You are viewing a plain text version of this content. The canonical link for it is here.
Posted to user@hive.apache.org by Vijay <te...@gmail.com> on 2011/02/03 21:49:01 UTC
Hive queries consuming 100% cpu
Hi,
The simplest of hive queries seem to be consuming 100% cpu. This is
with a small 4-node cluster. The machines are pretty beefy (16 cores
per machine, tons of RAM, 16 M+R maximum tasks configured, 1GB RAM for
mapred.child.java.opts, etc). A simple query like "select count(1)
from events" where the events table has daily partitions of log files
in gzipped file format). While this is probably too generic a question
and there is a bunch of investigation we need to, are there any
specific areas for me to look at? Has anyone see anything like this
before? Also, are there any tools or easy options to profile hive
query execution?
Thanks in advance,
Vijay
Re: Hive queries consuming 100% cpu
Posted by Vijay <te...@gmail.com>.
Sorry i should've given more details.
The query was limited by a partition range; I just omitted the WHERE
clause in the mail.
The table is not that big. For each day, there is one gzipped file.
The largest file is about 250MB (close to 2GB uncompressed).
I did intend to count and that was just to test since I wanted to run
a query that did the most minimal logic/processing.
Here's a test I ran now. The query is getting count(1) for 8 days. It
spawned 8 maps as expected. The maps run for anywhere between 42 to 69
seconds (which may or may not be right; I need to check that). It
spawned only one reduce task. The reducer ran for 117 seconds, which
seems long for this query.
On Thu, Feb 3, 2011 at 2:31 PM, Viral Bajaria <vi...@gmail.com> wrote:
> Hey Vijay,
> You can go to the mapred ui, normally it runs on port 50030 of the namenode
> and see how many map jobs got created for your submitted query.
> You said that the events table has daily partitions but the example query
> that you have does not prune the partitions by specifying a WHERE clause. So
> I have the following questions
> 1) how big is the table (you can just do a hadoop dfs -dus
> <hdfs-dir-for-table> ? how many partitions ?
> 2) do you really intend to count the number of events across all days ?
> 3) could you build a query which computes over 1-5 day(s) and persists the
> data in a separate table for consumption later on ?
> Based on your node configuration, I am just guessing the amount of data to
> process is too large and hence the high CPU.
> Thanks,
> Viral
> On Thu, Feb 3, 2011 at 12:49 PM, Vijay <te...@gmail.com> wrote:
>>
>> Hi,
>>
>> The simplest of hive queries seem to be consuming 100% cpu. This is
>> with a small 4-node cluster. The machines are pretty beefy (16 cores
>> per machine, tons of RAM, 16 M+R maximum tasks configured, 1GB RAM for
>> mapred.child.java.opts, etc). A simple query like "select count(1)
>> from events" where the events table has daily partitions of log files
>> in gzipped file format). While this is probably too generic a question
>> and there is a bunch of investigation we need to, are there any
>> specific areas for me to look at? Has anyone see anything like this
>> before? Also, are there any tools or easy options to profile hive
>> query execution?
>>
>> Thanks in advance,
>> Vijay
>
>
Re: Hive queries consuming 100% cpu
Posted by Vijay <te...@gmail.com>.
Sorry i should've given more details.
The query was limited by a partition range; I just omitted the WHERE
clause in the mail.
The table is not that big. For each day, there is one gzipped file.
The largest file is about 250MB (close to 2GB uncompressed).
I did intend to count and that was just to test since I wanted to run
a query that did the most minimal logic/processing.
Here's a test I ran now. The query is getting count(1) for 8 days. It
spawned 8 maps as expected. The maps run for anywhere between 42 to 69
seconds (which may or may not be right; I need to check that). It
spawned only one reduce task. The reducer ran for 117 seconds, which
seems long for this query.
On Thu, Feb 3, 2011 at 2:31 PM, Viral Bajaria <vi...@gmail.com> wrote:
> Hey Vijay,
> You can go to the mapred ui, normally it runs on port 50030 of the namenode
> and see how many map jobs got created for your submitted query.
> You said that the events table has daily partitions but the example query
> that you have does not prune the partitions by specifying a WHERE clause. So
> I have the following questions
> 1) how big is the table (you can just do a hadoop dfs -dus
> <hdfs-dir-for-table> ? how many partitions ?
