Allocation of containers to tasks in Hadoop

classic Classic list List threaded Threaded
8 messages Options
Reply | Threaded
Open this post in threaded view
|

Allocation of containers to tasks in Hadoop

Or Raz

How can I change/suggest a different allocation of containers to tasks in Hadoop? Regarding a native Hadoop (2.9.1) cluster on AWS.

I am running a native Hadoop cluster (2.9.1) on AWS (with EC2, not EMR) and I want the scheduling/allocating of the containers (Mappers/Reducers) would be more balanced than it is currently. It seems like RM is assigning the Mappers in a Bin Packing way (where the data resides) and for the reducers, it looks more balanced. My setup includes three Machines with replication rate three (all the data is on every machine), and I run my jobs with mapreduce.job.reduce.slowstart.completedmaps=0 to start shuffle as fast as possible (It is vital for me that all the containers are working in concurrency, it is a must condition). Also, according to the EC2 instances I have chosen and my settings of the YARN cluster, I can run at most 93 containers (31 each).

For example, if I want to have nine reducers then (93-9-1=83), 83 containers could be left for the mappers, and one is for the AM. I have played with the size of split input (mapreduce.input.fileinputformat.split.minsize, mapreduce.input.fileinputformat.split.maxsize) to find the right balance where all of the machines have the same "work" for the map phase. But it seems like the first 31 mappers would be allocated in one computer, the next 31 to the second one and the last 31 in the last machine. Thus, I can try to use 87 mappers where 31 of them in Machine #1, another 31 in Machine #2 and another 25 in Machine #3 and the rest is left for the reducers and as Machine #1 and Machine #2 are fully occupied then the reducers would have to be placed in Machine #3. This way I get an almost balanced allocation of mappers at the expense of unbalanced reducers allocation. And this is not what I want...

# of mappers = size_input / split size [Bytes]

split size =max(mapreduce.input.fileinputformat.split.minsize,min(mapreduce.input.fileinputformat.split.maxsize, dfs.blocksize))

Reply | Threaded
Open this post in threaded view
|

Re: Allocation of containers to tasks in Hadoop

Aaron Eng-2
Have you checked the yarn.scheduler.fair.assignmultiple and yarn.scheduler.fair.max.assign parameters for the ResourceManager configuration?

On Wed, Jan 9, 2019 at 9:49 AM Or Raz <[hidden email]> wrote:

How can I change/suggest a different allocation of containers to tasks in Hadoop? Regarding a native Hadoop (2.9.1) cluster on AWS.

I am running a native Hadoop cluster (2.9.1) on AWS (with EC2, not EMR) and I want the scheduling/allocating of the containers (Mappers/Reducers) would be more balanced than it is currently. It seems like RM is assigning the Mappers in a Bin Packing way (where the data resides) and for the reducers, it looks more balanced. My setup includes three Machines with replication rate three (all the data is on every machine), and I run my jobs with mapreduce.job.reduce.slowstart.completedmaps=0 to start shuffle as fast as possible (It is vital for me that all the containers are working in concurrency, it is a must condition). Also, according to the EC2 instances I have chosen and my settings of the YARN cluster, I can run at most 93 containers (31 each).

For example, if I want to have nine reducers then (93-9-1=83), 83 containers could be left for the mappers, and one is for the AM. I have played with the size of split input (mapreduce.input.fileinputformat.split.minsize, mapreduce.input.fileinputformat.split.maxsize) to find the right balance where all of the machines have the same "work" for the map phase. But it seems like the first 31 mappers would be allocated in one computer, the next 31 to the second one and the last 31 in the last machine. Thus, I can try to use 87 mappers where 31 of them in Machine #1, another 31 in Machine #2 and another 25 in Machine #3 and the rest is left for the reducers and as Machine #1 and Machine #2 are fully occupied then the reducers would have to be placed in Machine #3. This way I get an almost balanced allocation of mappers at the expense of unbalanced reducers allocation. And this is not what I want...

