User demands

With the increasing development of communication technology, the size and data volume of the communication system have all explosively grown. Statistical analysis and mining on the massive communication data has become a technical difficulty for the present industry.

The system has about 56 sheets, the required performance is 7GB/m, 40 million of records/m, in the system, the data recorded per minute on the largest sheet is about 8 millions of records, all these records need a storage capacity about 1800MB, the overall storage capacity of the system is about 900TB per year. The main contents of applications include data generation, data loading and data inquiry.

Solutions

The test system includes 4 data nodes and 4 service nodes. The 4 database nodes adopt Sugon A620 server, the 4 service nodes adopt HP servers, including 2 loaded service nodes and 2 inquiry service nodes, one loaded service node also serves as service manager, one inquiry node also serves as configuration server. Figure 2-1 shows the hardware topology of the test environment. All servers are interconnected through gigabit Ethernet.

Figure 2-1: Hardware Structure

Model of the 4 database nodes applied is Sugon A620 server, RAID controller was applied and the nodes were configurated with 8 (12 to the most) 2TB SATA disks.

Each database node configurations the 8 SATA disks into one Raid5, including one 120GB logical partition as the operating system partition, Cent OS 6.2 operating system was installed, the remained were data partitions. Divided one large partition from the data part and place it under the directory /mnt/raid. To achieve load balance of data reading and writing, the control files, log fils and data files of the database are all distributed on the disk array partition (the bottom layer is distributed to multiple disk devices).

Application effects

1. Loading effects

The customer program analyzes the data files and send that to the DRAC loading servers through the DRAC loading interface, such files will be separately written to the database by the loading server. The performance data about data loading test is listed below:

Implementing continuous data loading tests on the CDR-LU form, the stable aggregation writing performance of the entire system is around 267000 per second.

Due to the non-sharing structure, the aggregation loading performance of the DRAC system is therefore proportional to the quantity of database nodes. Data in each sheet is evenly distributed to all database nodes on the bottom layer in a way of Round-Robin, thus the balance between various node data may be achieved.

2. Inquiry performance

Inquiry adopts the DRAC sequence, a large quantity of tests will be implemented in 8 types of typical SQL sentences, including the filtration type, statistic type and associated inquiry, etc, inquiry tests are also conducted on the single customer end and multiple customer end. Conclusion drawn from inquiry tests is described below:

In case of single customer inquiry, the average inquiry response time is within 2 seconds.

In case of 10 customer inquiries, the average inquiry response time is slightly extended, to observe from the test results, most concurrent inquiries may be finished within 5 seconds.  The multiple-user concurrent inquiry performance has far exceeded the current data volume and inquiry demands of the application system.

For associated inquiries, DRAC data partition strategy may be used for optimization, thus most associated inquiries may be responded within 5 seconds. The inquiry results can meet the user’s performance demands on the associated inquiry business.