-
Notifications
You must be signed in to change notification settings - Fork 28.3k
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Broadcast task sql test [DO NOT MERGE] #1654
Conversation
QA tests have started for PR 1654. This patch merges cleanly. |
…very task). Currently (as of Spark 1.0.1), Spark sends RDD object (which contains closures) using Akka along with the task itself to the executors. This is inefficient because all tasks in the same stage use the same RDD object, but we have to send RDD object multiple times to the executors. This is especially bad when a closure references some variable that is very large. The current design led to users having to explicitly broadcast large variables. The patch uses broadcast to send RDD objects and the closures to executors, and use Akka to only send a reference to the broadcast RDD/closure along with the partition specific information for the task. For those of you who know more about the internals, Spark already relies on broadcast to send the Hadoop JobConf every time it uses the Hadoop input, because the JobConf is large. The user-facing impact of the change include: 1. Users won't need to decide what to broadcast anymore, unless they would want to use a large object multiple times in different operations 2. Task size will get smaller, resulting in faster scheduling and higher task dispatch throughput. In addition, the change will simplify some internals of Spark, eliminating the need to maintain task caches and the complex logic to broadcast JobConf (which also led to a deadlock recently). A simple way to test this: ```scala val a = new Array[Byte](1000*1000); scala.util.Random.nextBytes(a); sc.parallelize(1 to 1000, 1000).map { x => a; x }.groupBy { x => a; x }.count ``` Numbers on 3 r3.8xlarge instances on EC2 ``` master branch: 5.648436068 s, 4.715361895 s, 5.360161877 s with this change: 3.416348793 s, 1.477846558 s, 1.553432156 s ``` Author: Reynold Xin <rxin@apache.org> Closes apache#1452 from rxin/broadcast-task and squashes the following commits: 762e0be [Reynold Xin] Warn large broadcasts. ade6eac [Reynold Xin] Log broadcast size. c3b6f11 [Reynold Xin] Added a unit test for clean up. 754085f [Reynold Xin] Explain why broadcasting serialized copy of the task. 04b17f0 [Reynold Xin] [SPARK-2521] Broadcast RDD object once per TaskSet (instead of sending it for every task). (cherry picked from commit 7b8cd17) Signed-off-by: Reynold Xin <rxin@apache.org>
QA results for PR 1654: |
QA tests have started for PR 1654. This patch merges cleanly. |
QA results for PR 1654: |
QA tests have started for PR 1654. This patch merges cleanly. |
QA results for PR 1654: |
This is the same as #1498 but adds a dummy change to sql so I can trigger a SQL test on Jenkins.
This is not meant to be merged.