[jira] [Commented] (SPARK-27651) Avoid the network when block manager fetches shuffle blocks from the same host
[ https://issues.apache.org/jira/browse/SPARK-27651?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17052348#comment-17052348 ] Attila Zsolt Piros commented on SPARK-27651: well in case of dynamic allocation and a recalculation the executors could be already gone. > Avoid the network when block manager fetches shuffle blocks from the same host > -- > > Key: SPARK-27651 > URL: https://issues.apache.org/jira/browse/SPARK-27651 > Project: Spark > Issue Type: Improvement > Components: Block Manager >Affects Versions: 3.0.0 >Reporter: Attila Zsolt Piros >Assignee: Attila Zsolt Piros >Priority: Major > Fix For: 3.0.0 > > > When a shuffle block (content) is fetched the network is always used even > when it is fetched from the external shuffle service running on the same > host. This can be avoided by getting the local directories of the same host > executors from the external shuffle service and accessing those blocks from > the disk directly. -- This message was sent by Atlassian Jira (v8.3.4#803005) - To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-27651) Avoid the network when block manager fetches shuffle blocks from the same host
[ https://issues.apache.org/jira/browse/SPARK-27651?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17052251#comment-17052251 ] Thomas Graves commented on SPARK-27651: --- thanks, that makes sense. I can look in more details at the code, but I assume the executors could ask the other executors for the directory list rather than going to the external shuffle service if we wanted to support it. > Avoid the network when block manager fetches shuffle blocks from the same host > -- > > Key: SPARK-27651 > URL: https://issues.apache.org/jira/browse/SPARK-27651 > Project: Spark > Issue Type: Improvement > Components: Block Manager >Affects Versions: 3.0.0 >Reporter: Attila Zsolt Piros >Assignee: Attila Zsolt Piros >Priority: Major > Fix For: 3.0.0 > > > When a shuffle block (content) is fetched the network is always used even > when it is fetched from the external shuffle service running on the same > host. This can be avoided by getting the local directories of the same host > executors from the external shuffle service and accessing those blocks from > the disk directly. -- This message was sent by Atlassian Jira (v8.3.4#803005) - To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-27651) Avoid the network when block manager fetches shuffle blocks from the same host
[ https://issues.apache.org/jira/browse/SPARK-27651?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17051995#comment-17051995 ] Attila Zsolt Piros commented on SPARK-27651: Yes, the final implementation works only when the external shuffle service is used as the local directories of the other host local executors are asked from the external shuffle service. The initial implementation when the PR was opened was using the driver to get the host local directories. The technical reasons of asking the external shuffle service was: * decreasing network pressure on the driver (main reason). * getting rid of an unbounded (or bounded but in that case complex fall back logic at the fetcher) map which maps the executors to local dirs. In addition does that redundantly as this information is already available at the external shuffle service just stored in distributed way I mean at a running ext shuffle service process only for those executor data are stored which are on the same host. > Avoid the network when block manager fetches shuffle blocks from the same host > -- > > Key: SPARK-27651 > URL: https://issues.apache.org/jira/browse/SPARK-27651 > Project: Spark > Issue Type: Improvement > Components: Block Manager >Affects Versions: 3.0.0 >Reporter: Attila Zsolt Piros >Assignee: Attila Zsolt Piros >Priority: Major > Fix For: 3.0.0 > > > When a shuffle block (content) is fetched the network is always used even > when it is fetched from an executor (or the external shuffle service) running > on the same host. -- This message was sent by Atlassian Jira (v8.3.4#803005) - To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-27651) Avoid the network when block manager fetches shuffle blocks from the same host
[ https://issues.apache.org/jira/browse/SPARK-27651?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17051630#comment-17051630 ] Thomas Graves commented on SPARK-27651: --- It looks like this only works when using the external shuffle service, is that correct? The way I read the description implies it works from both "from an executor (or the external shuffle service)" so perhaps we should clarify. In both this Jira and the config descriptions. Also was there any technical reasons we didn't support it for executor to executor shuffle? > Avoid the network when block manager fetches shuffle blocks from the same host > -- > > Key: SPARK-27651 > URL: https://issues.apache.org/jira/browse/SPARK-27651 > Project: Spark > Issue Type: Improvement > Components: Block Manager >Affects Versions: 3.0.0 >Reporter: Attila Zsolt Piros >Assignee: Attila Zsolt Piros >Priority: Major > Fix For: 3.0.0 > > > When a shuffle block (content) is fetched the network is always used even > when it is fetched from an executor (or the external shuffle service) running > on the same host. -- This message was sent by Atlassian Jira (v8.3.4#803005) - To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org
[jira] [Commented] (SPARK-27651) Avoid the network when block manager fetches shuffle blocks from the same host
[ https://issues.apache.org/jira/browse/SPARK-27651?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16834832#comment-16834832 ] Attila Zsolt Piros commented on SPARK-27651: I am already working on this. > Avoid the network when block manager fetches shuffle blocks from the same host > -- > > Key: SPARK-27651 > URL: https://issues.apache.org/jira/browse/SPARK-27651 > Project: Spark > Issue Type: Improvement > Components: Block Manager >Affects Versions: 3.0.0 >Reporter: Attila Zsolt Piros >Priority: Major > > When a shuffle block (content) is fetched the network is always used even > when it is fetched from an executor (or the external shuffle service) running > on the same host. -- This message was sent by Atlassian JIRA (v7.6.3#76005) - To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org