[jira] [Commented] (GIRAPH-45) Improve the way to keep outgoing messages
[ https://issues.apache.org/jira/browse/GIRAPH-45?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanelfocusedCommentId=13158131#comment-13158131 ] Hyunsik Choi commented on GIRAPH-45: Claudio, Thank for a nice suggestion. That seems a cool idea. However, I concern with the platform dependency of leveldb. The leveldb is built in C++ language. It may give us a burden of the distribution of Giraph. What does anyone else think? Improve the way to keep outgoing messages - Key: GIRAPH-45 URL: https://issues.apache.org/jira/browse/GIRAPH-45 Project: Giraph Issue Type: Improvement Components: bsp Reporter: Hyunsik Choi Assignee: Hyunsik Choi As discussed in GIRAPH-12(http://goo.gl/CE32U), I think that there is a potential problem to cause out of memory when the rate of message generation is higher than the rate of message flush (or network bandwidth). To overcome this problem, we need more eager strategy for message flushing or some approach to spill messages into disk. The below link is Dmitriy's suggestion. https://issues.apache.org/jira/browse/GIRAPH-12?focusedCommentId=13116253page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel#comment-13116253 -- This message is automatically generated by JIRA. If you think it was sent incorrectly, please contact your JIRA administrators: https://issues.apache.org/jira/secure/ContactAdministrators!default.jspa For more information on JIRA, see: http://www.atlassian.com/software/jira
[jira] [Commented] (GIRAPH-68) Implement a Graph Generator
[ https://issues.apache.org/jira/browse/GIRAPH-68?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanelfocusedCommentId=13152185#comment-13152185 ] Hyunsik Choi commented on GIRAPH-68: I missed javadoc. I will reattach the patch including javadoc. Implement a Graph Generator --- Key: GIRAPH-68 URL: https://issues.apache.org/jira/browse/GIRAPH-68 Project: Giraph Issue Type: New Feature Components: benchmark Affects Versions: 0.70.0 Reporter: Hyunsik Choi Assignee: Hyunsik Choi Attachments: GIRAPH-68_1.patch, GIRAPH-68_2.patch To provide users with benchmark environments and to deeply test the input/output system of giraph, we need a graph generator. We will enable the graph generator to generate various kinds of graph data sets by specifying a VertexInputFormat and a VertexOutputFormat. -- This message is automatically generated by JIRA. If you think it was sent incorrectly, please contact your JIRA administrators: https://issues.apache.org/jira/secure/ContactAdministrators!default.jspa For more information on JIRA, see: http://www.atlassian.com/software/jira
[jira] [Commented] (GIRAPH-77) Coordinator should expose a web interface with progress, vertex region assignments, etc.
[ https://issues.apache.org/jira/browse/GIRAPH-77?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanelfocusedCommentId=13152583#comment-13152583 ] Hyunsik Choi commented on GIRAPH-77: I also think that this feature is necessary because we would not depend on MapReduce anymore after we port Giraph to Yarn. Coordinator should expose a web interface with progress, vertex region assignments, etc. Key: GIRAPH-77 URL: https://issues.apache.org/jira/browse/GIRAPH-77 Project: Giraph Issue Type: New Feature Reporter: Jakob Homan It would be nice if the coordinator worker had a web interface that showed progress, splits, etc. during job execution. Right now it would duplicate information currently being exposed through task status, but with the move to YARN, it will be a necessity. It would be great if we could do this in a modern way to avoid the screen-scraping, etc. currently used to get information from most other Hadoop project's web interfaces. The coordinator could announce its address at the beginning or via status updates. -- This message is automatically generated by JIRA. If you think it was sent incorrectly, please contact your JIRA administrators: https://issues.apache.org/jira/secure/ContactAdministrators!default.jspa For more information on JIRA, see: http://www.atlassian.com/software/jira
[jira] [Commented] (GIRAPH-45) Improve the way to keep outgoing messages
