tengyong
2003-02-20 11:29

谢谢二位的答复和建议,受益良多!

除了我们以上谈到的这些考虑的因素,还需要考虑什么?我们能不能做一些归纳呢?

tengyong
2003-02-20 15:24

Designing Your Application for Best Performance

原文出处:http://e-docs.bea.com/wls/docs70/jdbc/performance.html

Most performance gains or losses in a database application is not determined by the application language, but by how the application is designed. The number and location of clients, size and structure of DBMS tables and indexes, and the number and types of queries all affect application performance.

The following are general hints that apply to all DBMSs. It is also important to be familiar with the performance documentation of the specific DBMS that you use in your application.

1. Process as Much Data as Possible Inside the Database

Most serious performance problems in DBMS applications come from moving raw data around needlessly, whether it is across the network or just in and out of cache in the DBMS. A good method for minimizing this waste is to put your logic where the data is―in the DBMS, not in the client ―even if the client is running on the same box as the DBMS. In fact, for some DBMSs a fat client and a fat DBMS sharing one CPU is a performance disaster.

Most DBMSs provide stored procedures, an ideal tool for putting your logic where your data is. There is a significant difference in performance between a client that calls a stored procedure to update 10 rows, and another client that fetches those rows, alters them, and sends update statements to save the changes to the DBMS.

Also review the DBMS documentation on managing cache memory in the DBMS. Some DBMSs (Sybase, for example) provide the means to partition the virtual memory allotted to the DBMS, and to guarantee certain objects exclusive use of some fixed areas of cache. This means that an important table or index can be read once from disk and remain available to all clients without having to access the disk again.

2. Use Built-in DBMS Set-based Processing

SQL is a set processing language. DBMSs are designed from the ground up to do set-based processing. Accessing a database one row at a time is, without exception, slower than set-based processing and, on some DBMSs is poorly implemented. For example, it will always be faster to update each of four tables one at a time for all the 100 employees represented in the tables than to alter each table 100 times, once for each employee.

Many complicated processes that were originally thought too complex to do any other way but row-at-a-time have been rewritten using set-based processing, resulting in improved performance. For example, a major payroll application was converted from a huge slow COBOL application to four stored procedures running in series, and what took hours on a multi-CPU machine now takes fifteen minutes with many fewer resources used.

3. Make Your Queries Smart

Frequently customers ask how to tell how many rows will be coming back in a given result set. The only way to find out without fetching all the rows is by issuing the same query using the count keyword:

SELECT count(*) from myTable, yourTable where ...

This returns the number of rows the original query would have returned, assuming no change in relevant data. The actual count may change when the query is issued if other DBMS activity has occurred that alters the relevant data.

Be aware, however, that this is a resource-intensive operation. Depending on the original query, the DBMS may perform nearly as much work to count the rows as it will to send them.

Make your application queries as specific as possible about what data it actually wants. For example, tailor your application to select into temporary tables, returning only the count, and then sending a refined second query to return only a subset of the rows in the temporary table.

Learning to select only the data you really want at the client is crucial. Some applications ported from ISAM (a pre-relational database architecture) will unnecessarily send a query selecting all the rows in a table when only the first few rows are required. Some applications use a 'sort by' clause to get the rows they want to come back first. Database queries like this cause unnecessary degradation of performance.

Proper use of SQL can avoid these performance problems. For example, if you only want data about the top three earners on the payroll, the proper way to make this query is with a correlated subquery. Table 3-1 shows the entire table returned by the SQL statement

select * from payroll

Table 3-1 Full Results Returned

Name

Salary

Joe

10

Mikes

20

Sam

30

Tom

40

Jan

50

Ann

60

Sue

70

Hal

80

May

80

A correlated subquery

select p.name, p.salary from payroll p

where 3 >= (select count(*) from payroll pp

where pp.salary >= p.salary);

returns a much smaller result, shown in Table 3-2.

Table 3-2 Results from Subquery

Name

Salary

Sue

70

Hal

80

May

80

This query returns only three rows, with the name and salary of the top three earners. It scans through the payroll table, and for every row, it goes through the whole payroll table again in an inner loop to see how many salaries are higher than the current row of the outer scan. This may look complicated, but DBMSs are designed to use SQL efficiently for this type of operation.

4. Make Transactions Single-batch

Whenever possible, collect a set of data operations and submit an update transaction in one statement in the form:

BEGIN TRANSACTION

UPDATE TABLE1...

INSERT INTO TABLE2

DELETE TABLE3

COMMIT

This approach results in better performance than using separate statements and commits. Even with conditional logic and temporary tables in the batch, it is preferable because the DBMS obtains all the locks necessary on the various rows and tables, and uses and releases them in one step. Using separate statements and commits results in many more client-to-DBMS transmissions and holds the locks in the DBMS for much longer. These locks will block out other clients from accessing this data, and, depending on whether different updates can alter tables in different orders, may cause deadlocks.

Warning: If any individual statement in the preceding transaction fails, due, for instance, to violating a unique key constraint, you should put in conditional SQL logic to detect statement failure and to roll back the transaction rather than commit. If, in the preceding example, the insert failed, most DBMSs return an error message about the failed insert, but behave as if you got the message between the second and third statement, and decided to commit anyway! Microsoft SQL Server offers a connection option enabled by executing the SQL set xact_abort on, which automatically rolls back the transaction if any statement fails.

