduler_auth,
             toolchain_cache_size,
+            max_per_core_load,
             num_cpus_to_ignore,
         }) => {
             let bind_addr = bind_addr.unwrap_or(public_addr);
@@ -307,6 +308,7 @@ fn run(command: Command) -> Result<i32> {
                 bind_addr,
                 scheduler_url.to_url(),
                 scheduler_auth,
+                max_per_core_load,
                 num_cpus_to_ignore,
                 server,
             )
@@ -605,8 +607,7 @@ impl SchedulerIncoming for Scheduler {
                 })
                 // Sort servers by least load and oldest error
                 .sorted_by(|(_, details_a, load_a), (_, details_b, load_b)| {
-                    let (penalty_a, penalty_b) = match (details_a.last_error, details_b.last_error)
-                    {
+                    let (score_a, score_b) = match (details_a.last_error, details_b.last_error) {
                         // If neither server has a recent error, prefer the one with lowest load
                         (None, None) => (*load_a, *load_b),
                         // Prefer servers with no recent errors over servers with recent errors
@@ -618,7 +619,7 @@ impl SchedulerIncoming for Scheduler {
                             score_server(load_b, now - err_b),
                         ),
                     };
-                    penalty_a.total_cmp(&penalty_b)
+                    score_a.total_cmp(&score_b)
                 })
                 .find_or_first(|_| true)
                 .map(|(server_id, _, _)| *server_id)
diff --git a/src/config.rs b/src/config.rs
index 722316425..fc6d9d4d0 100644
--- a/src/config.rs
+++ b/src/config.rs
@@ -1190,6 +1190,14 @@ pub mod server {
         TEN_GIGS
     }
 
+    pub fn default_max_per_core_load() -> f64 {
+        std::env::var("SCCACHE_DIST_MAX_PER_CORE_LOAD")
+            .ok()
+            .and_then(|s| s.parse().ok())
+            // Default to 1 to match the server's default thread pool multiple
+            .unwrap_or(1f64)
+    }
+
     fn default_num_cpus_to_ignore() -> usize {
         std::env::var("SCCACHE_DIST_NUM_CPUS_TO_IGNORE")
             .ok()
@@ -1268,6 +1276,7 @@ pub mod server {
         pub scheduler_url: HTTPUrl,
         pub scheduler_auth: SchedulerAuth,
         pub toolchain_cache_size: u64,
+        pub max_per_core_load: f64,
         pub num_cpus_to_ignore: usize,
     }
 
@@ -1283,6 +1292,7 @@ pub mod server {
                 ),
                 scheduler_auth: SchedulerAuth::Insecure,
                 toolchain_cache_size: default_toolchain_cache_size(),
+                max_per_core_load: default_max_per_core_load(),
                 num_cpus_to_ignore: default_num_cpus_to_ignore(),
             }
         }
diff --git a/src/dist/http.rs b/src/dist/http.rs
index bd5125416..5095cef14 100644
--- a/src/dist/http.rs
+++ b/src/dist/http.rs
@@ -928,6 +928,7 @@ mod server {
         jwt_key: Vec<u8>,
         // Randomly generated nonce to allow the scheduler to detect server restarts
         server_nonce: ServerNonce,
+        max_per_core_load: f64,
         num_cpus_to_ignore: usize,
         handler: S,
     }
@@ -938,6 +939,7 @@ mod server {
             bind_addr: SocketAddr,
             scheduler_url: reqwest::Url,
             scheduler_auth: String,
+            max_per_core_load: f64,
             num_cpus_to_ignore: usize,
             handler: S,
         ) -> Result<Self> {
@@ -958,6 +960,7 @@ mod server {
                 privkey_pem,
                 jwt_key,
                 server_nonce,
+                max_per_core_load,
                 num_cpus_to_ignore,
                 handler,
             })
@@ -974,6 +977,7 @@ mod server {
                 privkey_pem,
                 jwt_key,
                 server_nonce,
+                max_per_core_load,
                 num_cpus_to_ignore,
                 handler,
             } = self;
@@ -1099,7 +1103,7 @@ mod server {
             // This limit is rouille's default for `start_server_with_pool`, which
             // we would use, except that interface doesn't permit any sort of
             // error handling to be done.
-            let server = server.pool_size(num_cpus);
+            let server = server.pool_size((num_cpus as f64 * max_per_core_load).floor() as usize);
             server.run();
 
             panic!("Rouille server terminated")
diff --git a/tests/harness/mod.rs b/tests/harness/mod.rs
index eb689f1a4..fdde68e0a 100644
--- a/tests/harness/mod.rs
+++ b/tests/harness/mod.rs
@@ -419,6 +419,7 @@ impl DistSystem {
             server_addr,
             self.scheduler_url().to_url(),
             token,
+            1f64,
             0,
             handler,
         )