[ https://issues.apache.org/jira/browse/GIRAPH-45?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanelfocusedCommentId=13151038#comment-13151038 ] Hyunsik Choi commented on GIRAPH-45: I'm in another time zone. I'm sad to miss the hot party. I consider this problem as Giraph becomes slow, but works well or Giraph cannot deal with some problems or data when the volume of generated messages exceeds the memory capacity. As you mentioned, apparently spilling data to disk is the simplest way to solve this problem. In addition, this way does not affect usual cases if spilling data is started only when the memory is getting tight. Anyway, the discussion is concluded as follows? - Each worker sends outgoing messages in an eager manner (immediately or periodically). - The receiving side spills incoming messages into disk only when the memory is getting tight. Avery, I also agree that storing partitions to disk is a good way to mitigate the memory problem. Also, I think that both ways are compatible and have different effects. The storing partitioning is more efficient if the volume of graph data is very large. Later, if Giraph enables users to choose the options (i.e., spilling, storing to partitions, or both), users can choose some of them according to their programs. Improve the way to keep outgoing messages - Key: GIRAPH-45 URL: https://issues.apache.org/jira/browse/GIRAPH-45 Project: Giraph Issue Type: Improvement Components: bsp Reporter: Hyunsik Choi Assignee: Hyunsik Choi As discussed in GIRAPH-12(http://goo.gl/CE32U), I think that there is a potential problem to cause out of memory when the rate of message generation is higher than the rate of message flush (or network bandwidth). To overcome this problem, we need more eager strategy for message flushing or some approach to spill messages into disk. The below link is Dmitriy's suggestion. https://issues.apache.org/jira/browse/GIRAPH-12?focusedCommentId=13116253page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel#comment-13116253 -- This message is automatically generated by JIRA. If you think it was sent incorrectly, please contact your JIRA administrators: https://issues.apache.org/jira/secure/ContactAdministrators!default.jspa For more information on JIRA, see: http://www.atlassian.com/software/jira
[jira] [Commented] (GIRAPH-79) Change the menu layout of the site
[ https://issues.apache.org/jira/browse/GIRAPH-79?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanelfocusedCommentId=13149304#comment-13149304 ] Hyunsik Choi commented on GIRAPH-79: Gianmarco, I misunderstood your mention. You did not the remove of the project reports. I'm agree that the report is placed deeper in the site. Change the menu layout of the site -- Key: GIRAPH-79 URL: https://issues.apache.org/jira/browse/GIRAPH-79 Project: Giraph Issue Type: Task Components: site Reporter: Hyunsik Choi Labels: site Attachments: GIRAPH-79_1.patch The current site has the basic menu layout generated by maven site plugin. This layout is restricted to embrace new contents. I would like to suggest the following menu layout. http://people.apache.org/~hyunsik/giraph/site/index.html Although the layout includes most existing contents, it has two addition categories, Giraph and Documentation. I think that this layout is simpler and is good to add new contents. Anyone has any other suggestions? -- This message is automatically generated by JIRA. If you think it was sent incorrectly, please contact your JIRA administrators: https://issues.apache.org/jira/secure/ContactAdministrators!default.jspa For more information on JIRA, see: http://www.atlassian.com/software/jira
[jira] [Commented] (GIRAPH-11) Improve the graph distribution of Giraph
[ https://issues.apache.org/jira/browse/GIRAPH-11?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanelfocusedCommentId=13149412#comment-13149412 ] Hyunsik Choi commented on GIRAPH-11: Avery, I'm sorry for delaying the review. Now, I'm digging your patch. That looks great! Based on this work, we can consider some advanced graph partitioner based on the number of edge-cuts on graph partitions. I need about one more day for more investigation because the patch is somewhat complicated for me :) Besides, for the deeper review, I would like to execute the some tests and trace them. Your patch needs the rebase. Could you rebase the patch? Thank you :) Improve the graph distribution of Giraph Key: GIRAPH-11 URL: https://issues.apache.org/jira/browse/GIRAPH-11 Project: Giraph Issue Type: Improvement Affects Versions: 0.70.0 Reporter: Avery Ching Assignee: Avery Ching Attachments: GIRAPH-11.diff Currently, Giraph assumes that the data from the VertexInputFormat is sorted. If the user data is not sorted by the vertex id, they must first run a MapReduce or Pig job to generate a sorted dataset. This is often a bit inconvenient. Giraph graph partitioning is currently range based and there are some advantages and disadvantages of this approach. The proposal of this JIRA would be to allow for both range and hash based partitioning and provide more flexibility to the user. Design goals for the graph distribution: * Allow vertices to be unordered or unordered * Ability to repartition * Select the partitioning scheme based on user needs (i.e. hash or range based) * Ability to provide user-specific hints about partitions Hash-based partitioning * Good vertex balancing across ranges for random data * Bad at vertex id locality Range-based partitioning * Good at vertex id locality * Ability to split ranges easily * Can cause hotspots for hot ranges -- This message is automatically generated by JIRA. If you think it was sent incorrectly, please contact your JIRA administrators: https://issues.apache.org/jira/secure/ContactAdministrators!default.jspa For more information on JIRA, see: http://www.atlassian.com/software/jira