# of mappers = size_input / split size [Bytes]

split size =max(mapreduce.input.fileinputformat.split.minsize,min(mapreduce.input.fileinputformat.split.maxsize, dfs.blocksize))

Reply | Threaded
Open this post in threaded view
|

Re: Allocation of containers to tasks in Hadoop

Or Raz
As far as I know, the scheduler in YARN is only scheduling the jobs and not the containers inside each job. Therefore, I don't believe it is relevant.
Also, I haven't used or set those two parameters, and I haven't picked nor set any particular schedule for my research (Fair, FIFO or Capacity).
Please correct if I am wrong.
P.S. currently I have no interest in a situation when I run a few jobs concurrently, my case is much simpler with one job that I would like that allocation of containers will be more balanced...
Or


‫בתאריך יום ד׳, 9 בינו׳ 2019 ב-19:11 מאת ‪Aaron Eng‬‏ <‪[hidden email]‬‏>:‬
Have you checked the yarn.scheduler.fair.assignmultiple and yarn.scheduler.fair.max.assign parameters for the ResourceManager configuration?

On Wed, Jan 9, 2019 at 9:49 AM Or Raz <[hidden email]> wrote:

How can I change/suggest a different allocation of containers to tasks in Hadoop? Regarding a native Hadoop (2.9.1) cluster on AWS.

I am running a native Hadoop cluster (2.9.1) on AWS (with EC2, not EMR) and I want the scheduling/allocating of the containers (Mappers/Reducers) would be more balanced than it is currently. It seems like RM is assigning the Mappers in a Bin Packing way (where the data resides) and for the reducers, it looks more balanced. My setup includes three Machines with replication rate three (all the data is on every machine), and I run my jobs with mapreduce.job.reduce.slowstart.completedmaps=0 to start shuffle as fast as possible (It is vital for me that all the containers are working in concurrency, it is a must condition). Also, according to the EC2 instances I have chosen and my settings of the YARN cluster, I can run at most 93 containers (31 each).

For example, if I want to have nine reducers then (93-9-1=83), 83 containers could be left for the mappers, and one is for the AM. I have played with the size of split input (mapreduce.input.fileinputformat.split.minsize, mapreduce.input.fileinputformat.split.maxsize) to find the right balance where all of the machines have the same "work" for the map phase. But it seems like the first 31 mappers would be allocated in one computer, the next 31 to the second one and the last 31 in the last machine. Thus, I can try to use 87 mappers where 31 of them in Machine #1, another 31 in Machine #2 and another 25 in Machine #3 and the rest is left for the reducers and as Machine #1 and Machine #2 are fully occupied then the reducers would have to be placed in Machine #3. This way I get an almost balanced allocation of mappers at the expense of unbalanced reducers allocation. And this is not what I want...

# of mappers = size_input / split size [Bytes]

split size =max(mapreduce.input.fileinputformat.split.minsize,min(mapreduce.input.fileinputformat.split.maxsize, dfs.blocksize))

Reply | Threaded
Open this post in threaded view
|

Re: Allocation of containers to tasks in Hadoop

Aaron Eng-2
The settings are very relevant to having an equal number of containers running on each node if you have an idle cluster and want to distribute containers for a single job.  An application master submits requests for container allocations to the ResourceManager.  The MRAppMaster will request all the map containers at once, the FairScheduler will find NodeManagers with capacity to fulfill the container requests.  If assign multiple is enabled then you generally won't get an even number of containers assigned to each node +/- 1 container.  Before you say it's not relevant, you should check if your environment uses the FairScheduler and whether multiple assignment is enabled.  If so, that's likely why there isn't an even assignment +/- 1 container.  If not using FairScheduler and/or multiple assign, then you should look at locality settings, which can cause containers to be preferentially run on a subset of nodes, resulting in an uneven container assignment per node.  

On Wed, Jan 9, 2019 at 2:19 PM Or Raz <[hidden email]> wrote:
As far as I know, the scheduler in YARN is only scheduling the jobs and not the containers inside each job. Therefore, I don't believe it is relevant.
Also, I haven't used or set those two parameters, and I haven't picked nor set any particular schedule for my research (Fair, FIFO or Capacity).
Please correct if I am wrong.
P.S. currently I have no interest in a situation when I run a few jobs concurrently, my case is much simpler with one job that I would like that allocation of containers will be more balanced...
Or


‫בתאריך יום ד׳, 9 בינו׳ 2019 ב-19:11 מאת ‪Aaron Eng‬‏ <‪[hidden email]‬‏>:‬
Have you checked the yarn.scheduler.fair.assignmultiple and yarn.scheduler.fair.max.assign parameters for the ResourceManager configuration?

