[
https://issues.apache.org/jira/browse/HDFS-17935?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=18089898#comment-18089898
]
ASF GitHub Bot commented on HDFS-17935:
---------------------------------------
kokonguyen191 commented on code in PR #8548:
URL: https://github.com/apache/hadoop/pull/8548#discussion_r3435326285
##########
hadoop-hdfs-project/hadoop-hdfs-rbf/src/main/java/org/apache/hadoop/hdfs/server/federation/router/RouterRpcServer.java:
##########
@@ -510,44 +515,75 @@ public RouterRpcServer(Configuration conf, Router router,
public void initAsyncThreadPools(Configuration configuration) {
Set<String> allConfiguredNS =
FederationUtil.getAllConfiguredNS(configuration);
allConfiguredNS.add(CONCURRENT_NS);
- Map<String, Integer> nsAsyncActiveHandlerCount =
parseNsAsyncHandlerCount(configuration);
- initAsyncHandlerThreadPools(configuration, allConfiguredNS,
nsAsyncActiveHandlerCount);
+ Map<String, Integer> nsAsyncActiveHandlerCount =
parseNsAsyncHandlerCount(configuration,
+ DFS_ROUTER_ASYNC_RPC_NS_HANDLER_COUNT_KEY);
+ initAsyncHandlerThreadPools(configuration, allConfiguredNS, false,
+ DFS_ROUTER_ASYNC_RPC_HANDLER_COUNT_KEY,
DFS_ROUTER_ASYNC_RPC_HANDLER_COUNT_DEFAULT,
+ nsAsyncActiveHandlerCount, asyncRouterHandlerExecutors);
+
+ Map<String, Integer> nsAsyncObserverHandlerCount =
parseNsAsyncHandlerCount(configuration,
+ DFS_ROUTER_ASYNC_RPC_NS_OBSERVER_HANDLER_COUNT_KEY);
+ initAsyncHandlerThreadPools(configuration, allConfiguredNS, true,
+ DFS_ROUTER_ASYNC_RPC_OBSERVER_HANDLER_COUNT_KEY,
+ DFS_ROUTER_ASYNC_RPC_OBSERVER_HANDLER_COUNT_DEFAULT,
nsAsyncObserverHandlerCount,
+ asyncRouterOBHandlerExecutors);
+
initAsyncResponderThreadPools(configuration);
}
- private void initAsyncHandlerThreadPools(Configuration configuration,
- Set<String> allConfiguredNS, Map<String, Integer> nsAsyncHandlerCount) {
- LOG.info("Initializing asynchronous handler thread pools");
+ /**
+ * Initializes async handler executors for each configured namespace.
+ *
+ * @param configuration router configuration
+ * @param allConfiguredNS set of all configured namespace IDs
+ * @param useObserver true if these handlers serve observer namenodes
+ * @param handlerCountKey key for the default handler count
+ * @param handlerCountDefault default handler count
+ * @param nsAsyncHandlerCount per-namespace handler counts
+ * @param executors executor map to populate
+ */
+ private void initAsyncHandlerThreadPools(Configuration configuration,
Set<String> allConfiguredNS,
+ boolean useObserver, String handlerCountKey, int handlerCountDefault,
+ Map<String, Integer> nsAsyncHandlerCount, Map<String,
ThreadPoolExecutor> executors) {
+ String namenodeTypeForLogging = useObserver ? "Observer Namenode" :
"Active Namenode";
int asyncQueueSize = configuration.getInt(DFS_ROUTER_ASYNC_RPC_QUEUE_SIZE,
DFS_ROUTER_ASYNC_RPC_QUEUE_SIZE_DEFAULT);
if (asyncQueueSize < 1) {
throw new IllegalArgumentException("Async queue size must be at least
1");
}
- int asyncHandlerCountDefault =
configuration.getInt(DFS_ROUTER_ASYNC_RPC_HANDLER_COUNT_KEY,
- DFS_ROUTER_ASYNC_RPC_HANDLER_COUNT_DEFAULT);
+ int asyncHandlerCountDefault = configuration.getInt(handlerCountKey,
handlerCountDefault);
+ // Skip separate observer handlers and let them share with active if not
configured
+ if (asyncHandlerCountDefault < 1 && nsAsyncHandlerCount.isEmpty() &&
useObserver) {
+ LOG.info("Async observer handlers are not configured, skipping...");
+ return;
+ } else {
+ useSeparateAsyncRouterOBHandlerExecutors = true;
+ }
Review Comment:
oof, good catch. Fixed and added unit test
> [ARR] Separate router async handlers for active and observer
> ------------------------------------------------------------
>
> Key: HDFS-17935
> URL: https://issues.apache.org/jira/browse/HDFS-17935
> Project: Hadoop HDFS
> Issue Type: Improvement
> Components: hdfs, rbf
> Reporter: Felix N
> Assignee: Felix N
> Priority: Minor
> Labels: pull-request-available
>
> Most router traffic are reads, and currently all traffic to a namespace is
> handled by one same executor, so high observer load can increase competition
> for active ops. It would be beneficial to split the handler pools into 2
> separate sets, one for active, one for observer.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
