引题：受人之托，写了这篇文章，貌似这个问题也困扰了许多人，希望可以给大家一点启发
Test
1. 创建一个用户leonarding并授予dba权限
SYS@LEO> create user leonarding identified by leonarding default tablespace users;
SYS@LEO> grant dba to leonarding;
2. 创建一个t表，只要结构信息
SYS@LEO> conn leonarding/leonarding
LEONARDING@LEO> create table t as select * from all_objects where 1=0;
Table created.
3. 启动执行计划，查看统计报告
LEONARDING@LEO> set autotrace on;
LEONARDING@LEO> select * from t;
no rows selected
Execution Plan 执行计划
----------------------------------------------------------
Plan hash value: 1601196873
--------------------------------------------------------------------------
| Id | Operation | Name | Rows | Bytes | Cost (%CPU)| Time |
--------------------------------------------------------------------------
| 0 | SELECT STATEMENT | | 1 | 128 | 2 (0)| 00:00:01 |
| 1 | TABLE ACCESS FULL| T | 1 | 128 | 2 (0)| 00:00:01 | 全表扫描
--------------------------------------------------------------------------
Note
-----
- dynamic sampling used for this statement
Statistics 统计报告
----------------------------------------------------------
264 recursive calls
0 db block gets
30 consistent gets 一致性读，I/O数量
0 physical reads
0 redo size
995 bytes sent via SQL*Net to client
370 bytes received via SQL*Net from client
1 SQL*Net roundtrips to/from client
0 sorts (memory)
0 sorts (disk)
0 rows processed
LEONARDING@LEO>
4. 关闭执行计划
LEONARDING@LEO> set autotrace off;
5. 向表t插入记录但不提交
LEONARDING@LEO> insert into t select * from all_objects; 已经插入9681row
9681 rows created.
LEONARDING@LEO> select count(*) from t; 此时表中已经有9681行数据了
COUNT(*)
----------------------------
9681
6. Rollback回滚操作
LEONARDING@LEO> rollback;
Rollback complete.
LEONARDING@LEO> select count(*) from t; 此时表中为0行
COUNT(*)
------------------
0
7. 第二次查询表t统计报告
LEONARDING@LEO> set autotrace traceonly statistics;
LEONARDING@LEO> select * from t;
no rows selected
Statistics
----------------------------------------------------------
0 recursive calls
0 db block gets
141 consistent gets 一致性读，I/O数量明显上升
0 physical reads
0 redo size
995 bytes sent via SQL*Net to client
370 bytes received via SQL*Net from client
1 SQL*Net roundtrips to/from client
0 sorts (memory)
0 sorts (disk)
0 rows processed
说明：
我们可以看出第二次一致性读I/O数量明显高过了第一次I/O数量，我说明一下流程大家就都明白了，当我们insert插入大量数据的时候oracle已经预先分配块空间，此时HWM已经上浮（注：我们检索HWM以下的块）。虽然我们又做了rollback操作，但rollback只是回滚undo快照并没有重设HWM功能，那么块空间还是存在的只不过为空。select扫描HWM以下的块，当然也包括了这些空块，增加了一致性读I/O次数，所以一致性读I/O从30上升到141，多出来的111就是扫描这些空块。下面我们来看看truncate操作对HWM影响
8. truncate操作对HWM影响
LEONARDING@LEO> insert into t select * from all_objects;
9681 rows created.
Statistics
----------------------------------------------------------
8 recursive calls
923 db block gets
16277 consistent gets 插入时就已经分配了块空间，HWM上浮
0 physical reads
976484 redo size
674 bytes sent via SQL*Net to client
571 bytes received via SQL*Net from client
4 SQL*Net roundtrips to/from client
1 sorts (memory)
0 sorts (disk)
9681 rows processed
LEONARDING@LEO> commit;
Commit complete.
LEONARDING@LEO> truncate table t;
Table truncated.
LEONARDING@LEO> select * from t;
no rows selected
Statistics
----------------------------------------------------------
1 recursive calls
1 db block gets
6 consistent gets 明显降低了一致性读I/O数量，说明truncate操作能够有效降低HWM并且不会生成undo信息
0 physical reads
96 redo size
995 bytes sent via SQL*Net to client
370 bytes received via SQL*Net from client
1 SQL*Net roundtrips to/from client
0 sorts (memory)
0 sorts (disk)
0 rows processed
小结：当我们在删除一个表中所有记录时，如果想有效降低I/O开销，那么可以直接使用truncate方式，但注意此方式不能生成undo信息恢复成本较高谨慎操作。
oracle视频教程请关注:http://u.youku.com/user_video/id_UMzAzMjkxMjE2.html