On Wed, Jan 9, 2019 at 9:49 AM Or Raz <[hidden email]> wrote:

How can I change/suggest a different allocation of containers to tasks in Hadoop? Regarding a native Hadoop (2.9.1) cluster on AWS.

I am running a native Hadoop cluster (2.9.1) on AWS (with EC2, not EMR) and I want the scheduling/allocating of the containers (Mappers/Reducers) would be more balanced than it is currently. It seems like RM is assigning the Mappers in a Bin Packing way (where the data resides) and for the reducers, it looks more balanced. My setup includes three Machines with replication rate three (all the data is on every machine), and I run my jobs with mapreduce.job.reduce.slowstart.completedmaps=0 to start shuffle as fast as possible (It is vital for me that all the containers are working in concurrency, it is a must condition). Also, according to the EC2 instances I have chosen and my settings of the YARN cluster, I can run at most 93 containers (31 each).

For example, if I want to have nine reducers then (93-9-1=83), 83 containers could be left for the mappers, and one is for the AM. I have played with the size of split input (mapreduce.input.fileinputformat.split.minsize, mapreduce.input.fileinputformat.split.maxsize) to find the right balance where all of the machines have the same "work" for the map phase. But it seems like the first 31 mappers would be allocated in one computer, the next 31 to the second one and the last 31 in the last machine. Thus, I can try to use 87 mappers where 31 of them in Machine #1, another 31 in Machine #2 and another 25 in Machine #3 and the rest is left for the reducers and as Machine #1 and Machine #2 are fully occupied then the reducers would have to be placed in Machine #3. This way I get an almost balanced allocation of mappers at the expense of unbalanced reducers allocation. And this is not what I want...

# of mappers = size_input / split size [Bytes]

split size =max(mapreduce.input.fileinputformat.split.minsize,min(mapreduce.input.fileinputformat.split.maxsize, dfs.blocksize))

Reply | Threaded
Open this post in threaded view
|

Re: Allocation of containers to tasks in Hadoop

Or Raz
Thanks for the tips!
Because I haven't set any scheduler (on purpose) for YARN then, I am using the default one (Capacity).
I have looked in yarn-site.xml and in the configuration tab (using JobHistory UI), and both of the parameters that you have mentioned weren't there (so they haven't been set).
You said that I should look at "locality settings" can you be more specific on what and where to look? 
Also, it is worth mentioning that I am using three computers and the replication factor (of HDFS) is three too. Thus, every data (even input) would be on every computer, and the memory of each computer is the same (two t2.xlarge and one m4.xlarge) while I am using DefaultResourceCalculator.

Or

‫בתאריך יום ד׳, 9 בינו׳ 2019 ב-23:28 מאת ‪Aaron Eng‬‏ <‪[hidden email]‬‏>:‬
The settings are very relevant to having an equal number of containers running on each node if you have an idle cluster and want to distribute containers for a single job.  An application master submits requests for container allocations to the ResourceManager.  The MRAppMaster will request all the map containers at once, the FairScheduler will find NodeManagers with capacity to fulfill the container requests.  If assign multiple is enabled then you generally won't get an even number of containers assigned to each node +/- 1 container.  Before you say it's not relevant, you should check if your environment uses the FairScheduler and whether multiple assignment is enabled.  If so, that's likely why there isn't an even assignment +/- 1 container.  If not using FairScheduler and/or multiple assign, then you should look at locality settings, which can cause containers to be preferentially run on a subset of nodes, resulting in an uneven container assignment per node.  

On Wed, Jan 9, 2019 at 2:19 PM Or Raz <[hidden email]> wrote:
As far as I know, the scheduler in YARN is only scheduling the jobs and not the containers inside each job. Therefore, I don't believe it is relevant.
Also, I haven't used or set those two parameters, and I haven't picked nor set any particular schedule for my research (Fair, FIFO or Capacity).
Please correct if I am wrong.
P.S. currently I have no interest in a situation when I run a few jobs concurrently, my case is much simpler with one job that I would like that allocation of containers will be more balanced...
Or


‫בתאריך יום ד׳, 9 בינו׳ 2019 ב-19:11 מאת ‪Aaron Eng‬‏ <‪[hidden email]‬‏>:‬
Have you checked the yarn.scheduler.fair.assignmultiple and yarn.scheduler.fair.max.assign parameters for the ResourceManager configuration?

