diff --git a/README.md b/README.md
index 758fb29..e6fdc7a 100644
--- a/README.md
+++ b/README.md
@@ -480,8 +480,9 @@ Available methods are different depending on `worker_type`. ServerEngine support
   - **server_detach_wait** sets wait time before starting live restart (default: 10.0) [dynamic reloadable]
 - Multithread server and multiprocess server: available only when `worker_type` is thread or process
   - **workers** sets number of workers (default: 1) [dynamic reloadable]
-  - **start_worker_delay** sets wait time before starting a new worker (default: 0) [dynamic reloadable]
+  - **start_worker_delay** sets the delay between each worker-start when starting/restarting multiple workers at once (default: 0) [dynamic reloadable]
   - **start_worker_delay_rand** randomizes start_worker_delay at this ratio (default: 0.2) [dynamic reloadable]
+  - **restart_worker_interval** sets wait time before restarting a stopped worker (default: 0) [dynamic reloadable]
 - Multiprocess server: available only when `worker_type` is "process"
   - **worker_process_name** changes process name ($0) of workers [dynamic reloadable]
   - **worker_heartbeat_interval** sets interval of heartbeats in seconds (default: 1.0) [dynamic reloadable]
diff --git a/lib/serverengine/multi_process_server.rb b/lib/serverengine/multi_process_server.rb
index 19cd72a..41761b4 100644
--- a/lib/serverengine/multi_process_server.rb
+++ b/lib/serverengine/multi_process_server.rb
@@ -105,9 +105,11 @@ def initialize(worker, wid, pmon, reload_signal = Signals::RELOAD, unrecoverable
         @unrecoverable_exit_codes = unrecoverable_exit_codes
         @unrecoverable_exit = false
         @exitstatus = nil
+        @restart_at = nil
       end
 
       attr_reader :exitstatus
+      attr_accessor :restart_at
 
       def send_stop(stop_graceful)
         @stop = true
diff --git a/lib/serverengine/multi_thread_server.rb b/lib/serverengine/multi_thread_server.rb
index 0b3e2d1..6615937 100644
--- a/lib/serverengine/multi_thread_server.rb
+++ b/lib/serverengine/multi_thread_server.rb
@@ -39,8 +39,11 @@ class WorkerMonitor
       def initialize(worker, thread)
         @worker = worker
         @thread = thread
+        @restart_at = nil
       end
 
+      attr_accessor :restart_at
+
       def send_stop(stop_graceful)
         Thread.new do
           begin
diff --git a/lib/serverengine/multi_worker_server.rb b/lib/serverengine/multi_worker_server.rb
index 60b650f..1fef343 100644
--- a/lib/serverengine/multi_worker_server.rb
+++ b/lib/serverengine/multi_worker_server.rb
@@ -55,8 +55,8 @@ def reload
 
     def run
       while true
-        num_alive = keepalive_workers
-        break if num_alive == 0
+        num_alive_or_restarting = keepalive_workers
+        break if num_alive_or_restarting == 0
         wait_tick
       end
     end
@@ -85,6 +85,7 @@ def reload_config
 
       @start_worker_delay = @config[:start_worker_delay] || 0
       @start_worker_delay_rand = @config[:start_worker_delay_rand] || 0.2
+      @restart_worker_interval = @config[:restart_worker_interval] || 0
 
       scale_workers(@config[:workers] || 1)
 
@@ -96,12 +97,12 @@ def wait_tick
     end
 
     def keepalive_workers
-      num_alive = 0
+      num_alive_or_restarting = 0
 
       @monitors.each_with_index do |m,wid|
         if m && m.alive?
           # alive
-          num_alive += 1
+          num_alive_or_restarting += 1
 
         elsif m && m.respond_to?(:recoverable?) && !m.recoverable?
           # exited, with unrecoverable exit code
@@ -116,8 +117,12 @@ def keepalive_workers
         elsif wid < @num_workers
           # scale up or reboot
           unless @stop
-            @monitors[wid] = delayed_start_worker(wid)
-            num_alive += 1
+            if m
+              restart_worker(wid)
+            else
+              start_new_worker(wid)
+            end
+            num_alive_or_restarting += 1
           end
 
         elsif m
@@ -126,7 +131,27 @@ def keepalive_workers
         end
       end
 
-      return num_alive
+      return num_alive_or_restarting
+    end
+
+    def start_new_worker(wid)
+      delayed_start_worker(wid)
+    end
+
+    def restart_worker(wid)
+      m = @monitors[wid]
+
+      is_already_restarting = !m.restart_at.nil?
+      if is_already_restarting
+        delayed_start_worker(wid) if m.restart_at <= Time.now()
+        return
+      end
+
+      if @restart_worker_interval > 0
+        m.restart_at = Time.now() + @restart_worker_interval
+      else
+        delayed_start_worker(wid)
+      end
     end
 
     def delayed_start_worker(wid)
@@ -143,7 +168,7 @@ def delayed_start_worker(wid)
         @last_start_worker_time = now
       end
 
