关于Oracle11g数据库的redo并行机制

Yong Huang

> 由于组提交的原因，比如某一session执行了commit操作的话，那么其他session所属的private strand中的redo信息也都要随之一起被写入redo 文件中吧？

Are you sure when one session commits, another session's private redo will also be flushed and written to the logfile? I just did a simple test. It's Oracle 11.2.0.1, not RAC, no supplemental logging, no flashback database:

Session 2 records session 1's redo size statistic, 0. Session 1 makes a small (< 128kb on 64-bit Oracle) transaction (insert a row into a one number column table). Session 2 says session 1's redo size remains 0 (the redo is in private redo, not visible in v$sesstat). Session 3 makes a transaction and commits. Session 2 reports session 1's redo size is still 0. It changes to non-zero when it commits (or rolls back) its own transaction.

If my session 1 had flushed its private redo to public redo when session 3 committed, session 1's redo size statistic would be immediately shown in v$sesstat.

Your second message is about reduced redo allocation latch gets because of private redo. I'm sure that's true. I don't know the details. Think of some good tests and share with us.

Yong Huang

> Redo entries in private strands are not directly written to the current logfile.
> They're copied to the public redo log buffer first, and then immediately written
> ...

I need to study a little more about that. There's a concept of "zero copy redo". Tanel Poder has an old message at
http://www.freelists.org/post/oracle-l/latchfree-SCN-scheme-10103,5
and Stephan Haisley has
http://www.teamycc.com/RMOUG_2008_Conference/PDF/Haisley.pdf (see "Improvement #3.5")

It's possible there's no actual copying operation from private to public redo buffer. It may be just some virtual memory address (pointer) manipulation.

[2012-08-24 update] If the parameter _log_private_parallelism ("Number of private log buffer strands for zero-copy redo") only controls whether to enable *zero-copy redo* of private strands, not whether to enable private strands, then zero-copy redo is disabled by default. You have to set _log_private_parallelism to true to enable it. It's a parameter in Oracle 10.1 to 11.1. I don't know if there's a replacement for it in 11.2.

ccsnmoracle

Yong Huang 发表于 2012-8-24 12:24
> 由于组提交的原因，比如某一session执行了commit操作的话，那么其他session所属的private strand中的redo ...

>Session 2 records session 1's redo size statistic, 0. Session 1 makes a small (< 128kb on 64->bit Oracle) transaction (insert a row into a one number column table). Session 2 says session >1's redo size remains 0 (the redo is in private redo, not visible in v$sesstat). Session 3 makes >a transaction and commits. Session 2 reports session 1's redo size is still 0. It changes to non->zero when it commits (or rolls back) its own transaction.

Yong Huang sir,我会按照你的这个实验思路，在我的测试环境了试一试！

sundog315

Yong Huang 发表于 2012-8-24 12:51
> Redo entries in private strands are not directly written to the current logfile.
> They're copie ...

就我的理解zero copy redo是在内存之间进行了映射，从log_buffer中空出一部分空间，将private stand映射过来，当lgwr写出时，逻辑上还是从log_buffer读，只是有一部分内容其实是从private stand映射过来的。

这个理解对么？

Yong Huang

> 就我的理解zero copy redo是在内存之间进行了映射，从log_buffer中空出一部分空间，将private stand映射过来，当lgwr写出时，
> 逻辑上还是从log_buffer读，只是有一部分内容其实是从private stand映射过来的。

Your understanding is exactly mine! I'm still not sure whether zero copy redo has to be enabled separately. That is, it may or may not be enabled even if private redo is already enabled.