[jira] [Commented] (GIRAPH-11) Improve the graph distribution of Giraph
[ https://issues.apache.org/jira/browse/GIRAPH-11?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanelfocusedCommentId=13149471#comment-13149471 ] Hyunsik Choi commented on GIRAPH-11: Thank you for rebase. Improve the graph distribution of Giraph Key: GIRAPH-11 URL: https://issues.apache.org/jira/browse/GIRAPH-11 Project: Giraph Issue Type: Improvement Affects Versions: 0.70.0 Reporter: Avery Ching Assignee: Avery Ching Attachments: GIRAPH-11.diff Currently, Giraph assumes that the data from the VertexInputFormat is sorted. If the user data is not sorted by the vertex id, they must first run a MapReduce or Pig job to generate a sorted dataset. This is often a bit inconvenient. Giraph graph partitioning is currently range based and there are some advantages and disadvantages of this approach. The proposal of this JIRA would be to allow for both range and hash based partitioning and provide more flexibility to the user. Design goals for the graph distribution: * Allow vertices to be unordered or unordered * Ability to repartition * Select the partitioning scheme based on user needs (i.e. hash or range based) * Ability to provide user-specific hints about partitions Hash-based partitioning * Good vertex balancing across ranges for random data * Bad at vertex id locality Range-based partitioning * Good at vertex id locality * Ability to split ranges easily * Can cause hotspots for hot ranges -- This message is automatically generated by JIRA. If you think it was sent incorrectly, please contact your JIRA administrators: https://issues.apache.org/jira/secure/ContactAdministrators!default.jspa For more information on JIRA, see: http://www.atlassian.com/software/jira
[jira] [Commented] (GIRAPH-11) Improve the graph distribution of Giraph
[ https://issues.apache.org/jira/browse/GIRAPH-11?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanelfocusedCommentId=13148187#comment-13148187 ] Hyunsik Choi commented on GIRAPH-11: That's a huge patch :) I have just started to explore your patch. I will leave some comments (maybe tomorrow). Improve the graph distribution of Giraph Key: GIRAPH-11 URL: https://issues.apache.org/jira/browse/GIRAPH-11 Project: Giraph Issue Type: Improvement Affects Versions: 0.70.0 Reporter: Avery Ching Assignee: Avery Ching Attachments: GIRAPH-11.diff Currently, Giraph assumes that the data from the VertexInputFormat is sorted. If the user data is not sorted by the vertex id, they must first run a MapReduce or Pig job to generate a sorted dataset. This is often a bit inconvenient. Giraph graph partitioning is currently range based and there are some advantages and disadvantages of this approach. The proposal of this JIRA would be to allow for both range and hash based partitioning and provide more flexibility to the user. Design goals for the graph distribution: * Allow vertices to be unordered or unordered * Ability to repartition * Select the partitioning scheme based on user needs (i.e. hash or range based) * Ability to provide user-specific hints about partitions Hash-based partitioning * Good vertex balancing across ranges for random data * Bad at vertex id locality Range-based partitioning * Good at vertex id locality * Ability to split ranges easily * Can cause hotspots for hot ranges -- This message is automatically generated by JIRA. If you think it was sent incorrectly, please contact your JIRA administrators: https://issues.apache.org/jira/secure/ContactAdministrators!default.jspa For more information on JIRA, see: http://www.atlassian.com/software/jira
[jira] [Commented] (GIRAPH-64) Create VertexRunner to make it easier to run users' computations
[
https://issues.apache.org/jira/browse/GIRAPH-64?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanelfocusedCommentId=13146023#comment-13146023
]
Hyunsik Choi commented on GIRAPH-64:
In my case, 'mvn package' is ok, but 'mvn assembly:assembly' incurs the error
as I mentioned above.