-      start_worker(wid)
+      @monitors[wid] = start_worker(wid)
     end
   end
 
diff --git a/serverengine.gemspec b/serverengine.gemspec
index 91b7f0b..955d9c1 100644
--- a/serverengine.gemspec
+++ b/serverengine.gemspec
@@ -27,6 +27,8 @@ Gem::Specification.new do |gem|
   gem.add_development_dependency 'rake-compiler-dock', ['~> 0.5.0']
   gem.add_development_dependency 'rake-compiler', ['~> 0.9.4']
 
+  gem.add_development_dependency "timecop", ["~> 0.9.5"]
+
   # build gem for a certain platform. see also Rakefile
   fake_platform = ENV['GEM_BUILD_FAKE_PLATFORM'].to_s
   gem.platform = fake_platform unless fake_platform.empty?
diff --git a/spec/multi_spawn_server_spec.rb b/spec/multi_spawn_server_spec.rb
index 194e234..13441d1 100644
--- a/spec/multi_spawn_server_spec.rb
+++ b/spec/multi_spawn_server_spec.rb
@@ -1,25 +1,207 @@
 require 'timeout'
+require 'timecop'
 
 describe ServerEngine::MultiSpawnServer do
   include_context 'test server and worker'
 
-  context 'with command_sender=pipe' do
-    it 'starts worker processes' do
-      config = {workers: 2, command_sender: 'pipe', log_stdout: false, log_stderr: false}
+  describe 'starts worker processes' do
+    context 'with command_sender=pipe' do
+      it do
+        config = {workers: 2, command_sender: 'pipe', log_stdout: false, log_stderr: false}
 
-      s = ServerEngine::MultiSpawnServer.new(TestWorker) { config.dup }
-      t = Thread.new { s.main }
+        s = ServerEngine::MultiSpawnServer.new(TestWorker) { config.dup }
+        t = Thread.new { s.main }
 