> 2) do you really intend to count the number of events across all days ?
> 3) could you build a query which computes over 1-5 day(s) and persists the
> data in a separate table for consumption later on ?
> Based on your node configuration, I am just guessing the amount of data to
> process is too large and hence the high CPU.
> Thanks,
> Viral
> On Thu, Feb 3, 2011 at 12:49 PM, Vijay <te...@gmail.com> wrote:
>>
>> Hi,
>>
>> The simplest of hive queries seem to be consuming 100% cpu. This is
>> with a small 4-node cluster. The machines are pretty beefy (16 cores
>> per machine, tons of RAM, 16 M+R maximum tasks configured, 1GB RAM for
>> mapred.child.java.opts, etc). A simple query like "select count(1)
>> from events" where the events table has daily partitions of log files
>> in gzipped file format). While this is probably too generic a question
>> and there is a bunch of investigation we need to, are there any
>> specific areas for me to look at? Has anyone see anything like this
>> before? Also, are there any tools or easy options to profile hive
>> query execution?
>>
>> Thanks in advance,
>> Vijay
>
>
Re: Hive queries consuming 100% cpu
Posted by Viral Bajaria <vi...@gmail.com>.
Hey Vijay,
You can go to the mapred ui, normally it runs on port 50030 of the namenode
and see how many map jobs got created for your submitted query.
You said that the events table has daily partitions but the example query
that you have does not prune the partitions by specifying a WHERE clause. So
I have the following questions
1) how big is the table (you can just do a hadoop dfs -dus
<hdfs-dir-for-table> ? how many partitions ?
2) do you really intend to count the number of events across all days ?
3) could you build a query which computes over 1-5 day(s) and persists the
data in a separate table for consumption later on ?
Based on your node configuration, I am just guessing the amount of data to
process is too large and hence the high CPU.
Thanks,
Viral
On Thu, Feb 3, 2011 at 12:49 PM, Vijay <te...@gmail.com> wrote:
> Hi,
>
> The simplest of hive queries seem to be consuming 100% cpu. This is
> with a small 4-node cluster. The machines are pretty beefy (16 cores
> per machine, tons of RAM, 16 M+R maximum tasks configured, 1GB RAM for
> mapred.child.java.opts, etc). A simple query like "select count(1)
> from events" where the events table has daily partitions of log files
> in gzipped file format). While this is probably too generic a question
> and there is a bunch of investigation we need to, are there any
> specific areas for me to look at? Has anyone see anything like this
> before? Also, are there any tools or easy options to profile hive
> query execution?
>
> Thanks in advance,
> Vijay
>
Re: Hive queries consuming 100% cpu
Posted by Viral Bajaria <vi...@gmail.com>.
Hey Vijay,
You can go to the mapred ui, normally it runs on port 50030 of the namenode
and see how many map jobs got created for your submitted query.
You said that the events table has daily partitions but the example query
that you have does not prune the partitions by specifying a WHERE clause. So
I have the following questions
1) how big is the table (you can just do a hadoop dfs -dus
<hdfs-dir-for-table> ? how many partitions ?
2) do you really intend to count the number of events across all days ?
3) could you build a query which computes over 1-5 day(s) and persists the
data in a separate table for consumption later on ?
Based on your node configuration, I am just guessing the amount of data to
process is too large and hence the high CPU.
Thanks,
Viral
On Thu, Feb 3, 2011 at 12:49 PM, Vijay <te...@gmail.com> wrote:
> Hi,
>
> The simplest of hive queries seem to be consuming 100% cpu. This is
> with a small 4-node cluster. The machines are pretty beefy (16 cores
> per machine, tons of RAM, 16 M+R maximum tasks configured, 1GB RAM for
> mapred.child.java.opts, etc). A simple query like "select count(1)
> from events" where the events table has daily partitions of log files
> in gzipped file format). While this is probably too generic a question
> and there is a bunch of investigation we need to, are there any
> specific areas for me to look at? Has anyone see anything like this
> before? Also, are there any tools or easy options to profile hive
> query execution?
>
> Thanks in advance,
> Vijay
>