An Oracle blog about Openomics

  • sun
    March 29, 2012

Talend Enterprise Data Integration overperforms on Oracle SPARC T4

The SPARC T microprocessor,
released in 2005 by Sun Microsystems, and now continued at Oracle,
has a good track record in parallel execution and multi-threaded performance. However it was less suited for pure single-threaded workloads. The new SPARC T4 processor is now filling that gap by
offering a 5x better single-thread performance over previous

Following our long-term
relationship with Talend, a fast growing ISV positioned by Gartner in
the “Visionaries” quadrant of the “Magic Quadrant for Data
Integration Tools”, we decided to test some of their integration
components with the T4 chip, more precisely on a T4-1 system, in
order to verify first hand if this new processor stands up to its

Several tests were performed,
mainly focused on:

  • Single-thread performance of

    the new SPARC T4 processor compared to an older SPARC

    T2+ processor
  • Overall throughput of the

    SPARC T4-1 server using multiple threads

The tests consisted in reading
large amounts of data --ten's of gigabytes--, processing and writing
them back to a file or an Oracle 11gR2 database table. They are CPU,
memory and IO bound tests. Given the main focus of this project --CPU
performance--, bottlenecks were removed as much as possible on the memory
and IO sub-systems. When possible, the data to process was put
into the ZFS filesystem cache, for instance. Also, two external storage devices
were directly attached to the servers under test, each one divided
in two ZFS pools for read and write operations.

Test Configuration

Multi-thread: Testing throughput on the Oracle

The tests were performed with
different number of simultaneous threads (1, 2, 4, 8, 12, 16, 32, 48
and 64) and using different storage devices: Flash, Fibre Channel
storage, two stripped internal disks and one single internal disk.
All storage devices used ZFS as filesystem and volume management.

Each thread read a dedicated
1GB-large file containing 12.5M lines with the following

1;Ronald;Reagan;South Highway;Santa Fe;Montana;98756;A;04-06-2006;09-08-2008
2;Theodore;Roosevelt;Timberlane Drive;Columbus;Louisiana;75677;A;10-05-2009;27-05-2008
3;Andrew;Madison;S Rustle St;Santa Fe;Arkansas;75677;A;29-04-2005;09-02-2008
4;Dwight;Adams;South Roosevelt Drive;Baton Rouge;Vermont;75677;A;15-02-2004;26-01-2007

The following graphs present the
results of our tests:

Results 1

Unsurprisingly up to 16 threads,
all files fit in the ZFS cache a.k.a L2ARC : once the cache is hot
there is no performance difference depending on the underlying
storage. From 16 threads upwards however, it is clear that IO becomes
a bottleneck, having a good IO subsystem is thus key. Single-disk performance collapses whereas the Sun F5100 and ST6180 arrays allow the
T4-1 to scale quite seamlessly. From 32 to 64 threads, the
performance is almost constant with just a slow decline.

For the database load tests, only
the best IO configuration --using external storage devices-- were
used, hosting the Oracle table spaces and redo log files.

Results 2

Using the Sun Storage F5100 array allows the T4-1 server to scale up to 48 parallel JVM
processes before saturating the CPU. The final result is a
staggering 646K lines per second insertion in an Oracle table using
48 parallel threads.

Single-thread: Testing the single thread

Seven different tests were
performed on both servers. Given the fact that only one thread, thus
one file was read, no IO bottleneck was involved, all data being
served from the ZFS cache.

  • Read File → Filter → Write File: Read file, filter data, write the filtered data in a new file.

    The filter is set on the “Status” column: only lines with status

    set to “A” are selected. This limits each output file to about

    500 MB.

  • Read File → Load Database Table: Read file, insert into a single Oracle table.

  • Average: Read file, compute the

    average of a numeric column, write the result in a new file.
  • Division & Square Root: Read file, perform a division and square root on a numeric column, write

    the result data in a new file.

  • Oracle DB Dump: Dump the content of an Oracle table (12.5M rows) into a CSV file.
  • Transform: Read file, transform,

    write the result data in a new file. The transformations applied

    are: set the address column to upper case and add an extra column at

    the end, which is the concatenation of two columns.

  • Sort: Read file, sort a numeric

    and alpha numeric column, write the result data in a new file.

The following table and graph
present the final results of the tests:

  • Throughput unit is thousand

    lines per second processed (K lines/second).
  • Improvement is the % of

    improvement between the T5140 and T4-1.



(Time s.)


(Time s.)

























Division & Square Root







DB Dump


















Results 3

The improvement of single-thread performance is quite dramatic:
depending on the tests, the T4 is between 5.4 to 7 times faster than
the T2+. It seems clear that the SPARC T4 processor has gone a long
way filling the gap in single-thread performance, without
sacrifying the multi-threaded capability as it still shows a very impressive scaling on heavy-duty multi-threaded jobs.

Finally, as always at Oracle ISV
Engineering, we are happy to help our ISV partners test their own
applications on our platforms, so don't hesitate to contact us and
let's see what the SPARC T4-based systems can do for your application!

"As describe in this benchmark, Talend Enterprise Data Integration has overperformed on T4. I was generally happy to see that the T4 gave scaling opportunities for many scenarios like complex aggregations. Row by row insertion in Oracle DB is faster with more than 650,000 rows per seconds without using any bulk Oracle capabilities !"

Cedric Carbone, Talend CTO.

Join the discussion

Comments ( 4 )
  • Jukka Friday, March 30, 2012

    Nice writeup. Would be interesting to see how one of the M-series servers performs for a comparison.

  • martin francis k Sunday, April 1, 2012

    would be more beneficial to compare performance of an Intel/AMD released in same time frame (similar generation) to T4. Comparing T4 to T2 only proves that T4 can outperform T2, but its more like saying my 2012 Accord has better mileage than 2006 Accord.

  • Amir Javanshir Monday, April 2, 2012

    Hi Martin

    I don't fully agree with you when you say that this only proves that the T4 outperforms the T2+ (or even a T3) as a natural fact. Obviously newer chips always tend to outperform older generations. However the main point here was to verify the fact that the T4 has a much better SINGLE thread performance than the previous generation: We are not talking about 30% or 50% improvement but at least 5x better, which is quite a breakthough.

    The other point was to check if the single thread performance evolution, did not break the primary force of the T chips: multi-threaded workload processing.

    The tests where therfore focused on those two points and I do believe that it answeres to both questions.

    However, it would be, as you said, interesting to test the performance of the T4 compared to an Intel/AMD chip but this was not the aim of this particular test.

  • martin francis k Monday, April 2, 2012

    I agree with your point. Given the purpose of the test, it very well proves the point.

    Its also interesting to know, FC drive scales similar to a flash array in sequential read.

Please enter your name.Please provide a valid email address.Please enter a comment.CAPTCHA challenge response provided was incorrect. Please try again.