-      begin
-        wait_for_fork
+        begin
+          wait_for_fork
 
-        Timeout.timeout(5) do
-          sleep(0.5) until test_state(:worker_run) == 2
+          Timeout.timeout(5) do
+            sleep(0.5) until test_state(:worker_run) == 2
+          end
+          test_state(:worker_run).should == 2
+        ensure
+          s.stop(true)
+          t.join
+        end
+      end
+    end
+  end
+
+  describe 'keepalive_workers' do
+    let(:config) {
+      {
+        workers: workers,
+        command_sender: 'pipe',
+        log_stdout: false,
+        log_stderr: false,
+        start_worker_delay: start_worker_delay,
+        start_worker_delay_rand: 0,
+        restart_worker_interval: restart_worker_interval,
+      }
+    }
+    let(:workers) { 3 }
+    let(:server) { ServerEngine::MultiSpawnServer.new(TestWorker) { config.dup } }
+    let(:monitors) { server.instance_variable_get(:@monitors) }
+
+    context 'default' do
+      let(:start_worker_delay) { 0 }
+      let(:restart_worker_interval) { 0 }
+
+      it do
+        t = Thread.new { server.main }
+
+        begin
+          wait_for_fork
+
+          Timeout.timeout(5) do
+            sleep(0.5) until monitors.count { |m| m && m.alive? } == workers
+          end
+
+          monitors.each do |m|
+            m.send_stop(true)
+          end
+
+          # On Windows, it takes some time to stop all multiple processes, so allow leeway to judge.
+          -> {
+            Timeout.timeout(5) do
+              sleep(0.5) until monitors.count { |m| m.alive? } == workers
+            end
+          }.should_not raise_error, "Not all workers restarted correctly."
+        ensure
+          server.stop(true)
+          t.join
+        end
+      end
+    end
+
+    context 'with only restart_worker_interval' do
+      let(:start_worker_delay) { 0 }
+      let(:restart_worker_interval) { 10 }
+
+      it do
+        t = Thread.new { server.main }
+
+        begin
+          wait_for_fork
+
+          # Wait for initial starting
+          Timeout.timeout(5) do
+            sleep(0.5) until monitors.count { |m| m && m.alive? } == workers
+          end
+
+          monitors.each do |m|
+            m.send_stop(true)
+          end
+
+          # Wait for all workers to stop
+          Timeout.timeout(5) do
+            sleep(0.5) until monitors.count { |m| m && m.alive? } == 0
+          end
+
+          Timecop.freeze
+
+          mergin_time = 3
+
+          Timecop.freeze(Time.now + restart_worker_interval - mergin_time)
+          sleep(1.5)
+          monitors.count { |m| m.alive? }.should == 0
+
+          Timecop.freeze(Time.now + 2 * mergin_time)
+          # On Windows, it takes some time to stop all multiple processes, so allow leeway to judge.
+          -> {
+            Timeout.timeout(5) do
+              sleep(0.5) until monitors.count { |m| m.alive? } == workers
+            end
+          }.should_not raise_error, "Not all workers restarted correctly."
+        ensure
+          server.stop(true)
+          t.join
+        end
+      end
+    end
+
+    context 'with only start_worker_delay' do
+      let(:start_worker_delay) { 3 }
+      let(:restart_worker_interval) { 0 }
+
+      it do
+        t = Thread.new { server.main }
+
+        begin
+          wait_for_fork
+
+          # Initial starts are delayed too, so set longer timeout.
+          # (`start_worker_delay` uses `sleep` inside, so Timecop can't skip this wait.)
+          Timeout.timeout(start_worker_delay * workers) do
+            sleep(0.5) until monitors.count { |m| m && m.alive? } == workers
+          end
+
+          # Skip time to avoid getting a delay for the initial starts.
+          Timecop.travel(Time.now + start_worker_delay)
+
+          monitors.each do |m|
+            m.send_stop(true)
+          end
+
+          sleep(3)
+
+          # The first worker should restart immediately.
+          monitors.count { |m| m.alive? }.should satisfy { |c| 0 < c && c < workers }
+
+          # `start_worker_delay` uses `sleep` inside, so Timecop can't skip this wait.
+          sleep(start_worker_delay * workers)
+          monitors.count { |m| m.alive? }.should == workers
+        ensure
+          server.stop(true)
+          t.join
+        end
+      end
+    end
+
+    context 'with both options' do
+      let(:start_worker_delay) { 3 }
+      let(:restart_worker_interval) { 10 }
+
+      it do
+        t = Thread.new { server.main }
+
+        begin
+          wait_for_fork
+
+          # Initial starts are delayed too, so set longer timeout.
+          # (`start_worker_delay` uses `sleep` inside, so Timecop can't skip this wait.)
+          Timeout.timeout(start_worker_delay * workers) do
+            sleep(0.5) until monitors.count { |m| m && m.alive? } == workers
+          end
+
+          monitors.each do |m|
+            m.send_stop(true)
+          end
+
+          # Wait for all workers to stop
+          Timeout.timeout(5) do
+            sleep(0.5) until monitors.count { |m| m && m.alive? } == 0
+          end
+
+          Timecop.freeze
+
+          mergin_time = 3
+
+          Timecop.freeze(Time.now + restart_worker_interval - mergin_time)
+          sleep(1.5)
+          monitors.count { |m| m.alive? }.should == 0
+
+          Timecop.travel(Time.now + 2 * mergin_time)
+          sleep(1.5)
+          monitors.count { |m| m.alive? }.should satisfy { |c| 0 < c && c < workers }
+
+          # `start_worker_delay` uses `sleep` inside, so Timecop can't skip this wait.
+          sleep(start_worker_delay * workers)
+          monitors.count { |m| m.alive? }.should == workers
+        ensure
+          server.stop(true)
+          t.join
         end
-        test_state(:worker_run).should == 2
-      ensure
-        s.stop(true)
-        t.join
       end
     end
   end
diff --git a/spec/server_worker_context.rb b/spec/server_worker_context.rb
index 7ed2062..21bd71b 100644
--- a/spec/server_worker_context.rb
+++ b/spec/server_worker_context.rb
@@ -1,6 +1,7 @@
 
 require 'thread'
 require 'yaml'
+require 'timecop'
 
 def reset_test_state
   FileUtils.mkdir_p 'tmp'
@@ -165,8 +166,9 @@ def before_fork
 
   def run
     incr_test_state :worker_run
-    5.times do
-      # repeats 5 times because signal handlers
+    # This means this worker will automatically finish after 50 seconds.
+    10.times do
+      # repeats multiple times because signal handlers
       # interrupts wait
       @stop_flag.wait(5.0)
     end
@@ -252,6 +254,7 @@ def stop
 
 shared_context 'test server and worker' do
   before { reset_test_state }
+  after { Timecop.return }
 
   if ServerEngine.windows?
     WAIT_RATIO = 2