On Wed, Jan 9, 2019 at 9:49 AM Or Raz <[hidden email]> wrote:

How can I change/suggest a different allocation of containers to tasks in Hadoop? Regarding a native Hadoop (2.9.1) cluster on AWS.

I am running a native Hadoop cluster (2.9.1) on AWS (with EC2, not EMR) and I want the scheduling/allocating of the containers (Mappers/Reducers) would be more balanced than it is currently. It seems like RM is assigning the Mappers in a Bin Packing way (where the data resides) and for the reducers, it looks more balanced. My setup includes three Machines with replication rate three (all the data is on every machine), and I run my jobs with mapreduce.job.reduce.slowstart.completedmaps=0 to start shuffle as fast as possible (It is vital for me that all the containers are working in concurrency, it is a must condition). Also, according to the EC2 instances I have chosen and my settings of the YARN cluster, I can run at most 93 containers (31 each).

For example, if I want to have nine reducers then (93-9-1=83), 83 containers could be left for the mappers, and one is for the AM. I have played with the size of split input (mapreduce.input.fileinputformat.split.minsize, mapreduce.input.fileinputformat.split.maxsize) to find the right balance where all of the machines have the same "work" for the map phase. But it seems like the first 31 mappers would be allocated in one computer, the next 31 to the second one and the last 31 in the last machine. Thus, I can try to use 87 mappers where 31 of them in Machine #1, another 31 in Machine #2 and another 25 in Machine #3 and the rest is left for the reducers and as Machine #1 and Machine #2 are fully occupied then the reducers would have to be placed in Machine #3. This way I get an almost balanced allocation of mappers at the expense of unbalanced reducers allocation. And this is not what I want...

# of mappers = size_input / split size [Bytes]

split size =max(mapreduce.input.fileinputformat.split.minsize,min(mapreduce.input.fileinputformat.split.maxsize, dfs.blocksize))

Reply | Threaded
Open this post in threaded view
|

Re: Allocation of containers to tasks in Hadoop

Or Raz
I have googled more about it, and it seems like two parameters should define the "bin packing problem".
According to https://hadoop.apache.org/docs/r2.9.1/hadoop-yarn/hadoop-yarn-site/CapacityScheduler.html#Other_Properties  yarn.scheduler.capacity.per-node-heartbeat.multiple-assignments-enabled is by default set to true and with parameter yarn.scheduler.capacity.per-node-heartbeat.maximum-container-assignments r set to -1 it can assign all the containers the Node manager "said" it is capable of (which could somehow explain the bin packing problem for the first Nodemanager who answer with a Heartbeat message).
Following Apache's instructions, I have inserted to my capacity-scheduler.xml  in hadoop/etc/hadoop folder 

  <property>
    <name>yarn.scheduler.capacity.per-node-heartbeat.multiple-assignments-enabled</name>
    <value>true</value>
    <description>
        Whether to allow multiple container assignments in one NodeManager heartbeat. Defaults to true.
    </description>
  </property>
  <property>
    <name>yarn.scheduler.capacity.per-node-heartbeat.maximum-container-assignments</name>
    <value>2</value>
    <description>
        If multiple-assignments-enabled is true, the maximum amount of containers that can be assigned in one NodeManager heartbeat. Defaults to -1, which sets no limit.
    </description>
  </property>
I have checked the configuration file, and I am using the capacity scheduler (I have enabled yarn.scheduler.capacity.per-node-heartbeat.multiple-assignments-enabled again just to be sure).
Furthermore, after I have run "yarn rmadmin -refreshQueues" I haven't seen any change in the Mappers allocation nor Reducers.
hadoop2@master:~$ yarn rmadmin -refreshQueues
19/01/10 16:06:33 INFO client.RMProxy: Connecting to ResourceManager at master/172.31.24.83:8033

What am I missing over here?

Or


‫בתאריך יום ד׳, 9 בינו׳ 2019 ב-23:57 מאת ‪Or Raz‬‏ <‪[hidden email]‬‏>:‬
Thanks for the tips!
Because I haven't set any scheduler (on purpose) for YARN then, I am using the default one (Capacity).
I have looked in yarn-site.xml and in the configuration tab (using JobHistory UI), and both of the parameters that you have mentioned weren't there (so they haven't been set).
You said that I should look at "locality settings" can you be more specific on what and where to look? 
Also, it is worth mentioning that I am using three computers and the replication factor (of HDFS) is three too. Thus, every data (even input) would be on every computer, and the memory of each computer is the same (two t2.xlarge and one m4.xlarge) while I am using DefaultResourceCalculator.

