March 21, 2014

Performance testing of Hive, Impala and esProc Part 2

In the previous article, we've tested the grouping computing. In this article, we will test their performances and compare their results in associating computing.
Associating computing test on narrow tables
Data sample:
Associated table p_narrow.
Col. count: 11
Row count: 500 million
Space occupied if saving as text: 120. 6G.
Data structure: personid int,name string,sex int,cityid int,birthday int,degree int,col1 string,col2 int,col3 int,col4 int,col5 string

Dimension table d_narrow
Col. count: 9
Row count: 10 million rows
Space occupied if saving as text: 563 M.
Data structure: id int, parentid int, col1 int, col2 int, col3 int, col4 int, col5 int, col6 int, col7 int
         Description:
                   Associated table: It is similar to joining the table on the left with SQL, and there are quite a lot of rows, for example, the order table.
                   Dimension table: It is similar to joining the table on the right with SQL, and there are quite a lot of rows, for example, the client ID and client name table.
Test case:
Hive:
select sum(p_narrow. col3), count(p_narrow. col5), sum(d_narrow. col7), d_narrow. id%10000 from p_narrow join d_narrow on d_narrow. id=p_narrow. col7 group by d_narrow. id%10000

esProc: The codes can be divided into 3 parts. They are respectively: Program for summary machine, main program for node machine, and subprogram for node machine.
Impala:
select sum(p_narrow. col3), count(p_narrow. col5), sum(d_narrow. col7), d_narrow. id%10000 from p_narrow join d_narrow on d_narrow. id=p_narrow. col7 group by d_narrow. id%10000
Test results:
Hive
Impala
esProc
773s
262s
279s
Result description:
1.       esProc and Impala outperform Hive obviously, almost 3 times better.
2.       Impala is slightly better than esProc, but the difference is not great.

Associating computation test on narrow tables
Data sample:
Associated tablep
Col. count: 106
Row count: 60 million rows
Space occupied if saving as text: 127. 9G.
Data structure: personid int,name string,sex int,cityid int,birthday int,degree int,col1 int,col2 int,col3 int,col4 int,col5 int,col6 int,col7 int,col8 int,col9 int,col10 int,col11 int,col12 int,col13 int,col14 int,col15 int,col16 int,col17 int,col18 int,col19 int,col20 int,col21 int,col22 int,col23 int,col24 int,col25 int,col26 int,col27 int,col28 int,col29 int,col30 int,col31 int,col32 int,col33 int,col34 int,col35 int,col36 int,col37 int,col38 int,col39 int,col40 int,col41 int,col42 int,col43 int,col44 int,col45 int,col46 int,col47 int,col48 int,col49 int,col50 int,col51 int,col52 int,col53 int,col54 int,col55 int,col56 int,col57 int,col58 int,col59 int,col60 int,col61 int,col62 int,col63 int,col64 int,col65 int,col66 int,col67 int,col68 int,col69 int,col70 int,col71 int,col72 int,col73 int,col74 int,col75 int,col76 int,col77 int,col78 int,col79 int,col80 int,col81 int,col82 int,col83 int,col84 string,col85 string,col86 string,col87 string,col88 string,col89 string,col90 string,col91 string,col92 string,col93 string,col94 string,col95 string,col96 string,col97 string,col98 string,col99 string,col100 string

         Dimension table d
         Col. count: 102
         Row count: 10 million rows
         Space occupied if saving as text: 6. 8G
         Data structure: id int, parentid int,col1 int,col2 int,col3 int,col4 int,col5 int,col6 int,col7 int,col8 int,col9 int,col10 int,col11 int,col12 int,col13 int,col14 int,col15 int,col16 int,col17 int,col18 int,col19 int,col20 int,col21 int,col22 int,col23 int,col24 int,col25 int,col26 int,col27 int,col28 int,col29 int,col30 int,col31 int,col32 int,col33 int,col34 int,col35 int,col36 int,col37 int,col38 int,col39 int,col40 int,col41 int,col42 int,col43 int,col44 int,col45 int,col46 int,col47 int,col48 int,col49 int,col50 int,col51 int,col52 int,col53 int,col54 int,col55 int,col56 int,col57 int,col58 int,col59 int,col60 int,col61 int,col62 int,col63 int,col64 int,col65 int,col66 int,col67 int,col68 int,col69 int,col70 int,col71 int,col72 int,col73 int,col74 int,col75 int,col76 int,col77 int,col78 int,col79 int,col80 int,col81 int,col82 int,col83 int,col84 int,col85 int,col86 int,col87 int,col88 int,col89 int,col90 int,col91 int,col92 int,col93 int,col94 int,col95 int,col96 int,col97 int,col98 int,col99 int,col100 int         Description:
                   Associated table: It is similar to joining the table on the left with SQL, and there are quite a lot of rows, for example, the order table.
                   Dimension table: It is similar to joining the table on the right with SQL, and there are quite a lot of rows, for example, the client ID and client name table.
Test case:
Hive:
select sum(p. col3), count(p. col5), sum(d. col7), d. id%10000 from p join d on d. id=p. col7 group by d. id%10000
esProc: The codes can be divided into 3 parts. They are respectively: Program for summary machine, main program for node machine, and subprogram for node machine.
Impala:
select sum(p. col3), count(p. col5), sum(d. col7), d. id%10000 from p join d on d. id=p. col7 group by d. id%10000
Test results:
Hive
Impala
esProc
525s
269s
268s

