giraph-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Sebastian Schelter <...@apache.org>
Subject Re: GIRAPH-825 and GIRAPH-840
Date Thu, 13 Feb 2014 17:05:53 GMT
I'm using 24 workers to process a dataset of 50M vertices. Where can I 
see the number of partitions assigned to each worker?

Best,
Sebastian

On 02/12/2014 04:34 PM, Claudio Martella wrote:
> If you set giraph.maxInMemoryPartitions accordingly to a number larger to
> that max(), and you're not setting stickyPartitions, then my question is
> whether there are at least that number of partitions in the worker. Can you
> validate whether there are enough partitions in the worker. In other terms,
> assuming you're using hash partitioner, how many workers are you using?
>
>
> On Wed, Feb 12, 2014 at 2:59 PM, Sebastian Schelter <ssc@apache.org> wrote:
>
>> No. Should I have done that?
>>
>>
>> On 02/12/2014 02:57 PM, Claudio Martella wrote:
>>
>>> did you also set stickyPartitions to some numbers?
>>>
>>>
>>> On Wed, Feb 12, 2014 at 1:00 PM, Sebastian Schelter <ssc@apache.org>
>>> wrote:
>>>
>>>   Updating documentation is never a bad idea :)
>>>>
>>>> I reran my test with giraph.maxPartitionsInMemory >
>>>> max(giraph.numComputeThreads,giraph.numInputThreads,giraph.
>>>> numOutputThreads)
>>>> and still got the same behavior. I'll wait for the updated patch.
>>>>
>>>> Get well Armando!
>>>>
>>>>
>>>> On 02/12/2014 12:53 PM, Armando Miraglia wrote:
>>>>
>>>>
>>>>> btw: I as also thinking to update the documentation page on the Giraph
>>>>> website to better explain the sticky partition logic. What do  you
>>>>> think?
>>>>>
>>>>> Cheers,
>>>>> Armando
>>>>>
>>>>>
>>>>> On Wed, Feb 12, 2014 at 12:50:25PM +0100, Armando Miraglia wrote:
>>>>>
>>>>>   Indeed, yesterday I was fixing a couple of things and I think I missed
>>>>>> a
>>>>>> case that I have to exclude. Sorry for this, I have fever at the
>>>>>> momento
>>>>>> so it could be that yesterday I was under the effect of the fever
:D
>>>>>>
>>>>>> I checked that the tests were passing but I think a missed something.
>>>>>>
>>>>>> I'll come back to you very soon
>>>>>>
>>>>>> On Wed, Feb 12, 2014 at 10:26:12AM +0100, Claudio Martella wrote:
>>>>>>
>>>>>>   the problem is that you're running with more threads than in-memory
>>>>>>> partitions. increase the number of partitions in memory to be
at least
>>>>>>> the
>>>>>>> number of threads. i have no time right now to check the latest
code,
>>>>>>> but
>>>>>>> you should not set the number of stickypartitions by hand.
>>>>>>>
>>>>>>>
>>>>>>> On Wed, Feb 12, 2014 at 10:03 AM, Sebastian Schelter <ssc@apache.org>
>>>>>>> wrote:
>>>>>>>
>>>>>>>    I ran a first test with the new DiskBackedPartitionStore and
it
>>>>>>> didn't
>>>>>>>
>>>>>>>> work for me unfortunately. The job never leaves the input
phase
>>>>>>>> (superstep
>>>>>>>> -1). I sshd onto one of the workers and it seems to wait
forever on
>>>>>>>> DiskBackedPartitionStore.getOrCreatePartition:
>>>>>>>>
>>>>>>>>       java.lang.Thread.State: BLOCKED (on object monitor)
>>>>>>>>            at org.apache.giraph.partition.DiskBackedPartitionStore.
>>>>>>>> getOrCreatePartition(DiskBackedPartitionStore.java:226)
>>>>>>>>            - waiting to lock <0x00000000aeb757c8> (a
>>>>>>>> org.apache.giraph.partition.DiskBackedPartitionStore$MetaPartition)
>>>>>>>>
>>>>>>>> Here are the custom arguments for my run, let me know if
I should do
>>>>>>>> another run with a different config.
>>>>>>>>
>>>>>>>> giraph.oneToAllMsgSending=true
>>>>>>>> giraph.isStaticGraph=true
>>>>>>>> giraph.numComputeThreads=15
>>>>>>>> giraph.numInputThreads=15
>>>>>>>> giraph.numOutputThreads=15
>>>>>>>> giraph.maxNumberOfSupersteps=30
>>>>>>>> giraph.useOutOfCoreGraph=true
>>>>>>>> giraph.stickyPartitions=5
>>>>>>>>
>>>>>>>> I also ran the job without using oneToAllMsgSending and saw
the same
>>>>>>>> behavior.
>>>>>>>>
>>>>>>>> Best,
>>>>>>>> Sebastian
>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>> On 02/12/2014 12:44 AM, Claudio Martella wrote:
>>>>>>>>
>>>>>>>>    please give it a test. i've been working on this with
armando. i'll
>>>>>>>>
>>>>>>>>> give a
>>>>>>>>> review, but we have been testing it for while. we'd really
>>>>>>>>> appreciate
>>>>>>>>> if
>>>>>>>>> somebody else could run some additional tests as well.
thanks!
>>>>>>>>>
>>>>>>>>>
>>>>>>>>> On Wed, Feb 12, 2014 at 12:39 AM, Sebastian Schelter
<
>>>>>>>>> ssc@apache.org>
>>>>>>>>> wrote:
>>>>>>>>>
>>>>>>>>>     I'll test the patch from GIRAPH-825 this week.
>>>>>>>>>
>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>> On 02/12/2014 12:10 AM, Roman Shaposhnik wrote:
>>>>>>>>>>
>>>>>>>>>>     Hi!
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>>> Given how big the diff here are:
>>>>>>>>>>>          https://issues.apache.org/jira/browse/GIRAPH-825
>>>>>>>>>>>          https://issues.apache.org/jira/browse/GIRAPH-840
>>>>>>>>>>> I am wondering whether it is realistic
>>>>>>>>>>> to have them in 1.1.0.
>>>>>>>>>>>
>>>>>>>>>>> Would appreciate folks chiming in.
>>>>>>>>>>>
>>>>>>>>>>> Thanks,
>>>>>>>>>>> Roman.
>>>>>>>>>>>
>>>>>>>>>>>
>>>>>>>>>>>
>>>>>>>>>>>
>>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>
>>>>>>>>>
>>>>>>>>
>>>>>>> --
>>>>>>>       Claudio Martella
>>>>>>>
>>>>>>>
>>>>>>
>>>>
>>>
>>>
>>
>
>


Mime
View raw message