5. Never Have a DBMS Transaction Span User Input

If an application sends a 'BEGIN TRAN' and some SQL that locks rows or tables for an update, do not write your application so that it must wait on the user to press a key before committing the transaction. That user may go to lunch first and lock up a whole DBMS table until the user returns.

If you require user input to form or complete a transaction, use optimistic locking. Briefly, optimistic locking employs timestamps and triggers in queries and updates. Queries select data with timestamp values and prepare a transaction based on that data, without locking the data in a transaction.

When an update transaction is finally defined by the user input, it is sent as a single submission that includes timestamped safeguards to make sure the data is the same as originally fetched. A successful transaction automatically updates the relevant timestamps for changed data. If an interceding update from another client has altered data on which the current transaction is based, the timestamps change, and the current transaction is rejected. Most of the time, no relevant data has been changed so transactions usually succeed. When a transaction fails, the application can refetch the updated data to present to the user to reform the transaction if desired.

6. Use In-place Updates

Changing a data row in place is much faster than moving a row, which may be required if the update requires more space than the table design can accommodate. If you design your rows to have the space they need initially, updates will be faster, although the table may require more disk space. Because disk space is cheap, using a little more of it can be a worthwhile investment to improve performance.

7. Keep Operational Data Sets Small

Some applications store operational data in the same table as historical data. Over time and with accumulation of this historical data, all operational queries have to read through lots of useless (on a day-to-day basis) data to get to the more current data. Move non-current data to other tables and do joins to these tables for the rarer historical queries. If this can't be done, index and cluster your table so that the most frequently used data is logically and physically localized.

8. Use Pipelining and Parallelism

DBMSs are designed to work best when very busy with lots of different things to do. The worst way to use a DBMS is as dumb file storage for one big single-threaded application. If you can design your application and data to support lots of parallel processes working on easily distinguished subsets of the work, your application will be much faster. If there are multiple steps to processing, try to design your application so that subsequent steps can start working on the portion of data that any prior process has finished, instead of having to wait until the prior process is complete. This may not always be possible, but you can dramatically improve performance by designing your program with this in mind.

shyguy
2003-02-20 16:29

可以参考jive ent的配置啊。呵呵,据说每天可以应付百万点击。

我记忆中那边的人建议:

更高档机器:比如sun 3800/4500级别。

好的数据库设计,且数据库集群。

快速的jvm

快速的AppServer

好的网络负载平衡设备。

shyguy
2003-02-20 16:30

至于代码优化么。。。。嘿嘿,建议听banq大哥上课咯。。。。

alibaba2002
2003-02-20 18:12

请教一下weblogic专家:

在deploy时,遇到以下错误是什么原因。我用的是weblogic7+jbuilder8,编译的工程是jbuilder自带的例子employee.jds。

D:\bea\jdk131_03\bin\javaw -classpath D:\bea\weblogic700\server\lib\weblogic.jar;D:\JBuilder8\lib\jdsserver.jar;D:\bea\weblogic700\server\lib\weblogic_sp.jar;D:\bea\weblogic700\server\lib\webservices.jar; weblogic.Deployer -user system -adminurl http://huangqi:7001 -password weblogic -activate -name TestEntMod -upload -source D:/jbproject/TestEntProject/TestEntMod.jar -targets myserver

Operation started, waiting for notifications...

...

#TaskID Action Status Target Type Application Source

10 Activate Failed myserver Server TestEntMod D:\bea\user_projects\mydomain\myserver\upload\TestEntMod\TestEntMod.jar

Exception:weblogic.management.ApplicationException: activate failed forTestEntMod

Module: TestEntMod Error: Exception activating module: EJBModule(TestEntMod,status=PREPARED)

Unable to deploy EJB: Employee from TestEntMod.jar:

weblogic.ejb20.WLDeploymentException: The DataSource with the JNDI name: DataSource could not be located. Please ensure that the DataSource has been deployed successfully and that the JNDI name in your EJB Deployment descriptor is correct.

at weblogic.ejb20.cmp.rdbms.RDBMSPersistenceManager.setup(RDBMSPersistenceManager.java:130)

at weblogic.ejb20.manager.BaseEntityManager.setupPM(BaseEntityManager.java:214)

at weblogic.ejb20.manager.BaseEntityManager.setup(BaseEntityManager.java:186)

at weblogic.ejb20.manager.DBManager.setup(DBManager.java:161)

at weblogic.ejb20.deployer.ClientDrivenBeanInfoImpl.activate(ClientDrivenBeanInfoImpl.java:936)

at weblogic.ejb20.deployer.EJBDeployer.activate(EJBDeployer.java:1302)

at weblogic.ejb20.deployer.EJBModule.activate(EJBModule.java:342)

at weblogic.j2ee.J2EEApplicationContainer.activateModule(J2EEApplicationContainer.java:1534)

at weblogic.j2ee.J2EEApplicationContainer.activate(J2EEApplicationContainer.java:991)

at weblogic.j2ee.J2EEApplicationContainer.activate(J2EEApplicationContainer.java:978)

at weblogic.management.deploy.slave.SlaveDeployer.processPrepareTask(SlaveDeployer.java:1104)

at weblogic.management.deploy.slave.SlaveDeployer.prepareUpdate(SlaveDeployer.java:724)

at weblogic.drs.internal.SlaveCallbackHandler$1.execute(SlaveCallbackHandler.java:24)

at weblogic.kernel.ExecuteThread.execute(ExecuteThread.java:152)

at weblogic.kernel.ExecuteThread.run(ExecuteThread.java:133)

8Go 上一页 1 2 3 4 5 6 7 ... 8 下一页