{code}
hyunsik@code:~$ mvn --version
Apache Maven 3.0.3 (r1075438; 2011-03-01 02:31:09+0900)
Maven home: /home/hyunsik/Local/maven-3
Java version: 1.6.0_26, vendor: Sun Microsystems Inc.
Java home: /usr/lib/jvm/java-6-sun-1.6.0.26/jre
Default locale: en_US, platform encoding: UTF-8
OS name: linux, version: 3.0.0-12-generic, arch: amd64, family: unix
{code}
Create VertexRunner to make it easier to run users' computations
Key: GIRAPH-64
URL: https://issues.apache.org/jira/browse/GIRAPH-64
Project: Giraph
Issue Type: New Feature
Reporter: Jakob Homan
Assignee: Jakob Homan
Attachments: GIRAPH-64.patch
Currently, if a user wants to implement a Giraph algorithm by extending
{{Vertex}} they must also write all the boilerplate around the {{Tool}}
interface and bundle it with the Giraph jar (or get Giraph on the classpath
and playing nice with the implementation). For example, what is included in
the PageRankBenchmark and what Kohei has done:
https://github.com/smly/java-Giraph-LabelPropagation It would be better if
we had perhaps a Vertex implementation to be subclassed that already had all
the standard Tooling included such that all one had to run would be (assuming
the Giraph jar was already on the classpath):
{noformat}hadoop jar my-awesome-vertex.jar my.awesome.vertex -i jazz_input -o
jazz_output -if org.apache.giraph.lib.in.text.adjacency-list.LongDoubleDouble
-of org.apache.giraph.lib.out.text.adjacency-list.LongDoubleDouble{noformat}
This wouldn't work with every algorithm, but would be useful in a large
number of cases.
