You are viewing a plain text version of this content. The canonical link for it is here.
Posted to user@giraph.apache.org by Sonja Koenig <so...@uni-ulm.de> on 2015/08/20 12:09:30 UTC
Giraph Performance Tuning
Hey there everyone!
I am currently writing my bachelor thesis about Giraph and GraphX, where
I am trying to compare their scalability and features and bring them
into a context with different graph types.
In order to compare the two on a fair basis, I want to tune the
frameworks to get the most out of them :-)
I was hoping to get some tips and tricks from you all, where I can make
some configurations to impact my computations..
My set up:
10 machines, each 1 cpu with 1 3,3GHz core, 4GB RAM, 100GB HDD -> one is
designated master
Giraph 1.10
Hadoop 1.2.1
So far I haven't done any special configurations for hadoop or giraph
besides the basic ones during setup.
Performance-critical might be these:
In *mapred-site.xml*:
mapred.tasktracker.map.tasks.maximum = 4
mapred.map.tasks=4
In *dfs-site.xml*:
dfs.replication=3
If I am correctly informed, the default amount of heap is 1000MB, which
I haven't changed. I am also not sure where I can actually increase
memory usage. Any advice?
Also, I read somewhere that it is smarter to increase the amount of
threads per worker and not the amount of worker per machine? But I am
anyways somewhat handicapped with only one core per machine..
Lastly, has anyone notice any performance changes when using
checkointing, cominers, aggregators and so on?
Is the use of combiners and aggregators a choice of the application code
or my execution command?
I would appreciate any advice and comments greatly! :-)
Greetings from Ulm,
Sonja
Re: Giraph Performance Tuning
Posted by Sonja Koenig <so...@uni-ulm.de>.
Thank you Eric!
Your information gave me a good inpoint for further configuration :)
It helped immensely!
Also I agree with you. A small amount of powerful workers is quite
surely better than a large amount of week. Unfortunately I'm just doing
my bachelor's, so I'm not (yet) in a position to scream for "more power
pls!!!" ;D
But it's working out nicely, already got some good data.
Greetings!
Sonja
Am 24.08.2015 um 18:50 schrieb Eric Kimbrel:
> Hello,
>
> I am not at all an expert, but here¹s what I can tell you:
>
> I havent used giraph in PURE_YARN MODE, as I¹ve had problems getting it to
> compile correctly, but when
> Using hadoop_2 mapreduce you set memory and cores per worker with the
> hadoop map reduce options. Those
> Might be different from mine depending on your hadoop version but I use:
> (hadoop 2.6.0)
>
> -Dmapreduce.map.memory.mb=$WORKER_MEMORY
> -Dmapreduce.map.cpu.vcores=$WORKER_CORES
>
>
> As a General rule of thumb I have found that big workers with lots of
> memory and cores will outperform lots of small workers by having reduced
> network IO, but that isn¹t exactly always true. Finding the optimal
> configuration for a give algorithm / hardware can be somewhat of an art
> form. I would generally find 10 workers with 10 cores and a ample memory
> would outperform 100 workers with a single core, but again that¹s just my
> rule of thumb and not always the case. You are probably going to want to
> run 9 workers (saving one machine for the master node) with one core and
> 3GB each.
>
> One of the reasons this is such a hard task is because you have three big
> variables to consider (algorithms, datasets, hardware,
> platforms/frameworks). Giraph and GraphX have very different execution
> models, and because of that the implementation of the same algorithm on
> each may be very different, and perform very differently. Implementation
> details can make all the difference.
>
> Combiners/Aggregators: Combiners and aggregators are completely
> application specific. (Although to use a combiner you do specify it on the
> command line, but some applications will work with them while others
> won¹t).
>
>
> Good luck on your thesis work!
>
> -Eric
>
>
> On 8/24/15, 2:18 AM, "Sonja Koenig" <so...@uni-ulm.de> wrote:
>
>> Hey there everyone!
>>
>> On the user list, there was noone to help me, so I thought I'll just
>> start bugging devs..
>>
>> I am currently writing my bachelor thesis about Giraph and GraphX, where
>> I am trying to compare their scalability and features and bring them
>> into a context with different graph types.
>> In order to compare the two on a fair basis, I want to tune the
>> frameworks to get the most out of them :-)
>> I was hoping to get some tips and tricks from you all, where I can make
>> some configurations to impact my computations..
>>
>> My set up:
>> 10 machines, each 1 cpu with 1 3,3GHz core, 4GB RAM, 100GB HDD -> one is
>> designated master
>> Giraph 1.10
>> Hadoop 1.2.1
>>
>> So far I haven't done any special configurations for hadoop or giraph
>> besides the basic ones during setup.
>> Performance-critical might be these:
>> In *mapred-site.xml*:
>> mapred.tasktracker.map.tasks.maximum = 4
>> mapred.map.tasks=4
>> In *dfs-site.xml*:
>> dfs.replication=3
>>
>> If I am correctly informed, the default amount of heap is 1000MB, which
>> I haven't changed. I am also not sure where I can actually increase
>> memory usage. Any advice?
>> Also, I read somewhere that it is smarter to increase the amount of
>> threads per worker and not the amount of worker per machine? But I am
>> anyways somewhat handicapped with only one core per machine..
