[英]How to improve Neo4j 2.0 cypher/ExecutionResult performance under heavy load?
Background: We are noticing a drop in performance while retrieving data out of an ExecutionResult as number of concurrent threads go up. 背景:随着并发线程数的增加,我们注意到从ExecutionResult中检索数据时性能下降。 Our production application has 200 worker threads, uses Neo4j 2.0.0 Community in embedded mode.
我们的生产应用程序具有200个工作线程,以嵌入式模式使用Neo4j 2.0.0社区。 eg in milli-seconds.
例如,以毫秒为单位。
Sample output of program (filtering results for 1 of the threads): 程序的示例输出(过滤其中一个线程的结果):
2013-12-23 14:39:31,137 [main] INFO net.ahm.graph.CypherLab - >>>>>>>>>>>>>>>>>>>>>>>>>>>>> NUMBER OF PARALLEL CYPHER EXECUTIONS: 1
2013-12-23 14:39:31,137 [main] INFO net.ahm.graph.CypherLab - >>>> STARTED GRAPHDB
2013-12-23 14:39:39,203 [main] INFO net.ahm.graph.CypherLab - >>>> CREATED NODES
2013-12-23 14:39:43,510 [main] INFO net.ahm.graph.CypherLab - >>>> WARMED UP
2013-12-23 14:39:43,510 [pool-1-thread-1] INFO net.ahm.graph.CypherLab - >>>> CYPHER TOOK: 0 m-secs
2013-12-23 14:39:43,698 [pool-1-thread-1] INFO net.ahm.graph.CypherLab - >>>> GETTING RESULTS TOOK: 188 m-secs
2013-12-23 14:39:43,698 [pool-1-thread-1] INFO net.ahm.graph.CypherLab - >>>> CYPHER RETURNED ROWS: 50000
2013-12-23 14:39:43,698 [Thread-4] INFO net.ahm.graph.CypherLab - ### GRAPHDB SHUTDOWNHOOK INVOKED !!!
2013-12-23 14:40:10,470 [main] INFO net.ahm.graph.CypherLab - >>>>>>>>>>>>>>>>>>>>>>>>>>>>> NUMBER OF PARALLEL CYPHER EXECUTIONS: 10
...
2013-12-23 14:40:23,985 [pool-1-thread-1] INFO net.ahm.graph.CypherLab - >>>> CYPHER TOOK: 1 m-secs
2013-12-23 14:40:25,219 [pool-1-thread-1] INFO net.ahm.graph.CypherLab - >>>> GETTING RESULTS TOOK: 188 m-secs
2013-12-23 14:40:25,219 [pool-1-thread-1] INFO net.ahm.graph.CypherLab - >>>> CYPHER RETURNED ROWS: 50000
2013-12-23 14:40:25,234 [Thread-4] INFO net.ahm.graph.CypherLab - ### GRAPHDB SHUTDOWNHOOK INVOKED !!!
2013-12-23 14:41:28,850 [main] INFO net.ahm.graph.CypherLab - >>>>>>>>>>>>>>>>>>>>>>>>>>>>> NUMBER OF PARALLEL CYPHER EXECUTIONS: 50
...
2013-12-23 14:41:41,781 [pool-1-thread-1] INFO net.ahm.graph.CypherLab - >>>> CYPHER TOOK: 1 m-secs
2013-12-23 14:41:45,720 [pool-1-thread-1] INFO net.ahm.graph.CypherLab - >>>> GETTING RESULTS TOOK: 2481 m-secs
2013-12-23 14:41:45,720 [pool-1-thread-1] INFO net.ahm.graph.CypherLab - >>>> CYPHER RETURNED ROWS: 50000
2013-12-23 14:41:46,855 [Thread-4] INFO net.ahm.graph.CypherLab - ### GRAPHDB SHUTDOWNHOOK INVOKED !!!
2013-12-23 14:44:09,267 [main] INFO net.ahm.graph.CypherLab - >>>>>>>>>>>>>>>>>>>>>>>>>>>>> NUMBER OF PARALLEL CYPHER EXECUTIONS: 100
...
2013-12-23 14:44:22,077 [pool-1-thread-1] INFO net.ahm.graph.CypherLab - >>>> CYPHER TOOK: 1 m-secs
2013-12-23 14:44:30,915 [pool-1-thread-1] INFO net.ahm.graph.CypherLab - >>>> GETTING RESULTS TOOK: 4466 m-secs
2013-12-23 14:44:30,915 [pool-1-thread-1] INFO net.ahm.graph.CypherLab - >>>> CYPHER RETURNED ROWS: 50000
2013-12-23 14:44:31,680 [Thread-4] INFO net.ahm.graph.CypherLab - ### GRAPHDB SHUTDOWNHOOK INVOKED !!!
