hama-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Edward J. Yoon" <edwardy...@apache.org>
Subject Re: Question
Date Fri, 12 Sep 2014 04:15:44 GMT
Yes, partitioning job partitions the input, assigns the partition file
to proper task of graph job. GraphJobRunner.LoadVetices() method
extracts the Vertex structure from records, and loads vertices into
{Mem/Disk/Heap}VerticesInfo.

On Fri, Sep 12, 2014 at 12:38 AM, Peter Zhang <zhangpeiran1227@gmail.com> wrote:
> Hi,
> Actually, e ....I want to know more about the details of the graph
> model,especially the class GraphJobRunner. Unfortunately I have not read
> all the source codes.I want to know weather the input record of the method
> GraphJobRunner.LoadVetices() is the record in the temp files that the
> partitioning job product after the graph is partitioned???
> thanks !
>
> 2014-09-10 7:38 GMT+08:00 Edward J. Yoon <edwardyoon@apache.org>:
>
>> Hi,
>>
>> Internally, client library submits two jobs to Hama cluster: 1)
>> partitioning BSP job 2) graph job.
>>
>> On Wed, Sep 10, 2014 at 1:31 AM, Peter Zhang <zhangpeiran1227@gmail.com>
>> wrote:
>> > Hi:
>> >   Recently, I am confused with how does the hama graph model work.when I
>> > submit a hama graph job, is there a partition job starting first? And How
>> > dose the class GraphJobRunner is created when the actuall graph job
>> begins??
>> >
>> > 2014-08-21 18:44 GMT+08:00 Edward J. Yoon <edwardyoon@apache.org>:
>> >
>> >> Oh.. it seems you misunderstood. The meaning of superstep is equal to
>> >> the meaning of iteration.
>> >>
>> >> I was mean that the count of ITERATIONS is not equal to the count of
>> >> SUPERSTEPS. The additional supersteps is used for vertices' value
>> >> initialization and master aggregation.
>> >>
>> >> ..
>> >> 14/08/21 19:34:24 INFO bsp.BSPJobClient: Counters: 12
>> >> 14/08/21 19:34:24 INFO bsp.BSPJobClient:   GraphJobCounter
>> >> 14/08/21 19:34:24 INFO bsp.BSPJobClient:     ITERATIONS=30
>> >> 14/08/21 19:34:24 INFO bsp.BSPJobClient:     INPUT_VERTICES=400
>> >> 14/08/21 19:34:24 INFO bsp.BSPJobClient:   JobCounter
>> >> 14/08/21 19:34:24 INFO bsp.BSPJobClient:     SUPERSTEPS=60
>> >> 14/08/21 19:34:24 INFO bsp.BSPJobClient:     LAUNCHED_TASKS=3
>> >> ..
>> >>
>> >> On Wed, Aug 20, 2014 at 11:40 PM, 张沛然 <zhangpeiran1227@gmail.com>
>> wrote:
>> >> > Hello:
>> >> >  I  get confused about hama graph ,take the SSSP as example:
>> >> > input :4 vertexs and4 edges
>> >> >          1\t2:2\t3:3
>> >> >          2\t4:4
>> >> >          3\t4:1
>> >> >          4
>> >> >
>> >> > I guess it will process like below:
>> >> >   superstep 0:vertex1,2,3,4 do computer()
>> >> >                     that's :vertex1 sent message of value 2 to vertex2
>> >> >                                vertex1 sent message of value 3 to
>> vertes3
>> >> >                                 vertex1,2,3,4 voteToHalt() and get
>> >> inactive;
>> >> >  superstep 1:only vertex 2,3 o compute()
>> >> >                    that's  vertex 2,3 received message and update
>> value
>> >> >                              vertex 2 sent message of value 6 to
>> vertex4
>> >> >                              vertex 3 sent message of value 4 to
>> vertex4
>> >> >                             vertex 2,3 voteToHalt()
>> >> >  superstep 2 :only vertex 4 do computer()
>> >> >                     that's  vertex4 received message and update value
>> >> >  then ,all the vertexs is inactive ,the processing is over!  I run
it
>> in
>> >> > pseudo-distributed modes .
>> >> > 1)Is it right???
>> >> > 2)I think all the processing just need 1 iteration,but when I set the
>> >> > maxIteration to 1,the result is not right!! It show that the vertex4
>> has
>> >> > not update its value, I wonder know why their is no supertetp2 and
>> when a
>> >> > iteration is over???
>> >> > Thanks a lot.I just get mad......
>> >>
>> >>
>> >>
>> >> --
>> >> Best Regards, Edward J. Yoon
>> >> CEO at DataSayer Co., Ltd.
>> >>
>>
>>
>>
>> --
>> Best Regards, Edward J. Yoon
>> CEO at DataSayer Co., Ltd.
>>



-- 
Best Regards, Edward J. Yoon
CEO at DataSayer Co., Ltd.

Mime
View raw message