You are viewing a plain text version of this content. The canonical link for it is here.
Posted to user@hive.apache.org by Sungwoo Park <gl...@gmail.com> on 2025/04/22 05:44:05 UTC
Performance evaluation of Trino 468, Spark 4.0.0-RC2, and Hive 4 on Tez/MR3
Hello,
We published a blog that reports the performance evaluation of Trino 468,
Spark 4.0.0-RC2, and Hive 4 on Tez/MR3 2.0 using the TPC-DS Benchmark, 10TB
scale factor. Hope you find it useful.
https://mr3docs.datamonad.com/blog/2025-04-18-performance-evaluation-2.0
--- Sungwoo
Re:Re: Performance evaluation of Trino 468, Spark 4.0.0-RC2, and Hive 4 on Tez/MR3
Posted by lisoda <li...@yeah.net>.
Hello,Okumin.
Currently, from the TPCDS test results, we observe that, in some scenarios, neither HIVE nor HIVE ON MR3 seems to be able to achieve satisfactory performance. It appears that HIVE4 has not generated a good execution plan.
Generally, the typical scenarios for these problematic queries are: query 97, query 29, query 75, query 11, query 65, query 4, and query 78.
Among them, Trino has recently introduced some optimizations. The execution time for query 11 and query 4 has decreased by nearly half compared to Trino 440. Is it possible for us to transplant Trino's optimization strategies or investigate the reasons for the poor execution plans of these typical SQL queries?
Tks.
Lisoda.
在 2025-04-22 16:37:29，"Sungwoo Park" <gl...@gmail.com> 写道：
From average response time analysis:
For Spark, it performs better than its total execution time suggests, with an average response time significantly lower than Hive on Tez.
For long-running complex queries (like query 24) on large datasets, Hive on Tez can be a better choice than Spark, even with its initial overhead of starting YARN containers.
--- Sungwoo
On Tue, Apr 22, 2025 at 2:52 PM ypeng <yp...@t-online.de> wrote:
Thanks for the doc.
I am surprised to see spark 4 is even slower than hive on Tez.
[Total Execution Time (Sequential). Trino is the fastest, followed
closely by Hive on MR3, which significantly outperformed Hive on Tez.
Spark is the slowest, skewed by a few outlier queries.]
Sungwoo Park:
> We published a blog that reports the performance evaluation of Trino
> 468, Spark 4.0.0-RC2, and Hive 4 on Tez/MR3 2.0 using the TPC-DS
> Benchmark, 10TB scale factor. Hope you find it useful.
Re: Performance evaluation of Trino 468, Spark 4.0.0-RC2, and Hive 4 on Tez/MR3
Posted by ypeng <yp...@t-online.de>.
can your next testing have kudu + impala included?
Thanks.
Sungwoo Park:
> For Hive-Tez, DAGAppMaster is reused across queries. Only worker
> containers are not reused across queries.
Re: Re: Performance evaluation of Trino 468, Spark 4.0.0-RC2, and Hive 4 on Tez/MR3
Posted by Sungwoo Park <gl...@gmail.com>.
Evaluating Hive4-LLAP can be of interest to many users in this mailing
list, but it's a lot of work and we are not sure if we can finish parameter
tuning to achieve the best performance.
For Hive-Tez, DAGAppMaster is reused across queries. Only worker containers
are not reused across queries.
--- Sungwoo
On Tue, Apr 22, 2025 at 6:25 PM lisoda <li...@yeah.net> wrote:
> Hello Sungwoo
> BTW, would you consider adding HIVE4-LLAP as a control group for the trial?
> Tks.
> Lisoda
>
>
>
>
>
> 在 2025-04-22 16:37:29，"Sungwoo Park" <gl...@gmail.com> 写道：
>
> From average response time analysis:
>
> For Spark, it performs better than its total execution time suggests, with
> an average response time significantly lower than Hive on Tez.
>
> For long-running complex queries (like query 24) on large datasets, Hive
> on Tez can be a better choice than Spark, even with its initial overhead of
> starting YARN containers.
>
> --- Sungwoo
>
> On Tue, Apr 22, 2025 at 2:52 PM ypeng <yp...@t-online.de> wrote:
>
>>
>> Thanks for the doc.
>> I am surprised to see spark 4 is even slower than hive on Tez.
>>
>>
>> [Total Execution Time (Sequential). Trino is the fastest, followed
>> closely by Hive on MR3, which significantly outperformed Hive on Tez.
>> Spark is the slowest, skewed by a few outlier queries.]
>>
>>
>> Sungwoo Park:
>> > We published a blog that reports the performance evaluation of Trino
>> > 468, Spark 4.0.0-RC2, and Hive 4 on Tez/MR3 2.0 using the TPC-DS
>> > Benchmark, 10TB scale factor. Hope you find it useful.