Or

‫בתאריך יום ד׳, 9 בינו׳ 2019 ב-23:28 מאת ‪Aaron Eng‬‏ <‪[hidden email]‬‏>:‬
The settings are very relevant to having an equal number of containers running on each node if you have an idle cluster and want to distribute containers for a single job.  An application master submits requests for container allocations to the ResourceManager.  The MRAppMaster will request all the map containers at once, the FairScheduler will find NodeManagers with capacity to fulfill the container requests.  If assign multiple is enabled then you generally won't get an even number of containers assigned to each node +/- 1 container.  Before you say it's not relevant, you should check if your environment uses the FairScheduler and whether multiple assignment is enabled.  If so, that's likely why there isn't an even assignment +/- 1 container.  If not using FairScheduler and/or multiple assign, then you should look at locality settings, which can cause containers to be preferentially run on a subset of nodes, resulting in an uneven container assignment per node.  

On Wed, Jan 9, 2019 at 2:19 PM Or Raz <[hidden email]> wrote:
As far as I know, the scheduler in YARN is only scheduling the jobs and not the containers inside each job. Therefore, I don't believe it is relevant.
Also, I haven't used or set those two parameters, and I haven't picked nor set any particular schedule for my research (Fair, FIFO or Capacity).
Please correct if I am wrong.
P.S. currently I have no interest in a situation when I run a few jobs concurrently, my case is much simpler with one job that I would like that allocation of containers will be more balanced...
Or


‫בתאריך יום ד׳, 9 בינו׳ 2019 ב-19:11 מאת ‪Aaron Eng‬‏ <‪[hidden email]‬‏>:‬
Have you checked the yarn.scheduler.fair.assignmultiple and yarn.scheduler.fair.max.assign parameters for the ResourceManager configuration?

On Wed, Jan 9, 2019 at 9:49 AM Or Raz <[hidden email]> wrote:

How can I change/suggest a different allocation of containers to tasks in Hadoop? Regarding a native Hadoop (2.9.1) cluster on AWS.

I am running a native Hadoop cluster (2.9.1) on AWS (with EC2, not EMR) and I want the scheduling/allocating of the containers (Mappers/Reducers) would be more balanced than it is currently. It seems like RM is assigning the Mappers in a Bin Packing way (where the data resides) and for the reducers, it looks more balanced. My setup includes three Machines with replication rate three (all the data is on every machine), and I run my jobs with mapreduce.job.reduce.slowstart.completedmaps=0 to start shuffle as fast as possible (It is vital for me that all the containers are working in concurrency, it is a must condition). Also, according to the EC2 instances I have chosen and my settings of the YARN cluster, I can run at most 93 containers (31 each).

For example, if I want to have nine reducers then (93-9-1=83), 83 containers could be left for the mappers, and one is for the AM. I have played with the size of split input (mapreduce.input.fileinputformat.split.minsize, mapreduce.input.fileinputformat.split.maxsize) to find the right balance where all of the machines have the same "work" for the map phase. But it seems like the first 31 mappers would be allocated in one computer, the next 31 to the second one and the last 31 in the last machine. Thus, I can try to use 87 mappers where 31 of them in Machine #1, another 31 in Machine #2 and another 25 in Machine #3 and the rest is left for the reducers and as Machine #1 and Machine #2 are fully occupied then the reducers would have to be placed in Machine #3. This way I get an almost balanced allocation of mappers at the expense of unbalanced reducers allocation. And this is not what I want...

# of mappers = size_input / split size [Bytes]

split size =max(mapreduce.input.fileinputformat.split.minsize,min(mapreduce.input.fileinputformat.split.maxsize, dfs.blocksize))

Reply | Threaded
Open this post in threaded view
|

Re: Allocation of containers to tasks in Hadoop

Hariharan
Not an expert on capacity scheduler but the above two are not queue-level configurations, so I think the changes would not reflect on running refreshqueues. You would need to restart the RM for the new values to take effect.

