+/EE360P

distributed systems

git clone https://git.julia.blue/+/EE360P

vijaygarg1init07ca65d

master
4.4 KiB114 linesraw
1<html>
2
3<!--
4   Licensed to the Apache Software Foundation (ASF) under one or more
5   contributor license agreements.  See the NOTICE file distributed with
6   this work for additional information regarding copyright ownership.
7   The ASF licenses this file to You under the Apache License, Version 2.0
8   (the "License"); you may not use this file except in compliance with
9   the License.  You may obtain a copy of the License at
10
11       http://www.apache.org/licenses/LICENSE-2.0
12
13   Unless required by applicable law or agreed to in writing, software
14   distributed under the License is distributed on an "AS IS" BASIS,
15   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
16   See the License for the specific language governing permissions and
17   limitations under the License.
18-->
19
20<body>
21This package consists of 3 map/reduce applications for Hadoop to
22compete in the annual <a
23href="http://www.hpl.hp.com/hosted/sortbenchmark" target="_top">terabyte sort</a>
24competition.
25
26<ul>
27<li><b>TeraGen</b> is a map/reduce program to generate the data.
28<li><b>TeraSort</b> samples the input data and uses map/reduce to
29    sort the data into a total order.
30<li><b>TeraValidate</b> is a map/reduce program that validates the
31    output is sorted.
32</ul>
33
34<p>
35
36<b>TeraGen</b> generates output data that is byte for byte
37equivalent to the C version including the newlines and specific
38keys. It divides the desired number of rows by the desired number of
39tasks and assigns ranges of rows to each map. The map jumps the random
40number generator to the correct value for the first row and generates
41the following rows.
42
43<p>
44
45<b>TeraSort</b> is a standard map/reduce sort, except for a custom
46partitioner that uses a sorted list of <i>N-1</i> sampled keys that define
47the key range for each reduce. In particular, all keys such that
48<i>sample[i-1] &lt;= key &lt; sample[i]</i> are sent to reduce
49<i>i</i>. This guarantees that the output of reduce <i>i</i> are all
50less than the output of reduce <i>i+1</i>. To speed up the
51partitioning, the partitioner builds a two level trie that quickly
52indexes into the list of sample keys based on the first two bytes of
53the key. TeraSort generates the sample keys by sampling the input
54before the job is submitted and writing the list of keys into HDFS.
55The input and output format, which are used by all 3 applications,
56read and write the text files in the right format. The output of the
57reduce has replication set to 1, instead of the default 3, because the
58contest does not require the output data be replicated on to multiple
59nodes.  
60
61<p>
62
63<b>TeraValidate</b> ensures that the output is globally sorted. It
64creates one map per a file in the output directory and each map ensures that
65each key is less than or equal to the previous one. The map also generates
66records with the first and last keys of the file and the reduce
67ensures that the first key of file <i>i</i> is greater that the last key of
68file <i>i-1</i>. Any problems are reported as output of the reduce with the
69keys that are out of order.
70
71<p>
72
73In May 2008, Owen O'Malley ran this code on a 910 node cluster and
74sorted the 10 billion records (1 TB) in 209 seconds (3.48 minutes) to
75win the annual general purpose (daytona)
76<a href="http://www.hpl.hp.com/hosted/sortbenchmark/">terabyte sort
77benchmark</a>.
78
79<p>
80
81The cluster statistics were:
82<ul>
83<li> 910 nodes
84<li> 4 dual core Xeons @ 2.0ghz per a node
85<li> 4 SATA disks per a node
86<li> 8G RAM per a node
87<li> 1 gigabit ethernet on each node
88<li> 40 nodes per a rack
89<li> 8 gigabit ethernet uplinks from each rack to the core
90<li> Red Hat Enterprise Linux Server Release 5.1 (kernel 2.6.18)
91<li> Sun Java JDK 1.6.0_05-b13
92</ul>
93
94<p>
95
96The test was on Hadoop trunk (pre-0.18) patched with <a
97href="http://issues.apache.org/jira/browse/HADOOP-3443">HADOOP-3443</a>
98and <a
99href="http://issues.apache.org/jira/browse/HADOOP-3446">HADOOP-3446</a>,
100which were required to remove intermediate writes to disk.
101TeraGen used
1021800 tasks to generate a total of 10 billion rows in HDFS, with a
103block size of 1024 MB.
104TeraSort was configured with 1800 maps and 1800 reduces, and
105<i>mapreduce.task.io.sort.mb</i>,
106<i>mapreduce.task.io.sort.factor</i>,
107<i>fs.inmemory.size.mb</i>, and task heap size
108sufficient that transient data was never spilled to disk, other at the
109end of the map. The sampler looked at 100,000 keys to determine the
110reduce boundaries, which lead to imperfect balancing with reduce
111outputs ranging from 337 MB to 872 MB.
112
113</body>
114</html>