azagrebin commented on a change in pull request #9693: [FLINK-13984] Separate on-heap and off-heap managed memory pools URL: https://github.com/apache/flink/pull/9693#discussion_r336907657
########## File path: flink-runtime/src/main/java/org/apache/flink/runtime/util/KeyedBudgetManager.java ########## @@ -0,0 +1,238 @@ +/* + * Licensed to the Apache Software Foundation (ASF) under one + * or more contributor license agreements. See the NOTICE file + * distributed with this work for additional information + * regarding copyright ownership. The ASF licenses this file + * to you under the Apache License, Version 2.0 (the + * "License"); you may not use this file except in compliance + * with the License. You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +package org.apache.flink.runtime.util; + +import org.apache.flink.types.Either; +import org.apache.flink.util.Preconditions; + +import javax.annotation.concurrent.GuardedBy; +import javax.annotation.concurrent.ThreadSafe; + +import java.util.Collections; +import java.util.HashMap; +import java.util.Map; +import java.util.Map.Entry; + +/** + * Manages {@code long} available budget per key (allocation/release). + * + * <p>This manager gets a certain maximum {@code long} budget per key. + * Users can allocate some budget for some key and release it later. + * The manager keeps track of allocated/released budget and prevents from over-allocating. + * + * <p>There is also a paged type of allocation where a certain number of pages can be allocated from a set of keys. + * The page has its budget size. The manager allocates randomly from all keys of a given set. + * At the end, sum of pages allocated from each key is either requested number of pages or none. + * Only integer number of pages are allocated from each key respecting its available budget (no page spans two or more keys) + * or nothing is allocated reporting the maximum number of pages which could be allocated per each given key at the moment. + * + * @param <K> type of the budget key + */ +@ThreadSafe +public class KeyedBudgetManager<K> { + private final Map<K, Long> maxBudgetByKey; + + private final long defaultPageSize; + + private final long totalNumberOfPages; + + @GuardedBy("lock") + private final Map<K, Long> availableBudgetByKey; + + private final Object lock = new Object(); Review comment: This mixed allocation was specifically implemented for DataSet API, to avoid its regression, because it does not care about the memory type of the allocated segments and should be able to use everything transparently. This should be valid for all use cases where only segments are needed. True, at them moment, the lock will create contention for all types of allocation. We should see in future whether it is a performance problem because currently it is not on critical per record processing. I also thought about this approach with `ConcurrentHashMap`. The problem is that the semantics of the method is either allocate from all keys or nothing. This means that it has to be some kind of optimistic transaction where we firstly allocate then check whether the target number has been reach and then revert if not (this would probably require more concurrent tests). The current solution is simpler, I would keep it for now. ---------------------------------------------------------------- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: [email protected] With regards, Apache Git Services
