Added nakayoshi_fork option.

Reduce memory usage in preloaded cluster-mode apps by GCing before
fork and compacting, where available.

5.0-Upgrade.md

@@ -6,9 +6,9 @@ Puma 5 brings new experimental performance features, a few quality-of-life featu
 
 * **Better throughput and lower latency via new experimental option**
 * **Better memory usage via new experimental option**
+* **Better copy-on-write performance via new experimental option**
 * **Loads of bugfixes**
 * Faster phased restarts and worker timeouts 
-* GC Compact/friendly fork
 * pumactl now has a `thread-backtraces` command to print thread backtraces, bringing thread backtrace printing to all platforms, not just *BSD and Mac. (#2053)
 * Added incrementing `requests_count` to `Puma.stats`. (#2106)
 * Faster phased restart and worker timeout (#2220)

History.md

@@ -1,9 +1,9 @@
 ## 5.0.0
 
 * Features
-  * EXPERIMENTAL: Add `fork_worker` option and `refork` command for reduced memory usage. (#2099)
-  * EXPERIMENTAL: Reduce latency on MRI through inserting a small delay before re-listening on the socket if worker is busy (#2079).
-  * Possibly reduced memory usage by calling `GC.compact`  before fork if available (Ruby 2.7+) (#2093)
+  * EXPERIMENTAL: Add `fork_worker` option and `refork` command for reduced memory usage by forking from a worker process and eliminating the master process. (#2099)
+  * EXPERIMENTAL: Added `wait_for_less_busy_worker` config. This may reduce latency on MRI through inserting a small delay before re-listening on the socket if worker is busy (#2079).
+  * EXPERIMENTAL: Added `nakayoshi_fork` option. Reduce memory usage in preloaded cluster-mode apps by GCing before fork and compacting, where available. (#2093, #2256)
   * Added pumactl `thread-backtraces` command to print thread backtraces (#2054)
   * Added incrementing `requests_count` to `Puma.stats`. (#2106)
   * Increased maximum URI path length from 2048 to 8196 bytes (#2167)

lib/puma/cluster.rb

@@ -298,7 +298,7 @@ def worker(index, master)
                 restart_server.clear
                 server.begin_restart(true)
                 @launcher.config.run_hooks :before_refork, nil, @launcher.events
-                GC.compact if GC.respond_to?(:compact)
+                nakayoshi_gc
               end
             elsif idx == 0 # restart server
               restart_server << true << false
@@ -540,7 +540,7 @@ def run
       @master_read, @worker_write = read, @wakeup
 
       @launcher.config.run_hooks :before_fork, nil, @launcher.events
-      GC.compact if GC.respond_to?(:compact)
+      nakayoshi_gc
 
       spawn_workers
 
@@ -644,5 +644,16 @@ def timeout_workers
         end
       end
     end
+
+    def nakayoshi_gc
+      return unless @options[:nakayoshi_fork]
+      log "! Promoting existing objects to old generation..."
+      4.times { GC.start(full_mark: false) }
+      if GC.respond_to?(:compact)
+        log "! Compacting..."
+        GC.compact
+      end
+      log "! Friendly fork preparation complete."
+    end
   end
 end

lib/puma/dsl.rb

@@ -759,5 +759,18 @@ def set_remote_address(val=:socket)
     def fork_worker(after_requests=1000)
       @options[:fork_worker] = Integer(after_requests)
     end
+
+    # When enabled, Puma will GC 4 times before forking workers.
+    # If available (Ruby 2.7+), we will also call GC.compact.
+    # Not recommended for non-MRI Rubies.
+    #
+    # Based on the work of Koichi Sasada and Aaron Patterson, this option may
+    # decrease memory utilization of preload-enabled cluster-mode Pumas. It will
+    # also increase time to boot and fork. See your logs for details on how much
+    # time this adds to your boot process. For most apps, it will be less than one
+    # second.
+    def nakayoshi_fork(enabled=false)
+      @options[:nakayoshi_fork] = enabled
+    end
   end
 end

test/test_integration_cluster.rb

@@ -168,6 +168,22 @@ def test_refork
     refute_includes pids, get_worker_pids(1, WORKERS - 1)
   end
 
+  def test_nakayoshi
+    cli_server "-w #{WORKERS} test/rackup/hello.ru", config: <<RUBY
+    nakayoshi_fork true
+RUBY
+
+    output = nil
+    Timeout.timeout(10) do
+      until output
+        output = @server.gets[/Friendly fork preparation complete/]
+        sleep(0.01)
+      end
+    end
+
+    assert output, "Friendly fork didn't run"
+  end
+
   private
 
   def worker_timeout(timeout, iterations, config)