--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators:
https://issues.apache.org/jira/secure/ContactAdministrators!default.jspa
For more information on JIRA, see: http://www.atlassian.com/software/jira
[jira] [Commented] (GIRAPH-36) Ensure that subclassing BasicVertex is possible by user apps
[ https://issues.apache.org/jira/browse/GIRAPH-36?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanelfocusedCommentId=13139976#comment-13139976 ] Hyunsik Choi commented on GIRAPH-36: Looks great! Actually, I need more time to fully keep up with this patch. First of all, I have executed unit tests on real hadoop cluster running on local host. All tests are passed! Ensure that subclassing BasicVertex is possible by user apps Key: GIRAPH-36 URL: https://issues.apache.org/jira/browse/GIRAPH-36 Project: Giraph Issue Type: Improvement Components: graph Affects Versions: 0.70.0 Reporter: Jake Mannix Assignee: Jake Mannix Priority: Blocker Fix For: 0.70.0 Attachments: GIRAPH-36.diff Original assumptions in Giraph were that all users would subclass Vertex (which extended MutableVertex extended BasicVertex). Classes which wish to have application specific data structures (ie. not a TreeMapI, EdgeI,E) may need to extend either MutableVertex or BasicVertex. Unfortunately VertexRange extends ArrayListVertex, and there are other places where the assumption is that vertex classes are either Vertex, or at least MutableVertex. Let's make sure the internal APIs allow for BasicVertex to be the base class. -- This message is automatically generated by JIRA. If you think it was sent incorrectly, please contact your JIRA administrators: https://issues.apache.org/jira/secure/ContactAdministrators!default.jspa For more information on JIRA, see: http://www.atlassian.com/software/jira
[jira] [Commented] (GIRAPH-13) Port Giraph to YARN
[ https://issues.apache.org/jira/browse/GIRAPH-13?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanelfocusedCommentId=13136877#comment-13136877 ] Hyunsik Choi commented on GIRAPH-13: Probably, there are many prerequisite and difficult issues. I'm willing to wait for your update :) Port Giraph to YARN --- Key: GIRAPH-13 URL: https://issues.apache.org/jira/browse/GIRAPH-13 Project: Giraph Issue Type: New Feature Reporter: Jakob Homan Assignee: Jakob Homan Now that YARN (aka MR2 aka MAPREDUCE-279) has been merged into the Hadoop trunk, we should think about what it would take to separate out the graph processing bits of Giraph from the MR1-specific code so as to take advantage of the less-MR centric aspects of YARN, while still supporting both over the medium term. -- This message is automatically generated by JIRA. If you think it was sent incorrectly, please contact your JIRA administrators: https://issues.apache.org/jira/secure/ContactAdministrators!default.jspa For more information on JIRA, see: http://www.atlassian.com/software/jira
[jira] [Commented] (GIRAPH-12) Investigate communication improvements
[ https://issues.apache.org/jira/browse/GIRAPH-12?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanelfocusedCommentId=13120604#comment-13120604 ] Hyunsik Choi commented on GIRAPH-12: Thank you for review. I agree with your opinion. The virtual memory size seems very important in 32-bit JVMs. I only considered 64-bit JVMs. I overlooked that point. Anyway, this patch allows users to control the number of threads. It is more helpful in restricted environment (e.g., 32-bit JVM). Investigate communication improvements -- Key: GIRAPH-12 URL: https://issues.apache.org/jira/browse/GIRAPH-12 Project: Giraph Issue Type: Improvement Components: bsp Reporter: Avery Ching Assignee: Hyunsik Choi Priority: Minor Attachments: GIRAPH-12_1.patch, GIRAPH-12_2.patch, GIRAPH-12_3.patch Currently every worker will start up a thread to communicate with every other workers. Hadoop RPC is used for communication. For instance if there are 400 workers, each worker will create 400 threads. This ends up using a lot of memory, even with the option -Dmapred.child.java.opts=-Xss64k. It would be good to investigate using frameworks like Netty or custom roll our own to improve this situation. By moving away from Hadoop RPC, we would also make compatibility of different Hadoop versions easier. -- This message is automatically generated by JIRA. If you think it was sent incorrectly, please contact your JIRA administrators: https://issues.apache.org/jira/secure/ContactAdministrators!default.jspa For more information on JIRA, see: http://www.atlassian.com/software/jira
[jira] [Commented] (GIRAPH-12) Investigate communication improvements