Thanks,
Hari

On Thu, Jan 10, 2019 at 7:41 PM Or Raz <[hidden email]> wrote:
I have googled more about it, and it seems like two parameters should define the "bin packing problem".
According to https://hadoop.apache.org/docs/r2.9.1/hadoop-yarn/hadoop-yarn-site/CapacityScheduler.html#Other_Properties  yarn.scheduler.capacity.per-node-heartbeat.multiple-assignments-enabled is by default set to true and with parameter yarn.scheduler.capacity.per-node-heartbeat.maximum-container-assignments r set to -1 it can assign all the containers the Node manager "said" it is capable of (which could somehow explain the bin packing problem for the first Nodemanager who answer with a Heartbeat message).
Following Apache's instructions, I have inserted to my capacity-scheduler.xml  in hadoop/etc/hadoop folder 

  <property>
    <name>yarn.scheduler.capacity.per-node-heartbeat.multiple-assignments-enabled</name>
    <value>true</value>
    <description>
        Whether to allow multiple container assignments in one NodeManager heartbeat. Defaults to true.
    </description>
  </property>
  <property>
    <name>yarn.scheduler.capacity.per-node-heartbeat.maximum-container-assignments</name>
    <value>2</value>
    <description>
        If multiple-assignments-enabled is true, the maximum amount of containers that can be assigned in one NodeManager heartbeat. Defaults to -1, which sets no limit.
    </description>
  </property>
I have checked the configuration file, and I am using the capacity scheduler (I have enabled yarn.scheduler.capacity.per-node-heartbeat.multiple-assignments-enabled again just to be sure).
Furthermore, after I have run "yarn rmadmin -refreshQueues" I haven't seen any change in the Mappers allocation nor Reducers.
hadoop2@master:~$ yarn rmadmin -refreshQueues
19/01/10 16:06:33 INFO client.RMProxy: Connecting to ResourceManager at master/172.31.24.83:8033

What am I missing over here?

Or


‫בתאריך יום ד׳, 9 בינו׳ 2019 ב-23:57 מאת ‪Or Raz‬‏ <‪[hidden email]‬‏>:‬
Thanks for the tips!
Because I haven't set any scheduler (on purpose) for YARN then, I am using the default one (Capacity).
I have looked in yarn-site.xml and in the configuration tab (using JobHistory UI), and both of the parameters that you have mentioned weren't there (so they haven't been set).
You said that I should look at "locality settings" can you be more specific on what and where to look? 
Also, it is worth mentioning that I am using three computers and the replication factor (of HDFS) is three too. Thus, every data (even input) would be on every computer, and the memory of each computer is the same (two t2.xlarge and one m4.xlarge) while I am using DefaultResourceCalculator.

Or

‫בתאריך יום ד׳, 9 בינו׳ 2019 ב-23:28 מאת ‪Aaron Eng‬‏ <‪[hidden email]‬‏>:‬
The settings are very relevant to having an equal number of containers running on each node if you have an idle cluster and want to distribute containers for a single job.  An application master submits requests for container allocations to the ResourceManager.  The MRAppMaster will request all the map containers at once, the FairScheduler will find NodeManagers with capacity to fulfill the container requests.  If assign multiple is enabled then you generally won't get an even number of containers assigned to each node +/- 1 container.  Before you say it's not relevant, you should check if your environment uses the FairScheduler and whether multiple assignment is enabled.  If so, that's likely why there isn't an even assignment +/- 1 container.  If not using FairScheduler and/or multiple assign, then you should look at locality settings, which can cause containers to be preferentially run on a subset of nodes, resulting in an uneven container assignment per node.  

On Wed, Jan 9, 2019 at 2:19 PM Or Raz <[hidden email]> wrote:
As far as I know, the scheduler in YARN is only scheduling the jobs and not the containers inside each job. Therefore, I don't believe it is relevant.
Also, I haven't used or set those two parameters, and I haven't picked nor set any particular schedule for my research (Fair, FIFO or Capacity).
Please correct if I am wrong.
P.S. currently I have no interest in a situation when I run a few jobs concurrently, my case is much simpler with one job that I would like that allocation of containers will be more balanced...
Or


‫בתאריך יום ד׳, 9 בינו׳ 2019 ב-19:11 מאת ‪Aaron Eng‬‏ <‪[hidden email]‬‏>:‬
Have you checked the yarn.scheduler.fair.assignmultiple and yarn.scheduler.fair.max.assign parameters for the ResourceManager configuration?

