From e480d69cbd1a42e492b46e6936496304ec2d5b2d Mon Sep 17 00:00:00 2001
From: Julien Cuquemelle <j.cuquemelle@criteo.com>
Date: Thu, 30 Nov 2017 16:28:06 +0000
Subject: [PATCH] [SPARK-22683][DOC] document tasksPerExecutorSlot parameter

---
 docs/configuration.md | 14 ++++++++++++++
 1 file changed, 14 insertions(+)

diff --git a/docs/configuration.md b/docs/configuration.md
index c39c7580fd75a..081704f0fe061 100644
--- a/docs/configuration.md
+++ b/docs/configuration.md
@@ -1584,6 +1584,7 @@ Apart from these, the following properties are also available, and may be useful
     <code>spark.dynamicAllocation.minExecutors</code>,
     <code>spark.dynamicAllocation.maxExecutors</code>, and
     <code>spark.dynamicAllocation.initialExecutors</code>
+    <code>spark.dynamicAllocation.tasksPerExecutorSlots</code>
   </td>
 </tr>
 <tr>
@@ -1628,6 +1629,19 @@ Apart from these, the following properties are also available, and may be useful
     Lower bound for the number of executors if dynamic allocation is enabled.
   </td>
 </tr>
+<tr>
+  <td><code>spark.dynamicAllocation.tasksPerSlot</code></td>
+  <td>1</td>
+  <td>
+    Each executor can process a certain number of tasks in parallel (task slots).
+    The number of task slots per executor is: executor.cores / task.cpus.
+    The ExecutorAllocationManager will set a target number of running executors equal to:
+    nbCurrentTask / (taskSlots * tasksPerSlot), with nbCurrentTask being the total number
+    of running and backlogged tasks. With the default value of 1, each available task slot
+    will compute a single task in average, which gives the best latency. With small tasks
+    however, this setting wastes a lot of resources due to executor allocation overhead
+  </td>
+</tr>
 <tr>
   <td><code>spark.dynamicAllocation.schedulerBacklogTimeout</code></td>
   <td>1s</td>