[DISCUSSION] CarbonData Integration with Presto

[bhavya411]

We are using  presto in our company for querying the data , the query processing is slower than what is the expectation of the client , we have been looking at the Apache CarbonData as a storage format so to make a decision we ran TPCH benchmarking on presto with CarbonData and compared it with Parquet. CarbonData community ask us to share the test result to mailing list for discussion,  so please see below for the details of the benchmarking that we did.

Question : We were able to run only 16 query set out of 22 query set as 2 query set involved temporary tables and 4 query set involved Views which are also not supported by Carbondata. Is there a way to run all the queries.

Environment

Cluster : 3 Node Cluster (48 GB RAM , 8 CPU Core and 2 TB hard-disk each) 

Data

Data Set : 50 GB data was generated using TPCH 2.17.2 (Schema is attached)

Results  :  Results are attached in the xlsx with correcponding queries.

The Carbon Data is performing better than Parquet , but since I have to make a decision on it we are trying to do benchmarking on more data and also compare it with ORC.

[Erlu Chen]

Wonderful!

The performance of Carbondata is much better than Parquet.

Regards.
Chenerlu.

[linqer]

Can you give me your all configuration files (etc/) of coordinate and worker, What configuration tuning did you make for carbondata?

A lot of scenes in our company are using Presto,I have tested presto read orc vs carbondata  , Carbondata is clearly inferior to Orc;

During the testing, for performance reasons, we used replica join in particular, and we increased the JVM codecache size. but these are equally fair to both ORC and carbondata

Carbondata_vs_ORC_on_Presto_Benchmark_testing.docx

[bhavya411]

Hi,

Please find the configuration setting that we used attached with this email , we are running Presto Server 0.179 for our testing.

Thanks and regards

Bhavya

On Fri, Jun 30, 2017 at 8:23 AM, linqer <[hidden email]> wrote:

Can you give me your all configuration files (etc/) of coordinate and worker,
What configuration tuning did you make for carbondata?

A lot of scenes in our company are using Presto,I have tested presto read
orc vs carbondata  , Carbondata is clearly inferior to Orc;

During the testing, for performance reasons, we used replica join in
particular, and we increased the JVM codecache size. but these are equally
fair to both ORC and carbondata

Carbondata_vs_ORC_on_Presto_Benchmark_testing.docx
<http://apache-carbondata-dev-mailing-list-archive.1130556.n5.nabble.com/file/n16849/Carbondata_vs_ORC_on_Presto_Benchmark_testing.docx>






--
View this message in context: http://apache-carbondata-dev-mailing-list-archive.1130556.n5.nabble.com/DISCUSSION-CarbonData-Integration-with-Presto-tp16793p16849.html

Sent from the Apache CarbonData Dev Mailing List archive mailing list archive at Nabble.com.

[Liang Chen-2]

Hi Bhavya

Currently, 1.2.0 propose to support presto version with 0.166.
Is there any performance difference between 0.179 and 0.166?

Regards
Liang


2017-07-01 13:12 GMT+08:00 Bhavya Aggarwal <[hidden email]>:

> Hi,
>
> Please find the configuration setting that we used attached with this
> email , we are running Presto Server 0.179 for our testing.
>
> Thanks and regards
> Bhavya
>
> On Fri, Jun 30, 2017 at 8:23 AM, linqer <[hidden email]> wrote:
>
>> Can you give me your all configuration files (etc/) of coordinate and
>> worker,
>> What configuration tuning did you make for carbondata?
>>
>> A lot of scenes in our company are using Presto,I have tested presto read
>> orc vs carbondata  , Carbondata is clearly inferior to Orc;
>>
>> During the testing, for performance reasons, we used replica join in
>> particular, and we increased the JVM codecache size. but these are equally
>> fair to both ORC and carbondata
>>
>> Carbondata_vs_ORC_on_Presto_Benchmark_testing.docx
>> <http://apache-carbondata-dev-mailing-list-archive.1130556.n
>> 5.nabble.com/file/n16849/Carbondata_vs_ORC_on_Presto_Benchma
>> rk_testing.docx>
>>
>>
>>
>>
>>
>>
>> --
>> View this message in context: http://apache-carbondata-dev-m
>> ailing-list-archive.1130556.n5.nabble.com/DISCUSSION-CarbonD
>> ata-Integration-with-Presto-tp16793p16849.html
>> Sent from the Apache CarbonData Dev Mailing List archive mailing list
>> archive at Nabble.com.
>>
>
>