On Wed, Jan 9, 2019 at 9:49 AM Or Raz <[hidden email]> wrote:

How can I change/suggest a different allocation of containers to tasks in Hadoop? Regarding a native Hadoop (2.9.1) cluster on AWS.

I am running a native Hadoop cluster (2.9.1) on AWS (with EC2, not EMR) and I want the scheduling/allocating of the containers (Mappers/Reducers) would be more balanced than it is currently. It seems like RM is assigning the Mappers in a Bin Packing way (where the data resides) and for the reducers, it looks more balanced. My setup includes three Machines with replication rate three (all the data is on every machine), and I run my jobs with mapreduce.job.reduce.slowstart.completedmaps=0 to start shuffle as fast as possible (It is vital for me that all the containers are working in concurrency, it is a must condition). Also, according to the EC2 instances I have chosen and my settings of the YARN cluster, I can run at most 93 containers (31 each).

For example, if I want to have nine reducers then (93-9-1=83), 83 containers could be left for the mappers, and one is for the AM. I have played with the size of split input (mapreduce.input.fileinputformat.split.minsize, mapreduce.input.fileinputformat.split.maxsize) to find the right balance where all of the machines have the same "work" for the map phase. But it seems like the first 31 mappers would be allocated in one computer, the next 31 to the second one and the last 31 in the last machine. Thus, I can try to use 87 mappers where 31 of them in Machine #1, another 31 in Machine #2 and another 25 in Machine #3 and the rest is left for the reducers and as Machine #1 and Machine #2 are fully occupied then the reducers would have to be placed in Machine #3. This way I get an almost balanced allocation of mappers at the expense of unbalanced reducers allocation. And this is not what I want...

# of mappers = size_input / split size [Bytes]

split size =max(mapreduce.input.fileinputformat.split.minsize,min(mapreduce.input.fileinputformat.split.maxsize, dfs.blocksize))

Reply | Threaded
Open this post in threaded view
|

Re: Allocation of containers to tasks in Hadoop

Or Raz
Thanks a lot! 
It is now working better!
Such a small parameter that I didn't know that exists and is not so common to modify.

Or

‫בתאריך יום ה׳, 10 בינו׳ 2019 ב-16:31 מאת ‪Hariharan‬‏ <‪[hidden email]‬‏>:‬
Not an expert on capacity scheduler but the above two are not queue-level configurations, so I think the changes would not reflect on running refreshqueues. You would need to restart the RM for the new values to take effect.

Thanks,
Hari

On Thu, Jan 10, 2019 at 7:41 PM Or Raz <[hidden email]> wrote:
I have googled more about it, and it seems like two parameters should define the "bin packing problem".
According to https://hadoop.apache.org/docs/r2.9.1/hadoop-yarn/hadoop-yarn-site/CapacityScheduler.html#Other_Properties  yarn.scheduler.capacity.per-node-heartbeat.multiple-assignments-enabled is by default set to true and with parameter yarn.scheduler.capacity.per-node-heartbeat.maximum-container-assignments r set to -1 it can assign all the containers the Node manager "said" it is capable of (which could somehow explain the bin packing problem for the first Nodemanager who answer with a Heartbeat message).
Following Apache's instructions, I have inserted to my capacity-scheduler.xml  in hadoop/etc/hadoop folder 

  <property>
    <name>yarn.scheduler.capacity.per-node-heartbeat.multiple-assignments-enabled</name>
    <value>true</value>
    <description>
        Whether to allow multiple container assignments in one NodeManager heartbeat. Defaults to true.
    </description>
  </property>
  <property>
    <name>yarn.scheduler.capacity.per-node-heartbeat.maximum-container-assignments</name>
    <value>2</value>
    <description>
        If multiple-assignments-enabled is true, the maximum amount of containers that can be assigned in one NodeManager heartbeat. Defaults to -1, which sets no limit.
    </description>
  </property>
I have checked the configuration file, and I am using the capacity scheduler (I have enabled yarn.scheduler.capacity.per-node-heartbeat.multiple-assignments-enabled again just to be sure).
Furthermore, after I have run "yarn rmadmin -refreshQueues" I haven't seen any change in the Mappers allocation nor Reducers.
hadoop2@master:~$ yarn rmadmin -refreshQueues
19/01/10 16:06:33 INFO client.RMProxy: Connecting to ResourceManager at master/172.31.24.83:8033

What am I missing over here?

