Thursday Mar 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 generations.

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 promises.

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 T4-1

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 structure:

customerID;FirstName;LastName;StreetAddress;City;State;Zip;Cust_Status;Since_DT;Status_DT
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 performance

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.

Test

T4-1 (Time s.)

T5140 (Time s.)

Improvement

T4-1 (Throughput)

T5140 (Throughput)

Read/Filter/Write

125

806

645%

100

16

Read/Load Database

195

1111

570%

64

11

Average

96

557

580%

130

22

Division & Square Root

161

1054

655%

78

12

Oracle DB Dump

164

945

576%

76

13

Transform

159

1124

707%

79

11

Sort

251

1336

532%

50

9

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.

Monday Mar 12, 2012

Mobile Tornado adopts Solaris features for better RAS and TCO

Mobile Tornado provides instant communication services for mobile devices, with a focus on enterprise workforce management. Its solutions include Push-To-Talk and Instant Locate, Alert & Message applications.

As a software developer, Mobile Tornado's main challenges are up-time --the applications are largely sold today into the homeland security and defense markets-- and scalability --during network peak usage. With these challenges in mind, and as part of the on-going engineering collaboration between Mobile Tornado and Oracle's ISV Engineering, we investigated which Oracle Solaris technologies would improve the application's availability and scalability while reducing the solution's TCO.

We looked at the following Oracle Solaris technologies: Solaris Cluster, ZFS and Zones.

[Read More]

Tuesday Jan 17, 2012

ZFS secures your application

One of our ISV partner, a leading vendor in the financial services space, commonly recommends as deployment platform a commodity 2-socket Lintel server to minimize cost, but equipped with an internal RAID storage controller to increase the application uptime on such entry-level servers by mirroring the root disk. We recently worked with its professional services team to explore if we could improve on the solution, i.e. continuing to bring the cost down while increasing uptime.

The proposed solution was to use the Oracle Solaris 10 operating system and its ZFS file system in lieu of the hardware RAID to mirror the root hard drive. ZFS is a new kind of filesystem that provides simple administration, transactional semantics and immense scalability. ZFS natively supports all common RAID functionalities and also embeds advanced fonctionnalities in compression, encryption and snapshot, typically expected from proprietary high-end storage systems.

The benefits of the ZFS-based solution is …

[Read More]

Monday Jan 09, 2012

Infovista VistaInsight for Networks shows 3.7x performance on Oracle

System management vendor InfoVista markets the VistaInsight for Networks® application to enable telco operators, service providers and large enterprises effectively meet performance and service level agreements of converged and next-generation communication networks. As part of our on-going technology partnership, InfoVista and Oracle ISV Engineering together ran a performance test campaign of VistaInsight for Networks® over Oracle Solaris and Sun CMT hardware. The two companies shared many common objectives when starting this project.

The most obvious was to improve the scalability and performance of VistaInsight for Networks® over Oracle's SPARC T-Series systems and thereby provide customers with a better price/performance ratio and a better ROI. From the onset, virtualization was considered a promising technology to improve scalability, thus testing VistaInsight for Networks® in the context of Oracle Solaris Zones was also a major milestone.

Second, InfoVista was interested in setting new limits in terms of the workload that its application can sustain, in response to the evolving needs of its customers.

Lastly, as the first improvements on computing scalability were delivered, it became obvious that the storage was the next critical component for the performance of the entire solution. A decision was then made to test the Oracle Solaris ZFS file system, the Sun ZFS Storage Appliance, and the SSD technology from Oracle to move to the next level of performance.

The result of this performance test campaign is a new Reference Architecture whitepaper that provides detailed information about the configuration tested, the tests executed and the results obtained. It clearly shows that VistaInsight for Networks® takes full advantage of the server, storage and virtualization technology provided by Oracle. By leveraging the Oracle Solaris Zones, Oracle Solaris ZFS, SSD and Sun ZFS Appliance storage, Infovista increased the throughput performance by more than 370%, meeting the highest expectations in terms of workload and performance while maintaining the cost in a very attractive range.

Learn all the details about this new Reference Architecture published on OTN.

Monday Nov 14, 2011

Latency Matters

A lot of interest in low latencies has been expressed within the financial services segment, most especially in the stock trading applications where every millisecond directly influences the profitability of the trader. These days, much of the trading is executed by software applications which are trained to respond to each other almost instantaneously. In fact, you could say that we are in an arms race where traders are using any and all options to cut down on the delay in executing transactions, even by moving physically closer to the trading venue.

The Solaris OS network stack has traditionally been engineered for high throughput, at the expense of higher latencies. Knowledge of tuning parameters to redress the imbalance is critical for applications that are latency sensitive. We are presenting in this blog how to configure further a default Oracle Solaris 10 installation to reduce network latency.

[Read More]
About

How open innovation and technology adoption translates to business value, with stories from our developer support work at Oracle's ISV Engineering.

Subscribe

Search

Categories
Archives
« April 2014
SunMonTueWedThuFriSat
  
1
2
3
5
6
8
9
10
11
12
13
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
   
       
Today
Feeds