[ https://issues.apache.org/jira/browse/GIRAPH-12?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanelfocusedCommentId=13116960#comment-13116960 ] Hyunsik Choi commented on GIRAPH-12: Dmitriy, Thank you for your comments. Regardless of the problem caused by thread stack size, those approaches look promising. Especially, spilling messages to disk looks necessary so that Giraph deals with really large graph data. Otherwise, out of memory may occur when the message generating rate are higher than network bandwidth. I'll open a separate issue about this. Investigate communication improvements -- Key: GIRAPH-12 URL: https://issues.apache.org/jira/browse/GIRAPH-12 Project: Giraph Issue Type: Improvement Components: bsp Reporter: Avery Ching Assignee: Hyunsik Choi Priority: Minor Attachments: GIRAPH-12_1.patch, GIRAPH-12_2.patch Currently every worker will start up a thread to communicate with every other workers. Hadoop RPC is used for communication. For instance if there are 400 workers, each worker will create 400 threads. This ends up using a lot of memory, even with the option -Dmapred.child.java.opts=-Xss64k. It would be good to investigate using frameworks like Netty or custom roll our own to improve this situation. By moving away from Hadoop RPC, we would also make compatibility of different Hadoop versions easier. -- This message is automatically generated by JIRA. If you think it was sent incorrectly, please contact your JIRA administrators: https://issues.apache.org/jira/secure/ContactAdministrators!default.jspa For more information on JIRA, see: http://www.atlassian.com/software/jira
[jira] [Commented] (GIRAPH-12) Investigate communication improvements
[
https://issues.apache.org/jira/browse/GIRAPH-12?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanelfocusedCommentId=13116961#comment-13116961
]
Hyunsik Choi commented on GIRAPH-12:
Avery,
Thank you for your review. You are right. Runtime's totalMem() and freeMem()
methods doesn't measure stack sizes. I'm sure of it after testing the below
code.
https://gist.github.com/1249761
I have looked for how to measure the stack size of a java application. I could
not find about that. Still, I'm not sure how to show that thread stack memory
is reduced by the thread pool approach. Now, your way seems a only method to
prove them.
However, I'm curious to know how much thread overhead is in terms of memory
consumption. Before I try your approach. I conducted some simple experiments.
I used the above source code to investigate the memory usage of threads. This
is executed on a machine with intel i3, ubuntu 11.10 (64bit), and 8G memory. I
measure their memory by using 'top'. 'top' shows several columns including VIRT
and RES, and SHR. We only need to focus RES, resident memory. RES includes all
resident memory usages, such as heap and stack. I could know this from this
page (http://goo.gl/JE7fD).
Firstly, I executed the above code with 1000 threads and without a jvm option
'-Xss'. Accoring to this page (http://goo.gl/sz2qM), the default stack size
'Xss' is 1024k on the jvm of 64bit linux. After all threads are created, I
executed 'top' to print the memory usages as follows:
1k threads with default thread stack size.
{noformat}
VIRT RES SHR
9163 hyunsik 20 0 3366m 30m 8296 S 18 0.4 0:01.52 java
{noformat}
2k threads with default thread stack size.
{noformat}
VIRT RES SHR
11223 hyunsik 20 0 4434m 46m 8340 S 40 0.6 0:04.11 java
{noformat}
With 1k and 2k threads, that program consumes only 30 and 46 mega bytes
respectively. The memory usage of threads are smaller than I expected. I wonder
if thread stack size is the main cause of the memory problem that we have faced.
Besides, the default stack size is 1024k. The thread stack size seems to not
affect RES. I had more tests with 'Xss' in order to investigate more the thread
stack size.
1k threads with '-Xss4096k'.
{noformat}
28301 hyunsik 20 0 6380m 30m 8292 S 17 0.4 0:05.25 java
{noformat}
2k threads with '-Xss4096k'
{noformat}
29326 hyunsik 20 0 10.1g 46m 8300 S 38 0.6 0:03.42 java
{noformat}
VIRT surely is affected by '-Xss', but RES is not. 'Xss' seems the maximum
stack size of each thread because it doesn't affect RES.
What do you think about that?
Investigate communication improvements
--
Key: GIRAPH-12
URL: https://issues.apache.org/jira/browse/GIRAPH-12
Project: Giraph
Issue Type: Improvement
Components: bsp
Reporter: Avery Ching
Assignee: Hyunsik Choi
Priority: Minor
Attachments: GIRAPH-12_1.patch, GIRAPH-12_2.patch
Currently every worker will start up a thread to communicate with every other
workers. Hadoop RPC is used for communication. For instance if there are
400 workers, each worker will create 400 threads. This ends up using a lot
of memory, even with the option
-Dmapred.child.java.opts=-Xss64k.