Or


‫בתאריך יום ד׳, 9 בינו׳ 2019 ב-23:57 מאת ‪Or Raz‬‏ <‪[hidden email]‬‏>:‬
Thanks for the tips!
Because I haven't set any scheduler (on purpose) for YARN then, I am using the default one (Capacity).
I have looked in yarn-site.xml and in the configuration tab (using JobHistory UI), and both of the parameters that you have mentioned weren't there (so they haven't been set).
You said that I should look at "locality settings" can you be more specific on what and where to look? 
Also, it is worth mentioning that I am using three computers and the replication factor (of HDFS) is three too. Thus, every data (even input) would be on every computer, and the memory of each computer is the same (two t2.xlarge and one m4.xlarge) while I am using DefaultResourceCalculator.

Or

‫בתאריך יום ד׳, 9 בינו׳ 2019 ב-23:28 מאת ‪Aaron Eng‬‏ <‪[hidden email]‬‏>:‬
The settings are very relevant to having an equal number of containers running on each node if you have an idle cluster and want to distribute containers for a single job.  An application master submits requests for container allocations to the ResourceManager.  The MRAppMaster will request all the map containers at once, the FairScheduler will find NodeManagers with capacity to fulfill the container requests.  If assign multiple is enabled then you generally won't get an even number of containers assigned to each node +/- 1 container.  Before you say it's not relevant, you should check if your environment uses the FairScheduler and whether multiple assignment is enabled.  If so, that's likely why there isn't an even assignment +/- 1 container.  If not using FairScheduler and/or multiple assign, then you should look at locality settings, which can cause containers to be preferentially run on a subset of nodes, resulting in an uneven container assignment per node.  

On Wed, Jan 9, 2019 at 2:19 PM Or Raz <[hidden email]> wrote:
As far as I know, the scheduler in YARN is only scheduling the jobs and not the containers inside each job. Therefore, I don't believe it is relevant.
Also, I haven't used or set those two parameters, and I haven't picked nor set any particular schedule for my research (Fair, FIFO or Capacity).
Please correct if I am wrong.
P.S. currently I have no interest in a situation when I run a few jobs concurrently, my case is much simpler with one job that I would like that allocation of containers will be more balanced...
Or


‫בתאריך יום ד׳, 9 בינו׳ 2019 ב-19:11 מאת ‪Aaron Eng‬‏ <‪[hidden email]‬‏>:‬
Have you checked the yarn.scheduler.fair.assignmultiple and yarn.scheduler.fair.max.assign parameters for the ResourceManager configuration?

On Wed, Jan 9, 2019 at 9:49 AM Or Raz <[hidden email]> wrote:

How can I change/suggest a different allocation of containers to tasks in Hadoop? Regarding a native Hadoop (2.9.1) cluster on AWS.

I am running a native Hadoop cluster (2.9.1) on AWS (with EC2, not EMR) and I want the scheduling/allocating of the containers (Mappers/Reducers) would be more balanced than it is currently. It seems like RM is assigning the Mappers in a Bin Packing way (where the data resides) and for the reducers, it looks more balanced. My setup includes three Machines with replication rate three (all the data is on every machine), and I run my jobs with mapreduce.job.reduce.slowstart.completedmaps=0 to start shuffle as fast as possible (It is vital for me that all the containers are working in concurrency, it is a must condition). Also, according to the EC2 instances I have chosen and my settings of the YARN cluster, I can run at most 93 containers (31 each).

For example, if I want to have nine reducers then (93-9-1=83), 83 containers could be left for the mappers, and one is for the AM. I have played with the size of split input (mapreduce.input.fileinputformat.split.minsize, mapreduce.input.fileinputformat.split.maxsize) to find the right balance where all of the machines have the same "work" for the map phase. But it seems like the first 31 mappers would be allocated in one computer, the next 31 to the second one and the last 31 in the last machine. Thus, I can try to use 87 mappers where 31 of them in Machine #1, another 31 in Machine #2 and another 25 in Machine #3 and the rest is left for the reducers and as Machine #1 and Machine #2 are fully occupied then the reducers would have to be placed in Machine #3. This way I get an almost balanced allocation of mappers at the expense of unbalanced reducers allocation. And this is not what I want...

# of mappers = size_input / split size [Bytes]

split size =max(mapreduce.input.fileinputformat.split.minsize,min(mapreduce.input.fileinputformat.split.maxsize, dfs.blocksize))