Result description:
Let’s conclude the results of the four tests, and explain it one by one.
Grouping and Summarizing for Narrow Table
Test case
Hive
Impala
esProc
1 col. for grouping and 1 col. for summarizing
501s
256s
233s
1 col. for grouping and 4 col. for summarizing
508s
254s
237s
4 col. for grouping and 1 col. for summarizing
509s
253s
237s
4 col. for grouping and 4 col. for summarizing
536s
255s
237s
1.       esProc and Impala outperforms Hive obviously, almost 1 time or above.
2.       The performance of esProc is a bit stronger than Impala, but the superiority is not great.
3.       The column counts for grouping and summarizing do not have much impact on the performance of the three solutions.

Grouping and summarizing for wide table
Grouping col. * Summarizing col.
Hive
Impala
esProc
1 col. for grouping and 1 col. for summarizing
457s
272s
218s
1 col. for grouping and 4 col. for summarizing
458s
265s
218s
4 col. for grouping and 1 col. for summarizing
475s
266s
219s
4 col. for grouping and 4 col. for summarizing
488s
271s
218s
1.       esProc and Impala outperforms Hive obviously, almost 1 time or above.
2.       The performance of esProc is a bit stronger than Impala, but the superiority is not great.
3.       The column counts for grouping and summarizing do not have much impact on the performance of the three solutions.
4.       Compare with the data from narrow tables. You may find that the table columns make no difference on performance, while the volume of the whole table has direct impact on the performance. In addition, for the wide table, the performance of Impala will drop slightly, while the performance of Hive and esProc will increase a bit.

Associating computation on narrow tables
Hive
Impala
esProc
773s
262s
279s
1.       esProc and Impala outperform Hive obviously, almost 3 times better.
2.       The performance of Impala is slightly stronger than esProc, but the superiority is not great.

Associating computation on wide table
Hive
Impala
esProc
525s
269s
268s
1.       esProc and Impala outperform Hive greatly, almost 2 times higher.
2.       Impala performs slower than that of esProc by 1 second. Despite this slight difference, both of them can be regarded as performing equally good.
Interpretation and Analysis:
The performance of Hive is rather poor, which is easy to understand: as the infrastructure of Hive, MapReduce exchanges the data between computational nodes via files in external storage, so a great deal of time is spent on the hard disk IO. Impala and esProc offer the better performance because they exchange the intermediate result through memory directly. But, the performance of Impala is not as better than Hive for dozens of times as widely believed.

Exchanging data in the form of files do bring some benefits, which can actually ensure the reliability of intermediate result in the unstable environment of large cluster. esProc supports two ways to exchange the data (depend on programmer’s choice). Impala only supports the direct exchange, and Hive only supports the file exchange.

For grouping and summarizing, esProc performs better than Impala a bit. This is mainly because esProc enables the direct access to the local disk. By comparison, Impala must rely on HDFS to access to the hard disk. The process gets slow down naturally when there is a more layer of control.

However, in the associating computation, we may find that the data processing performances of esProc and Impala are contrary to that in grouping and summarizing. The performance of esProc is equal to or slightly stronger than Impala. It is probably because that the Impala implemented the technology of localizing the code generation. In CPU computing, its performance is slightly higher than esProc that executing codes by interpreting. So, although Impala relies on HDFS to access the hard disk, the high efficiency of CPU saves the time and situation. . As you can imagine, in grouping and summarizing, the time spent on hard disk access is much greater than CPU computing. While in the associating computation, the time spent on CPU computing gets greater, so that the Impala will overtake esProc. In addition, according to the analysis, it is not difficult to reach the conclusion that the workload ratio between the CPU computation and the hard disk access for narrow table operations is greater than that for wide table. The test data also tells that the advantage for Impala performance is much more obvious when handling the narrow table, which proves and verifies the above assumption from another perspective.

The column counts for grouping and summarizing do not have great impact on performance. This is because the syntax for this case is quite simple, and most time is spent on hard disk access but not the data computing. However, Hive and Impala are not the procedural languages like esProc. They cannot handle the complex computation and such idle CPU usage becomes common.

In addition, we limited the scope of computational results to a relatively small result set in the above tests. This is because Impala relies heavily on memory, and the big result set will cause the memory overflow. Hive only supports the external storage computation and there is no limitation on memory. Once modified, esProc algorithm can also implement the external storage computation. But the performance will be degraded.

No comments:

Post a Comment