[
https://issues.apache.org/jira/browse/HADOOP-5318?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12734321#action_12734321
]
Todd Lipcon commented on HADOOP-5318:
-------------------------------------
Scott: I think if we add buffering it should be at the "client" level - eg in
SequenceFileOutputStream or anywhere else that very small records will be
written at a high rate. Doing so in the FSDataOutputStream seems heavy-handed,
since we don't know what size the writes will usually be, and I'd imagine the
usual is much larger than this pathological test case we used for the benchmark.
I'm also not sure I see your point about the L1 cache - in between the small
writes there may be arbitrary amounts of computation (eg in the mapper
function) that would evict the buffer from cache.
> Poor IO Performance due to AtomicLong operations
> ------------------------------------------------
>
> Key: HADOOP-5318
> URL: https://issues.apache.org/jira/browse/HADOOP-5318
> Project: Hadoop Common
> Issue Type: Bug
> Affects Versions: 0.19.0
> Environment: 2x quad core xeon linux 64 bit
> Reporter: Ben Maurer
> Assignee: Todd Lipcon
> Attachments: buf.patch, buffer-output.patch, hadoop-5318.txt,
> hadoop-5318.txt, Rplot001.png, TestWriteConcurrency.java,
> TestWriteConcurrency.java, TestWriteConcurrency.java
>
>
> The AtomicLong operations in counting file system statistics can cause high
> levels of contention with multiple threads. This test demonstrates having
> multiple threads writing to different sequence files:
> {code:java}
> import java.io.IOException;
> import org.apache.hadoop.conf.Configuration;
> import org.apache.hadoop.fs.FileSystem;
> import org.apache.hadoop.fs.Path;
> import org.apache.hadoop.io.ByteWritable;
> import org.apache.hadoop.io.SequenceFile;
> import org.apache.hadoop.io.SequenceFile.Writer;
> import org.apache.hadoop.io.SequenceFile.CompressionType;
> public class Test {
> public static void main(String[] args) throws IOException {
> final Configuration c = new Configuration();
> final FileSystem fs = FileSystem.get(c);
>
> final int NUM = 1000*1000;
> for (int i = 0; i < Integer.valueOf(args[0]); i ++) {
> final int ii = i;
> new Thread(new Runnable() {
> @Override
> public void run() {
>
> try {
> Writer f =
> SequenceFile.createWriter(fs, c, new Path("/test/" + ii ),
> ByteWritable.class, ByteWritable.class, CompressionType.NONE);
> ByteWritable v = new
> ByteWritable();
>
> long time =
> System.currentTimeMillis();
> for (int i = 0; i < NUM; i ++)
> f.append(v, v);
> f.close();
> long end =
> System.currentTimeMillis();
>
> System.out.printf("%d opartions
> in %d msec. %f/second\n", NUM, end - time, (float)(1000 * NUM)/(end - time));
>
> } catch (Exception e) {
> // TODO Auto-generated catch
> block
> e.printStackTrace();
> }
>
> }
> }).start();
> }
> }
> }
> {code}
> The results of this benchmark are
> {code}
> ==== 1 threads ====
> 1000000 opartions in 1431 msec. 698812.000000/second
> ==== 2 threads ====
> 1000000 opartions in 3001 msec. 333222.250000/second
> 1000000 opartions in 2985 msec. 335008.375000/second
> ==== 3 threads ====
> 1000000 opartions in 4923 msec. 203128.171875/second
> 1000000 opartions in 4924 msec. 203086.921875/second
> 1000000 opartions in 4981 msec. 200762.906250/second
> ==== 4 threads ====
> 1000000 opartions in 6716 msec. 148898.156250/second
> 1000000 opartions in 7048 msec. 141884.218750/second
> 1000000 opartions in 7342 msec. 136202.671875/second
> 1000000 opartions in 7344 msec. 136165.578125/second
> ==== 5 threads ====
> 1000000 opartions in 10366 msec. 96469.226563/second
> 1000000 opartions in 11085 msec. 90212.000000/second
> 1000000 opartions in 11121 msec. 89919.968750/second
