giraph-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Puneet Agarwal <puagar...@yahoo.com>
Subject Re: Fw: Persistent Aggregator Forgets Previously Aggregated Value ????
Date Mon, 17 Nov 2014 10:05:51 GMT
Dear Matthew,
Many thanks. It seem to have worked. I will post detailed analysis later. Hurry ......Thanks
a ton. 
( again....  phew... I was stuck on this for long time)
I will try and repay this be explaining how Aggregators work in Giraph, so that nobody else
faces this issue in future.

Following is what I understand now.1. There is an aggregator at every worker - Let me call
this as local-aggregator
2. There is another aggregator that aggregates the values for every superstep - let me call
this ss-aggregator3. There is a global aggregator which remembers the aggregated value from
previous superstep. Let me call this master-aggregator.4. There is sometimes one extra and
sometimes more than one extra aggregator at every worker, not sure that's their role so let
me call them as extra-aggregators.
How I think it is working is:
a. The master-aggregator receives the value aggregated in previous superstep as new value
of the method aggregate(). In this method, we can access the  value sent by ss-aggregator
by calling getAggregatedValue(). Thus by merging these two values we get finally aggregated
value. 

b. local-aggregators send their aggregated values to ss-aggregators. It is hard to figure
out, which one out of various aggregators is the ss-aggregator.

c. Not sure what is the role of extra-aggregators.
d. It seems that the ss-aggregators get created in a superstep and get destroyed at the end.
They don't live beyond one superstep, but their value get passed to suitable aggregator, to
complete the functionality.
Is my understanding right?
Also -->

Can I print the currentSuperstep in aggregate method? ( currentSuperstep= the superstep after
which the messages are being aggregated)
Is there any way to access getConf in aggregate method, if yes how?
- Puneet
 

     On Monday, November 17, 2014 2:36 PM, Panagiotis Eustratiadis <ep.pan.dit@gmail.com>
wrote:
   

 Hello Puneet,
I am encountering a very similar problem, and I'm studying it. I will update this thread as
soon as I find something. I think it has something to do with the code that initializes the
aggregators, but I'm not quite sure.
Regards,Eustratiadis Panagiotis.
On 17 November 2014 05:25, Puneet Agarwal <puagarwal@yahoo.com> wrote:

I have no clue on how to resolve this issue. Any guidance in this situation can be helpful.
Claudio - is there any detailed description of how aggregator works in Giraph in your book?
Getting no response from anyone, should such questions be asked in developers' list?

-PuneetIIT Delhi, India


      On Sunday, November 16, 2014 12:11 PM, Puneet Agarwal <puagarwal@yahoo.com> wrote:
   

 Hi All,

I have been facing a problem in Giraph for last two weeks. Following is the detailed description,
sample code, corresponding log message and a few questions.
I will greatly appreciate any help in this regard.

Code Description:
=================
My algorithm assigns a score to select vertexes of the graph. I have to find the best K(=5)
vertexes such that their score is minimum.
I have written a Persistent Aggregator for this which takes Text value. The snippet of the
code is given below.

It works fine on my laptop (single worker), but not on the cluster of 4 computers, when running
with 10 workers.
I am using Hadoop 1.0.0 and Giraph 1.0.0.


Problem Description:
====================
In the aggregator one of the previously aggregated value suddenly gets lost.
Therefore, my algorithm generates wrong results.

You can observe this in the logs, given below, in following manner
a.    At line number 11, the vertex 8.4793-1 has weight 3.0, and it is the best so far.
b.    At line number 12, the aggregator receives a new value "6.1100-1,6.0", i.e., vertex=6.1100-1
has weight=6.0
c.    At line number 13, in the begining of aggregate method, the output of getAggregatedValue()
is also same as newValue.
This means it simply forgets the previously aggregated value, why so? 

I have printed object-ids also, in the log messages, in order to ensure that this is happening
in the same object.

Source Code Snippet:
====================

@Override
public void aggregate(Text value) {
    printLogMessage("aggregate", "MyAGG", "Entering - NewValue=", value);
    printLogMessage("aggregate", "MyAGG",
            "Entering - getAggregatedValue()=", getAggregatedValue());

        // Main Logic is here - Omitting for simplicity

    setAggregatedValue(new Text(myTopK));
    printLogMessage("aggregate", "MyAGG", "Exiting AggregatedValue=",
            getAggregatedValue().toString());
}

@Override
public Text createInitialValue() {
    printLogMessage("=====>createInitialValue", "MyAGG",
            "Entering getAggregatedValue()=", getAggregatedValue());
    Text initialValue = getAggregatedValue();
    if (initialValue == null) {
        initialValue = new Text();
    }
    printLogMessage("=====>createInitialValue", "MyAGG",
            "Exiting ReturnedValue=", getAggregatedValue());
    return initialValue;
}
private void printLogMessage(String methodName, String prefix, String msg, Object obj) {
    if (isDebugOn(methodName)) {
        if (obj == null) {
            obj = "";
        }
        System.out.println(java.lang.System.identityHashCode(this) + "--"+ prefix +
"==>" + 
                                methodName + "() - " + msg
+ obj.toString());
    }
}

Printed Log Messages:
=====================
1.  %%%%%%%%%%%%% Entered Master Compute %%%%%%%%%%%%%%%%%%%% ====> SupsetStep No.=2
2.  249768912--MyAGG==>=====>createInitialValue() - Entering getAggregatedValue()=
3.  249768912--MyAGG==>=====>createInitialValue() - Exiting ReturnedValue=
4.  249768912--MyAGG==>aggregate() - Entering - NewValue=8.4793-1,3.0
5.  249768912--MyAGG==>aggregate() - Entering - getAggregatedValue()=8.4793-1,3.0
6.  249768912--MyAGG==>aggregate() - Received AggregatedValue Again- Exiting CurrVal=8.4793-1,3.0
7.  249768912--MyAGG==>=====>createInitialValue() - Entering getAggregatedValue()=8.4793-1,3.0
8.  249768912--MyAGG==>=====>createInitialValue() - Exiting ReturnedValue=8.4793-1,3.0
9.  %%%%%%%%%%%%% Entered Master Compute %%%%%%%%%%%%%%%%%%%% ====> SupsetStep No.=3
10. 249768912--MyAGG==>=====>createInitialValue() - Entering getAggregatedValue()=8.4793-1,3.0
11. 249768912--MyAGG==>=====>createInitialValue() - Exiting ReturnedValue=8.4793-1,3.0
12. 249768912--MyAGG==>aggregate() - Entering - NewValue=6.1100-1,6.0
13. 249768912--MyAGG==>aggregate() - Entering - getAggregatedValue()=6.1100-1,6.0
14. 249768912--MyAGG==>aggregate() - Received AggregatedValue Again- Exiting CurrVal=6.1100-1,6.0
15. 249768912--MyAGG==>=====>createInitialValue() - Entering getAggregatedValue()=6.1100-1,6.0
16. 249768912--MyAGG==>=====>createInitialValue() - Exiting ReturnedValue=6.1100-1,6.0
17. %%%%%%%%%%%%% Entered Master Compute %%%%%%%%%%%%%%%%%%%% ====> SupsetStep No.=4
18. ...
...

My Queries:
===========
1.    Why does the createInitialValue method gets called twice, in the same superstep,
on the same object of the aggregator?
2.    Even if its gets called twice, why should it forget the previouly aggregated value?

Please help me resolve this issue. This is a major problem in my work. 
I am not able to proceed further because of this.

Thanks in anticipation.

- Puneet
IIT Delhi, India



    



   
Mime
View raw message