Test Program: 测试程序:
package net.ahm.graph;
import java.io.File;
import java.util.Map;
import java.util.concurrent.CountDownLatch;
import java.util.concurrent.ExecutorService;
import java.util.concurrent.Executors;
import java.util.concurrent.TimeUnit;
import org.apache.log4j.Logger;
import org.neo4j.cypher.javacompat.ExecutionEngine;
import org.neo4j.cypher.javacompat.ExecutionResult;
import org.neo4j.graphdb.DynamicLabel;
import org.neo4j.graphdb.GraphDatabaseService;
import org.neo4j.graphdb.Node;
import org.neo4j.graphdb.RelationshipType;
import org.neo4j.graphdb.Transaction;
import org.neo4j.graphdb.factory.GraphDatabaseFactory;
import org.neo4j.graphdb.factory.GraphDatabaseSettings;
import org.neo4j.graphdb.schema.IndexDefinition;
import org.neo4j.graphdb.schema.Schema;
import org.neo4j.kernel.impl.util.FileUtils;
import org.neo4j.kernel.impl.util.StringLogger;
public class CypherLab {
private static final Logger LOG = Logger.getLogger(CypherLab.class);
private final static int CONCURRENCY = 100;
public static void main(String[] args) throws Exception {
FileUtils.deleteRecursively(new File("graphdb"));
final GraphDatabaseService graphDb = new GraphDatabaseFactory().newEmbeddedDatabaseBuilder("graphdb")
.setConfig(GraphDatabaseSettings.use_memory_mapped_buffers, "true").setConfig(GraphDatabaseSettings.cache_type, "strong")
.newGraphDatabase();
registerShutdownHook(graphDb);
LOG.info(">>>>>>>>>>>>>>>>>>>>>>>>>>>>> NUMBER OF PARALLEL CYPHER EXECUTIONS: " + CONCURRENCY);
LOG.info(">>>> STARTED GRAPHDB");
createIndex("Parent", "name", graphDb);
createIndex("Child", "name", graphDb);
try (Transaction tx = graphDb.beginTx()) {
Node parent = graphDb.createNode(DynamicLabel.label("Parent"));
parent.setProperty("name", "parent");
for (int i = 0; i < 50000; i++) {
Node child = graphDb.createNode(DynamicLabel.label("Child"));
child.setProperty("name", "child" + i);
parent.createRelationshipTo(child, RelationshipTypes.PARENT_CHILD);
}
tx.success();
}
LOG.info(">>>> CREATED NODES");
final ExecutionEngine engine = new ExecutionEngine(graphDb, StringLogger.SYSTEM);
for (int i = 0; i < 10; i++) {
try (Transaction tx = graphDb.beginTx()) {
ExecutionResult result = engine.execute("match (n:Parent)-[:PARENT_CHILD]->(m:Child) return n.name, m.name");
for (Map<String, Object> row : result) {
assert ((String) row.get("n.name") != null);
assert ((String) row.get("m.name") != null);
}
tx.success();
}
}
LOG.info(">>>> WARMED UP");
ExecutorService es = Executors.newFixedThreadPool(CONCURRENCY);
final CountDownLatch cdl = new CountDownLatch(CONCURRENCY);
for (int i = 0; i < CONCURRENCY; i++) {
es.execute(new Runnable() {
@Override
public void run() {
try (Transaction tx = graphDb.beginTx()) {
long time = System.currentTimeMillis();
ExecutionResult result = engine.execute("match (n:Parent)-[:PARENT_CHILD]->(m:Child) return n.name, m.name");
LOG.info(">>>> CYPHER TOOK: " + (System.currentTimeMillis() - time) + " m-secs");
int count = 0;
time = System.currentTimeMillis();
for (Map<String, Object> row : result) {
assert ((String) row.get("n.name") != null);
assert ((String) row.get("m.name") != null);
count++;
}
LOG.info(">>>> GETTING RESULTS TOOK: " + (System.currentTimeMillis() - time) + " m-secs");
tx.success();
LOG.info(">>>> CYPHER RETURNED ROWS: " + count);
} catch (Throwable t) {
LOG.error(t);
} finally {
cdl.countDown();
}
}
});
}
cdl.await();
es.shutdown();
}
private static void createIndex(String label, String propertyName, GraphDatabaseService graphDb) {
IndexDefinition indexDefinition;
try (Transaction tx = graphDb.beginTx()) {
Schema schema = graphDb.schema();
indexDefinition = schema.indexFor(DynamicLabel.label(label)).on(propertyName).create();
tx.success();
}
try (Transaction tx = graphDb.beginTx()) {
Schema schema = graphDb.schema();
schema.awaitIndexOnline(indexDefinition, 10, TimeUnit.SECONDS);
tx.success();
}
}
private static void registerShutdownHook(final GraphDatabaseService graphDb) {
Runtime.getRuntime().addShutdownHook(new Thread() {
@Override
public void run() {
LOG.info("### GRAPHDB SHUTDOWNHOOK INVOKED !!!");
graphDb.shutdown();
}
});
}
private enum RelationshipTypes implements RelationshipType {
PARENT_CHILD
}
}
Should be better when this commit is merged in . 合并此提交时应该会更好。 Which will be released as part of 2.0.1 There are still some other, smaller choke points.
它将作为2.0.1的一部分发布。还有其他一些较小的阻塞点。
Can you try to limit your webserver-threads to the # of cores times (or # of cores * 2) ? 您可以尝试将您的Web服务器线程限制为内核次数(或内核数量* 2)吗? And see if that helps?
看看是否有帮助?
My understanding is that after warmup and having the hot dataset in the cache it is only CPU bound, not I/O bound anymore for reads. 我的理解是,在预热并且将热数据集存储在缓存中之后,它仅受CPU约束,不再受I / O绑定。 So you starve CPUs and workers with too many threads.
因此,您会使CPU和工作人员的线程过多。
If I run your test with 8 and 100 cores I get these distributions for executing the query and fetching 50k results: 如果我使用8个和100个内核运行测试,则会得到以下分布以执行查询并获取50k结果:
Code and detailed Histograms: https://gist.github.com/jexp/a164f6cf9686b8125872 代码和详细的直方图: https : //gist.github.com/jexp/a164f6cf9686b8125872
声明:本站的技术帖子网页,遵循CC BY-SA 4.0协议,如果您需要转载,请注明本站网址或者原文地址。任何问题请咨询:yoyou2525@163.com.