[linqer]

Hi
1 you can set -XX:ReservedCodeCacheSize in JVM.propertes, make it big
2 I don't see your discovery.uri IP,  make user Master act as coordinate and discovery server;
3 remove query.max-memory-per-node firstly.
4 you can test orc, parquet is not support very well on presto.

thanks

[linqer]

for your first testing, What configuration tuning did you make for carbondata?  and could you provide a detailed testing report including create table sqls, the row numbers of tables etc.

[bhavya411]

Hi Liang,

We went through the documentation of Presto and there were some issues with
Presto 0.166 version which were resolved in later versions. There is a lot
of performance improvement in 0.166 and 0.179  as the ways joins are
interpreted in Presto are different from 0.166. Please see below for two
most important reasons why we choose to ran it on 0.179.

1. Fixed issue which could cause incorrect results when processing
dictionary encoded data. If the expression can fail on bad input, the
results from filtered-out rows containing bad input may be included in the
query output.

2. The order in which joins are executed in a query can have a significant
impact on the query’s performance. The aspect of join ordering that has the
largest impact on performance is the size of the data being processed and
passed over the network. If a join is not a primary key-foreign key join,
the data produced can be much greater than the size of either table in the
join– up to |Table 1| x |Table 2| for a cross join. If a join that produces
a lot of data is performed early in the execution, then subsequent stages
will need to process large amounts of data for longer than necessary,
increasing the time and resources needed for the query causing query
failure. This issue has been fixed in the presto-version 11.3. Release
0.178 onwards.

Thanks and regards
Bhavya


On Sun, Jul 2, 2017 at 10:41 AM, Liang Chen <[hidden email]> wrote:

> Hi Bhavya
>
> Currently, 1.2.0 propose to support presto version with 0.166.
> Is there any performance difference between 0.179 and 0.166?
>
> Regards
> Liang
>
>
> 2017-07-01 13:12 GMT+08:00 Bhavya Aggarwal <[hidden email]>:
>
> > Hi,
> >
> > Please find the configuration setting that we used attached with this
> > email , we are running Presto Server 0.179 for our testing.
> >
> > Thanks and regards
> > Bhavya
> >
> > On Fri, Jun 30, 2017 at 8:23 AM, linqer <[hidden email]> wrote:
> >
> >> Can you give me your all configuration files (etc/) of coordinate and
> >> worker,
> >> What configuration tuning did you make for carbondata?
> >>
> >> A lot of scenes in our company are using Presto,I have tested presto
> read
> >> orc vs carbondata  , Carbondata is clearly inferior to Orc;
> >>
> >> During the testing, for performance reasons, we used replica join in
> >> particular, and we increased the JVM codecache size. but these are
> equally
> >> fair to both ORC and carbondata
> >>
> >> Carbondata_vs_ORC_on_Presto_Benchmark_testing.docx
> >> <http://apache-carbondata-dev-mailing-list-archive.1130556.n
> >> 5.nabble.com/file/n16849/Carbondata_vs_ORC_on_Presto_Benchma
> >> rk_testing.docx>
> >>
> >>
> >>
> >>
> >>
> >>
> >> --
> >> View this message in context: http://apache-carbondata-dev-m
> >> ailing-list-archive.1130556.n5.nabble.com/DISCUSSION-CarbonD
> >> ata-Integration-with-Presto-tp16793p16849.html
> >> Sent from the Apache CarbonData Dev Mailing List archive mailing list
> >> archive at Nabble.com.
> >>
> >
> >
>

[bhavya411]

Thanks Linquer,

I will try with the above option and we tried it with ORC as well , here are the comparison results for same query with 50 GB of data and same configuration that I mentioned earlier.

Thanks and regards

Bhavya

On Mon, Jul 3, 2017 at 9:39 AM, linqer <[hidden email]> wrote:

Hi
1 you can set -XX:ReservedCodeCacheSize in JVM.propertes, make it big
2 I don't see your discovery.uri IP,  make user Master act as coordinate and
discovery server;
3 remove query.max-memory-per-node firstly.
4 you can test orc, parquet is not support very well on presto.

thanks



--
View this message in context: http://apache-carbondata-dev-mailing-list-archive.1130556.n5.nabble.com/DISCUSSION-CarbonData-Integration-with-Presto-tp16793p17055.html

Sent from the Apache CarbonData Dev Mailing List archive mailing list archive at Nabble.com.