>>
>>
Re:Re: Performance evaluation of Trino 468, Spark 4.0.0-RC2, and Hive 4 on Tez/MR3
Posted by lisoda <li...@yeah.net>.
Hello Sungwoo
BTW, would you consider adding HIVE4-LLAP as a control group for the trial?
Tks.
Lisoda
在 2025-04-22 16:37:29，"Sungwoo Park" <gl...@gmail.com> 写道：
From average response time analysis:
For Spark, it performs better than its total execution time suggests, with an average response time significantly lower than Hive on Tez.
For long-running complex queries (like query 24) on large datasets, Hive on Tez can be a better choice than Spark, even with its initial overhead of starting YARN containers.
--- Sungwoo
On Tue, Apr 22, 2025 at 2:52 PM ypeng <yp...@t-online.de> wrote:
Thanks for the doc.
I am surprised to see spark 4 is even slower than hive on Tez.
[Total Execution Time (Sequential). Trino is the fastest, followed
closely by Hive on MR3, which significantly outperformed Hive on Tez.
Spark is the slowest, skewed by a few outlier queries.]
Sungwoo Park:
> We published a blog that reports the performance evaluation of Trino
> 468, Spark 4.0.0-RC2, and Hive 4 on Tez/MR3 2.0 using the TPC-DS
> Benchmark, 10TB scale factor. Hope you find it useful.
Re:Re: Performance evaluation of Trino 468, Spark 4.0.0-RC2, and Hive 4 on Tez/MR3
Posted by lisoda <li...@yeah.net>.
Maybe HIVE4 ON TEZ should enable the continer reuse feature so that the start/stop overhead of the application can be further reduced.
在 2025-04-22 16:37:29，"Sungwoo Park" <gl...@gmail.com> 写道：
From average response time analysis:
For Spark, it performs better than its total execution time suggests, with an average response time significantly lower than Hive on Tez.
For long-running complex queries (like query 24) on large datasets, Hive on Tez can be a better choice than Spark, even with its initial overhead of starting YARN containers.
--- Sungwoo
On Tue, Apr 22, 2025 at 2:52 PM ypeng <yp...@t-online.de> wrote:
Thanks for the doc.
I am surprised to see spark 4 is even slower than hive on Tez.
[Total Execution Time (Sequential). Trino is the fastest, followed
closely by Hive on MR3, which significantly outperformed Hive on Tez.
Spark is the slowest, skewed by a few outlier queries.]
Sungwoo Park:
> We published a blog that reports the performance evaluation of Trino
> 468, Spark 4.0.0-RC2, and Hive 4 on Tez/MR3 2.0 using the TPC-DS
> Benchmark, 10TB scale factor. Hope you find it useful.
Re: Performance evaluation of Trino 468, Spark 4.0.0-RC2, and Hive 4 on Tez/MR3
Posted by Sungwoo Park <gl...@gmail.com>.
From average response time analysis:
For Spark, it performs better than its total execution time suggests, with
an average response time significantly lower than Hive on Tez.
For long-running complex queries (like query 24) on large datasets, Hive on
Tez can be a better choice than Spark, even with its initial overhead of
starting YARN containers.
--- Sungwoo
On Tue, Apr 22, 2025 at 2:52 PM ypeng <yp...@t-online.de> wrote:
>
> Thanks for the doc.
> I am surprised to see spark 4 is even slower than hive on Tez.
>
>
> [Total Execution Time (Sequential). Trino is the fastest, followed
> closely by Hive on MR3, which significantly outperformed Hive on Tez.
> Spark is the slowest, skewed by a few outlier queries.]
>
>
> Sungwoo Park:
> > We published a blog that reports the performance evaluation of Trino
> > 468, Spark 4.0.0-RC2, and Hive 4 on Tez/MR3 2.0 using the TPC-DS
> > Benchmark, 10TB scale factor. Hope you find it useful.
>
>
Re: Performance evaluation of Trino 468, Spark 4.0.0-RC2, and Hive 4 on Tez/MR3
Posted by ypeng <yp...@t-online.de>.
Thanks for the doc.
I am surprised to see spark 4 is even slower than hive on Tez.
[Total Execution Time (Sequential). Trino is the fastest, followed
closely by Hive on MR3, which significantly outperformed Hive on Tez.
Spark is the slowest, skewed by a few outlier queries.]
Sungwoo Park:
> We published a blog that reports the performance evaluation of Trino
> 468, Spark 4.0.0-RC2, and Hive 4 on Tez/MR3 2.0 using the TPC-DS
> Benchmark, 10TB scale factor. Hope you find it useful.