>>
>> Lastly, has anyone noticed any performance changes when using
>> checkointing, combiners, aggregators and so on?
>> Is the use of combiners and aggregators a choice of the application code
>> or my execution command?
>>
>> I would appreciate any advice and comments greatly! :-)
>>
>> Greetings from Ulm,
>> Sonja
>>
>>
>>
Re: Giraph Performance Tuning
Posted by Eric Kimbrel <Er...@soteradefense.com>.
Hello,
I am not at all an expert, but here¹s what I can tell you:
I havent used giraph in PURE_YARN MODE, as I¹ve had problems getting it to
compile correctly, but when
Using hadoop_2 mapreduce you set memory and cores per worker with the
hadoop map reduce options. Those
Might be different from mine depending on your hadoop version but I use:
(hadoop 2.6.0)
-Dmapreduce.map.memory.mb=$WORKER_MEMORY
-Dmapreduce.map.cpu.vcores=$WORKER_CORES
As a General rule of thumb I have found that big workers with lots of
memory and cores will outperform lots of small workers by having reduced
network IO, but that isn¹t exactly always true. Finding the optimal
configuration for a give algorithm / hardware can be somewhat of an art
form. I would generally find 10 workers with 10 cores and a ample memory
would outperform 100 workers with a single core, but again that¹s just my
rule of thumb and not always the case. You are probably going to want to
run 9 workers (saving one machine for the master node) with one core and
3GB each.
One of the reasons this is such a hard task is because you have three big
variables to consider (algorithms, datasets, hardware,
platforms/frameworks). Giraph and GraphX have very different execution
models, and because of that the implementation of the same algorithm on
each may be very different, and perform very differently. Implementation
details can make all the difference.
Combiners/Aggregators: Combiners and aggregators are completely
application specific. (Although to use a combiner you do specify it on the
command line, but some applications will work with them while others
won¹t).
Good luck on your thesis work!
-Eric
On 8/24/15, 2:18 AM, "Sonja Koenig" <so...@uni-ulm.de> wrote:
>Hey there everyone!
>
>On the user list, there was noone to help me, so I thought I'll just
>start bugging devs..
>
>I am currently writing my bachelor thesis about Giraph and GraphX, where
>I am trying to compare their scalability and features and bring them
>into a context with different graph types.
>In order to compare the two on a fair basis, I want to tune the
>frameworks to get the most out of them :-)
>I was hoping to get some tips and tricks from you all, where I can make
>some configurations to impact my computations..
>
>My set up:
>10 machines, each 1 cpu with 1 3,3GHz core, 4GB RAM, 100GB HDD -> one is
>designated master
>Giraph 1.10
>Hadoop 1.2.1
>
>So far I haven't done any special configurations for hadoop or giraph
>besides the basic ones during setup.
>Performance-critical might be these:
>In *mapred-site.xml*:
> mapred.tasktracker.map.tasks.maximum = 4
> mapred.map.tasks=4
>In *dfs-site.xml*:
> dfs.replication=3
>
>If I am correctly informed, the default amount of heap is 1000MB, which
>I haven't changed. I am also not sure where I can actually increase
>memory usage. Any advice?
>Also, I read somewhere that it is smarter to increase the amount of
>threads per worker and not the amount of worker per machine? But I am
>anyways somewhat handicapped with only one core per machine..
>
>Lastly, has anyone noticed any performance changes when using
>checkointing, combiners, aggregators and so on?
>Is the use of combiners and aggregators a choice of the application code
>or my execution command?
>
>I would appreciate any advice and comments greatly! :-)
>
>Greetings from Ulm,
>Sonja
>
>
>
Fwd: Giraph Performance Tuning
Posted by Sonja Koenig <so...@uni-ulm.de>.
Hey there everyone!
On the user list, there was noone to help me, so I thought I'll just
start bugging devs..
I am currently writing my bachelor thesis about Giraph and GraphX, where
I am trying to compare their scalability and features and bring them
into a context with different graph types.
In order to compare the two on a fair basis, I want to tune the
frameworks to get the most out of them :-)
I was hoping to get some tips and tricks from you all, where I can make
some configurations to impact my computations..
My set up:
10 machines, each 1 cpu with 1 3,3GHz core, 4GB RAM, 100GB HDD -> one is
designated master
Giraph 1.10
Hadoop 1.2.1
So far I haven't done any special configurations for hadoop or giraph
besides the basic ones during setup.
Performance-critical might be these:
In *mapred-site.xml*:
mapred.tasktracker.map.tasks.maximum = 4
mapred.map.tasks=4
In *dfs-site.xml*:
dfs.replication=3
If I am correctly informed, the default amount of heap is 1000MB, which
I haven't changed. I am also not sure where I can actually increase
memory usage. Any advice?
Also, I read somewhere that it is smarter to increase the amount of
threads per worker and not the amount of worker per machine? But I am
anyways somewhat handicapped with only one core per machine..
Lastly, has anyone noticed any performance changes when using
checkointing, combiners, aggregators and so on?
Is the use of combiners and aggregators a choice of the application code
or my execution command?
I would appreciate any advice and comments greatly! :-)
Greetings from Ulm,
Sonja