[bhavya411]

Please see the image as there was some problem with the image in previous email.

Thanks and regards

Bhavya

On Mon, Jul 3, 2017 at 11:45 AM, Bhavya Aggarwal <[hidden email]> wrote:

Thanks Linquer,

I will try with the above option and we tried it with ORC as well , here are the comparison results for same query with 50 GB of data and same configuration that I mentioned earlier.

Thanks and regards

Bhavya

On Mon, Jul 3, 2017 at 9:39 AM, linqer <[hidden email]> wrote:

Hi
1 you can set -XX:ReservedCodeCacheSize in JVM.propertes, make it big
2 I don't see your discovery.uri IP,  make user Master act as coordinate and
discovery server;
3 remove query.max-memory-per-node firstly.
4 you can test orc, parquet is not support very well on presto.

thanks



--
View this message in context: http://apache-carbondata-dev-mailing-list-archive.1130556.n5.nabble.com/DISCUSSION-CarbonData-Integration-with-Presto-tp16793p17055.html

Sent from the Apache CarbonData Dev Mailing List archive mailing list archive at Nabble.com.

[linqer]

hi, did you test orc vs carbondata with configuration as I  suggested?

[bhavya411]

I am currently loading 500 GB data once that is completed will try with
your settings today itself.

Regards
Bhavya
On 06-Jul-2017 8:39 am, "linqer" <[hidden email]> wrote:

> hi, did you test orc vs carbondata with configuration as I  suggested?
>
>
>
> --
> View this message in context: http://apache-carbondata-dev-
> mailing-list-archive.1130556.n5.nabble.com/DISCUSSION-
> CarbonData-Integration-with-Presto-tp16793p17419.html
> Sent from the Apache CarbonData Dev Mailing List archive mailing list
> archive at Nabble.com.
>

[bhavya411]

Hi Linquer,

I am  trying to run with the configuration that you suggested on 500 GB data , can you please verify the below properties.

config.queries

coordinator=true
node-scheduler.include-coordinator=true
http-server.http.port=8086
query.max-memory=5GB
#query.max-memory-per-node=1GB
discovery-server.enabled=true
discovery.uri=http://xx.xx.xx.xx:8086

jvm.config

-server
-Xmx16G
-XX:+UseG1GC
-XX:G1HeapRegionSize=32M
-XX:+UseGCOverheadLimit
-XX:+ExplicitGCInvokesConcurrent
-XX:+HeapDumpOnOutOfMemoryError
-XX:OnOutOfMemoryError=kill -9 %p
-XX:ReservedCodeCacheSize=1G

Regards

Bhavya

On Mon, Jul 3, 2017 at 11:45 AM, Bhavya Aggarwal <[hidden email]> wrote:

Thanks Linquer,

I will try with the above option and we tried it with ORC as well , here are the comparison results for same query with 50 GB of data and same configuration that I mentioned earlier.

Thanks and regards

Bhavya

On Mon, Jul 3, 2017 at 9:39 AM, linqer <[hidden email]> wrote:

Hi
1 you can set -XX:ReservedCodeCacheSize in JVM.propertes, make it big
2 I don't see your discovery.uri IP,  make user Master act as coordinate and
discovery server;
3 remove query.max-memory-per-node firstly.
4 you can test orc, parquet is not support very well on presto.

thanks



--
View this message in context: http://apache-carbondata-dev-mailing-list-archive.1130556.n5.nabble.com/DISCUSSION-CarbonData-Integration-with-Presto-tp16793p17055.html

Sent from the Apache CarbonData Dev Mailing List archive mailing list archive at Nabble.com.

[linqer]

there is no obvious issue in your configuration,  did you have finish testing on ORC?  

[linqer]

you can add
-XX:+PrintGCDateStamps
-XX:+PrintGCApplicationStoppedTime
-XX:+PrintGCDetails
-Xloggc:/var/presto/data/var/log/presto-gc.log

into you jvm.properties, if you think there is a gc issue , you can check the GC logs.