It would be good to investigate using frameworks like Netty or custom roll
our own to improve this situation. By moving away from Hadoop RPC, we would
also make compatibility of different Hadoop versions easier.
--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators:
https://issues.apache.org/jira/secure/ContactAdministrators!default.jspa
For more information on JIRA, see: http://www.atlassian.com/software/jira
[jira] [Commented] (GIRAPH-12) Investigate communication improvements
[
https://issues.apache.org/jira/browse/GIRAPH-12?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanelfocusedCommentId=13116166#comment-13116166
]
Hyunsik Choi commented on GIRAPH-12:
Avery,
Thank you for your comments. I decided to use Runtime. It seems to be enough to
investigate this issue.
Again, I conducted a benchmark to measure memory consumption with
RandomMessageBenchmark as follows:
{noformat}
hadoop jar giraph-0.70-jar-with-dependencies.jar
org.apache.giraph.benchmark.RandomMessageBenchmark -e 2 -s 3 -w 20 -b 4 -n 150
-V ${V} -v -f ${f}
{noformat}
, where 'f' option indicates the number of threads of thread pool. And, I
changed the the thread executor as FixedThreadPool class.
I conducted two times for every experiment and I got the average of them. You
can see the results from the below link:
http://goo.gl/arP62
This experiments was conducted in two cluaster nodes, each of which has 24
cores and 64GB mem. They are connected each other over 1Gbps ethernet. I
measured the memory footprints from Runtime in GraphMapper as Avery recommended.
In sum, the thread pool approach is better than original approach in terms of
processing times. I guess that this is because the thread pool approach reduces
the context switching cost and narrow the synchronization area.
Unfortunately, however, the thread pool approach doesn't reduce the memory
consumption. This is the main focus of this issue. Rather, this approach needs
slightly more memory as shown in Figure 3 and 4. However, we need to note the
experiments with f = 5 and f = 20. In these experiments, the number of threads
has small effect on the memory consumption.
We have faced the memory problem. We may need to approach this problem from
another aspect.
I think that this problem may be mainly caused by the current message flushing
strategy.
In current implementation, outgoing messages are transmitted to other peers by
only two cases:
1) When the number of outgoing messages for a specific peer exceeds the a
threshold (i.e., maxSize), the outgoing messages for the peer are transmitted
to the peer.
2) When one super step is finished, the entire messages are flushed to other
peers.
Flush (case 2) is only triggered at the end of superstep. During processing,
the message flushing only depends on the case 1. This may be not effective
because the case 1 only consider the the number of messages for each specific
peer. It never take account of the real memory occupation. If destinations of
outgoing messages are uniform, out of memory may occur before any 'case 1' is
triggered.
To overcome this problem, we may need more eager message flushing strategy or
some approach to store overflow messages into disk.
Let me know what you think.
Investigate communication improvements
--
Key: GIRAPH-12
URL: https://issues.apache.org/jira/browse/GIRAPH-12
Project: Giraph
Issue Type: Improvement
Components: bsp
Reporter: Avery Ching
Assignee: Hyunsik Choi
Priority: Minor
Attachments: GIRAPH-12_1.patch, GIRAPH-12_2.patch
Currently every worker will start up a thread to communicate with every other
workers. Hadoop RPC is used for communication. For instance if there are
400 workers, each worker will create 400 threads. This ends up using a lot
of memory, even with the option
-Dmapred.child.java.opts=-Xss64k.
It would be good to investigate using frameworks like Netty or custom roll
our own to improve this situation. By moving away from Hadoop RPC, we would
also make compatibility of different Hadoop versions easier.
--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators:
https://issues.apache.org/jira/secure/ContactAdministrators!default.jspa
For more information on JIRA, see: http://www.atlassian.com/software/jira
