imbajin commented on code in PR #3355: URL: https://github.com/apache/hugegraph/pull/3355#discussion_r4300663949
########## hugegraph-store/hg-store-core/src/main/java/org/apache/hugegraph/store/options/RocksdbMemoryBudget.java: ########## @@ -0,0 +1,398 @@ +/* + * 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.hugegraph.store.options; + +import java.io.IOException; +import java.lang.management.ManagementFactory; +import java.nio.charset.StandardCharsets; +import java.nio.file.Files; +import java.nio.file.Path; +import java.nio.file.Paths; +import java.util.ArrayList; +import java.util.List; + +import com.sun.management.HotSpotDiagnosticMXBean; +import com.sun.management.OperatingSystemMXBean; + +import lombok.extern.slf4j.Slf4j; + +/** + * Resolves rocksdb.total_memory_size, the budget RaftRocksdbOptions splits into the + * write buffer manager's cache and the block cache. Both caches are off-heap, so an + * unset or "0" budget is sized from the memory left after the heap, direct memory and + * the fixed raft log cache, instead of a fixed number that can exceed a container limit. + */ +@Slf4j +public final class RocksdbMemoryBudget { + + public static final String KEY = "total_memory_size"; + + static final long MIB = 1024L * 1024L; + // The 1 GiB raft log block cache (RaftRocksdbOptions) plus thread stacks, metaspace + // and other native allocations that sit outside the budget + static final long RESERVED = 1024 * MIB + 512 * MIB; + static final long MIN_AUTO = 256 * MIB; + // Direct memory counted when -XX:MaxDirectMemorySize is unset. The JVM then allows + // as much as the heap, but the Store's gRPC/Netty buffers do not grow with the heap, + // and counting a 36 GB heap twice would leave rocksdb the floor on a large host + static final long UNSET_DIRECT = 1024 * MIB; + // RaftRocksdbOptions gives the write buffer manager 2/3 of the budget, and a 0-byte + // write buffer manager means no memtable limit at all + static final long MIN_EXPLICIT = MIB; + // The value shipped before auto-sizing; a large host never gets more than it had + static final long MAX_AUTO = 32_000_000_000L; + + private RocksdbMemoryBudget() { + } + + /** + * Returns the configured budget in bytes, or an automatic one when the value is + * missing, blank or 0. + */ + public static long resolve(String configured) { + long value = isAuto(configured) ? 0L : parse(configured); + long heap = Runtime.getRuntime().maxMemory(); + long physical = physicalMemory(); + if (value > 0) { + if (value < MIN_EXPLICIT) { + throw new IllegalArgumentException(String.format( + "rocksdb.total_memory_size must be 0 (auto) or at least %s bytes, " + + "got %s", MIN_EXPLICIT, value)); + } + if (value < MIN_AUTO) { + log.warn("rocksdb.total_memory_size={} is below the {} bytes auto-sizing " + + "never goes under", value, MIN_AUTO); + } + if (value == MAX_AUTO) { + // Most likely an application-pd.yml kept from before auto-sizing + log.warn("rocksdb.total_memory_size={} is the fixed value shipped before " + + "auto-sizing; set it to 0 or remove it to size the budget from " + + "the available memory", value); + } + long direct = maxDirectMemory(heap); + long left = physical - heap - direct - RESERVED; + if (physical > 0 && value > left) { + log.warn("rocksdb.total_memory_size={} exceeds the {} bytes left after the " + + "heap, direct memory and the {} bytes reserve (physical={}, heap={}, " + + "direct={}); the process may be OOM-killed", + value, left, RESERVED, physical, heap, direct); + } + return value; + } + long direct = maxDirectMemory(heap); + long budget = auto(physical, heap, direct); + log.info("rocksdb.total_memory_size is unset or 0, auto-sizing to {} bytes " + + "(physical={}, heap={}, direct={})", budget, physical, heap, direct); + if (physical > 0 && physical - heap - direct - RESERVED < 2 * MIN_AUTO) { + log.warn("Heap, direct memory and the {} bytes reserve leave too little of the " + + "{} bytes physical memory, rocksdb gets the {} bytes floor; lower -Xmx " + + "or -XX:MaxDirectMemorySize", RESERVED, physical, MIN_AUTO); + } + return budget; + } + + static boolean isAuto(String configured) { + return configured == null || configured.trim().isEmpty(); + } + + static long parse(String configured) { + try { + long value = Long.parseLong(configured.trim()); + if (value < 0) { + throw new NumberFormatException("negative"); + } + return value; + } catch (NumberFormatException e) { + throw new IllegalArgumentException(String.format( + "rocksdb.total_memory_size must be a number of bytes, got '%s'", + configured), e); + } + } + + /** + * Half of what is left after the heap, direct memory and the reserve, clamped to + * [MIN_AUTO, MAX_AUTO]. Falls back to the heap size, the previous default, when + * physical memory is unknown. + */ + static long auto(long physical, long heap, long direct) { + if (physical <= 0) { + return Math.min(Math.max(heap, MIN_AUTO), MAX_AUTO); + } + long available = physical - heap - direct - RESERVED; + return Math.min(Math.max(available / 2, MIN_AUTO), MAX_AUTO); + } + + /** + * The raft disruptor buffer size AppConfig derives from the budget when + * raft.disruptorBufferSize is 0: the budget in GB, at least 1, rounded to a power + * of two, times 32. + */ + public static int disruptorBufferSize(long totalMemory) { + // A budget under 1 GB would otherwise give log(0) and a zero-sized buffer + int size = (int) Math.max(1, totalMemory / 1000 / 1000 / 1000); + return (int) Math.pow(2, Math.round(Math.log(size) / Math.log(2))) * 32; + } + + /** + * The write_buffer_size AppConfig derives when the key is unset or 0: a thousandth + * of the budget, but never below the 1 MiB rocksdb.write_buffer_size accepts + * (RocksDBOptions.MEMTABLE_SIZE), which an auto-sized budget under 1 GiB would + * otherwise fall under. + */ + public static long writeBufferSize(long totalMemory) { + return Math.max(totalMemory / 1000, MIB); + } + + private static long physicalMemory() { + long jvm; + try { + // Reports the cgroup v1/v2 memory limit inside a container, but older JDK 17 + // updates may not see a limit set only on a parent cgroup + OperatingSystemMXBean os = ManagementFactory.getPlatformMXBean( + OperatingSystemMXBean.class); + jvm = os == null ? -1L : os.getTotalMemorySize(); + } catch (RuntimeException | LinkageError e) { + log.warn("Failed to read physical memory size", e); + jvm = -1L; + } + long cgroup = cgroupMemoryLimit(Paths.get("/proc/self/cgroup"), + Paths.get("/proc/self/mountinfo"), + Paths.get("/sys/fs/cgroup")); + return effectivePhysical(jvm, cgroup); + } + + /** + * The smaller of what the JVM reports and the cgroup limit, ignoring unknown (<= 0) + * values. + */ + static long effectivePhysical(long jvm, long cgroup) { + if (cgroup <= 0) { + return jvm; + } + return jvm <= 0 ? cgroup : Math.min(jvm, cgroup); + } + + /** + * The lowest memory limit on this process's cgroup or any visible ancestor, or -1 + * when there is none or it can't be read. The membership path from /proc/self/cgroup + * is rebased onto every mount of the controller in /proc/self/mountinfo whose root + * contains it, so a cgroupfs mounted from a subtree (root /kubepods at /sys/fs/cgroup) + * is walked from the right directory, and the lowest limit over all those mounts is + * taken: a leaf-only or unrelated mount listed first can't hide a fuller view listed + * later. Without a matching mount the default root is used. cgroup v2 + * walks memory.max from the process's cgroup up to the mount; cgroup v1 reads + * hierarchical_memory_limit, which already accounts for ancestors. A limit above the + * mount root, as inside a private cgroup namespace, stays invisible. + */ + static long cgroupMemoryLimit(Path procSelfCgroup, Path mountinfo, Path defaultRoot) { + List<String[]> v2Mounts = cgroupMounts(mountinfo, true); + List<String[]> v1Mounts = cgroupMounts(mountinfo, false); + List<String> lines; + try { + lines = Files.readAllLines(procSelfCgroup, StandardCharsets.UTF_8); + } catch (IOException | RuntimeException e) { + return -1L; + } + long limit = -1L; + for (String line : lines) { + String[] parts = line.split(":", 3); + if (parts.length < 3) { + continue; + } + boolean v2 = parts[0].equals("0") && parts[1].isEmpty(); + if (!v2 && !("," + parts[1] + ",").contains(",memory,")) { + continue; + } + boolean matched = false; + for (String[] mount : v2 ? v2Mounts : v1Mounts) { + if (!contains(mount[0], parts[2])) { + // A mount of another subtree says nothing about this cgroup + continue; + } + matched = true; + Path point = Paths.get(mount[1]); + String relative = rebase(parts[2], mount[0]); + limit = lowest(limit, v2 ? cgroupV2Limit(point, relative) : + cgroupV1Limit(point, relative)); + } + if (!matched) { Review Comment: ⚠️ Important: This fallback misses the process cgroup in the namespace layout documented by [cgroup_namespaces(7)](https://man7.org/linux/man-pages/man7/cgroup_namespaces.7.html): /proc/self/cgroup reports 0::/ while a cgroup2 mount created before the namespace reports root /..; contains rejects that mount, then this code reads only the default mount root. If the namespace root is a child with a tighter memory.max than its parent, the process limit is missed and auto-sizing can allocate an oversized RocksDB budget. Please resolve this mount layout to the process cgroup, or fail safe when it cannot be matched, and add a fixture with an unlimited parent and limited child. -- 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. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected] --------------------------------------------------------------------- To unsubscribe, e-mail: [email protected] For additional commands, e-mail: [email protected]
