[ https://issues.apache.org/jira/browse/SPARK-21176?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]
Ingo Schuster updated SPARK-21176: ---------------------------------- Description: In reverse proxy mode, Sparks exhausts the Jetty thread pool if the master node has too many cpus or the cluster has too many executers: For each connector, Jetty creates Selector threads: minimum 4, maximum half the number of available CPUs: {{ this(Math.max(1, Runtime.getRuntime().availableProcessors() / 2));}} (see https://github.com/eclipse/jetty.project/blob/0c8273f2ca1f9bf2064cd9c4c939d2546443f759/jetty-client/src/main/java/org/eclipse/jetty/client/http/HttpClientTransportOverHTTP.java) In reverse proxy mode, a connector is set up for each executor and one for the master UI. I have a system with 88 CPUs on the master node and 7 executors. Jetty tries to instantiate 8*44 = 352 selector threads, but since the QueuedThreadPool is initialized with 200 threads by default, the UI gets stuck. I have patched JettyUtils.scala to extend the thread pool ( {{val pool = new QueuedThreadPool*(400)* }}). With this hack, the UI works. Obviously, the Jetty defaults are meant for a real web server. If that has 88 CPUs, you do certainly expect a lot of traffic. For the Spark admin UI however, there will rarely be concurrent accesses for the same application or the same executor. I therefore propose to dramatically reduce the number of selector threads that get instantiated - at least by default. I will propose a fix in a pull request. was: In reverse proxy mode, Sparks exhausts the Jetty thread pool if the master node has too many cpus or the cluster has too many executers: For each connector, Jetty creates Selector threads: minimum 4, maximum half the number of available CPUs: {{selectors>0?selectors:Math.max(1,Math.min(4,Runtime.getRuntime().availableProcessors()/2)));}} (see https://github.com/eclipse/jetty.project/blob/jetty-9.3.x/jetty-server/src/main/java/org/eclipse/jetty/server/ServerConnector.java) In reverse proxy mode, a connector is set up for each executor and one for the master UI. I have a system with 88 CPUs on the master node and 7 executors. Jetty tries to instantiate 8*44 = 352 selector threads, but since the QueuedThreadPool is initialized with 200 threads by default, the UI gets stuck. I have patched JettyUtils.scala to extend the thread pool ( {{val pool = new QueuedThreadPool*(400)* }}). With this hack, the UI works. Obviously, the Jetty defaults are meant for a real web server. If that has 88 CPUs, you do certainly expect a lot of traffic. For the Spark admin UI however, there will rarely be concurrent accesses for the same application or the same executor. I therefore propose to dramatically reduce the number of selector threads that get instantiated - at least by default. I will propose a fix in a pull request. > Master UI hangs with spark.ui.reverseProxy=true if the master node has many > CPUs > -------------------------------------------------------------------------------- > > Key: SPARK-21176 > URL: https://issues.apache.org/jira/browse/SPARK-21176 > Project: Spark > Issue Type: Bug > Components: Web UI > Affects Versions: 2.1.0, 2.1.1, 2.2.0, 2.2.1 > Environment: ppc64le GNU/Linux, POWER8, only master node is reachable > externally other nodes are in an internal network > Reporter: Ingo Schuster > Labels: network, web-ui > > In reverse proxy mode, Sparks exhausts the Jetty thread pool if the master > node has too many cpus or the cluster has too many executers: > For each connector, Jetty creates Selector threads: minimum 4, maximum half > the number of available CPUs: > {{ this(Math.max(1, Runtime.getRuntime().availableProcessors() / 2));}} > (see > https://github.com/eclipse/jetty.project/blob/0c8273f2ca1f9bf2064cd9c4c939d2546443f759/jetty-client/src/main/java/org/eclipse/jetty/client/http/HttpClientTransportOverHTTP.java) > In reverse proxy mode, a connector is set up for each executor and one for > the master UI. > I have a system with 88 CPUs on the master node and 7 executors. Jetty tries > to instantiate 8*44 = 352 selector threads, but since the QueuedThreadPool is > initialized with 200 threads by default, the UI gets stuck. > I have patched JettyUtils.scala to extend the thread pool ( {{val pool = new > QueuedThreadPool*(400)* }}). With this hack, the UI works. > Obviously, the Jetty defaults are meant for a real web server. If that has 88 > CPUs, you do certainly expect a lot of traffic. > For the Spark admin UI however, there will rarely be concurrent accesses for > the same application or the same executor. > I therefore propose to dramatically reduce the number of selector threads > that get instantiated - at least by default. > I will propose a fix in a pull request. -- This message was sent by Atlassian JIRA (v6.4.14#64029) --------------------------------------------------------------------- To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org