> 1000000 opartions in 11464 msec. 87229.585938/second
> 1000000 opartions in 11538 msec. 86670.132813/second
> ==== 6 threads ====
> 1000000 opartions in 16513 msec. 60558.347656/second
> 1000000 opartions in 17704 msec. 56484.410156/second
> 1000000 opartions in 18219 msec. 54887.753906/second
> 1000000 opartions in 18550 msec. 53908.355469/second
> 1000000 opartions in 18605 msec. 53748.992188/second
> 1000000 opartions in 18663 msec. 53581.953125/second
> ==== 7 threads ====
> 1000000 opartions in 22207 msec. 45030.847656/second
> 1000000 opartions in 23275 msec. 42964.554688/second
> 1000000 opartions in 23484 msec. 42582.183594/second
> 1000000 opartions in 24378 msec. 41020.593750/second
> 1000000 opartions in 24425 msec. 40941.656250/second
> 1000000 opartions in 24533 msec. 40761.421875/second
> 1000000 opartions in 24645 msec. 40576.183594/second
> ==== 8 threads ====
> 1000000 opartions in 26375 msec. 37914.691406/second
> 1000000 opartions in 26420 msec. 37850.113281/second
> 1000000 opartions in 26532 msec. 37690.335938/second
> 1000000 opartions in 26670 msec. 37495.312500/second
> 1000000 opartions in 29772 msec. 33588.605469/second
> 1000000 opartions in 29859 msec. 33490.738281/second
> 1000000 opartions in 30098 msec. 33224.800781/second
> 1000000 opartions in 30082 msec. 33242.468750/second
> {code}
> However, if I comment out the file system statistics increments, the
> benchmark improves to:
> {code}
> ==== 1 threads ====
> 1000000 opartions in 1194 msec. 837520.937500/second
> ==== 2 threads ====
> 1000000 opartions in 1433 msec. 697836.687500/second
> 1000000 opartions in 1433 msec. 697836.687500/second
> ==== 3 threads ====
> 1000000 opartions in 1643 msec. 608642.750000/second
> 1000000 opartions in 1643 msec. 608642.750000/second
> 1000000 opartions in 1639 msec. 610128.125000/second
> ==== 4 threads ====
> 1000000 opartions in 1886 msec. 530222.687500/second
> 1000000 opartions in 1886 msec. 530222.687500/second
> 1000000 opartions in 1886 msec. 530222.687500/second
> 1000000 opartions in 1899 msec. 526592.937500/second
> ==== 5 threads ====
> 1000000 opartions in 2065 msec. 484261.500000/second
> 1000000 opartions in 2066 msec. 484027.093750/second
> 1000000 opartions in 2067 msec. 483792.937500/second
> 1000000 opartions in 2066 msec. 484027.093750/second
> 1000000 opartions in 2066 msec. 484027.093750/second
> ==== 6 threads ====
> 1000000 opartions in 2151 msec. 464900.031250/second
> 1000000 opartions in 2111 msec. 473709.156250/second
> 1000000 opartions in 2153 msec. 464468.187500/second
> 1000000 opartions in 2114 msec. 473036.906250/second
> 1000000 opartions in 2113 msec. 473260.781250/second
> 1000000 opartions in 2112 msec. 473484.843750/second
> ==== 7 threads ====
> 1000000 opartions in 2368 msec. 422297.312500/second
> 1000000 opartions in 2334 msec. 428449.000000/second
> 1000000 opartions in 2332 msec. 428816.468750/second
> 1000000 opartions in 2330 msec. 429184.562500/second
> 1000000 opartions in 2332 msec. 428816.468750/second
> 1000000 opartions in 2375 msec. 421052.625000/second
> 1000000 opartions in 2394 msec. 417710.937500/second
> ==== 8 threads ====
> 1000000 opartions in 2517 msec. 397298.375000/second
> 1000000 opartions in 2538 msec. 394011.031250/second
> 1000000 opartions in 2538 msec. 394011.031250/second
> 1000000 opartions in 2538 msec. 394011.031250/second
> 1000000 opartions in 2539 msec. 393855.843750/second
> 1000000 opartions in 2614 msec. 382555.468750/second
> 1000000 opartions in 2666 msec. 375093.781250/second
> 1000000 opartions in 2701 msec. 370233.250000/second
> {code}
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
