请教参数db_file_multiblock_read_count的作用

Yong Huang

原帖由 king3171 于 2009-7-28 21:58 发表
http://www-900.ibm.com/cn/suppor ... DocId=2411083L10001

Thanks for this update. I looked at the IBM document, then searched Metalink for "socket buffer overflows" and found many bugs. But all are incomplete or duplicate bugs. I think if there's really a bug acknowledged by Oracle, it most likely is port specific (limited to AIX only). We run Oracle 10gR2 on Linux (Red Hat 4 and 5.3, with kernel 2.6). I guess I'm lucky.

On my Linux boxes, netstat -s does not have an entry for "socket buffer overflows". The closest is "packets pruned from receive queue because of socket buffer overrun", but it's under the heading of TcpExt. Under Udp, I only have

Udp:
    3212588471 packets received
    2753777 packets to unknown port received.
    4070 packet receive errors
    3129298945 packets sent

I'm not sure if Linux simply does not record the socket buffer overrun or overflow statistics.

Yong Huang

[ 本帖最后由 Yong Huang 于 2009-7-29 13:46 编辑 ]

火星鼠

学习了

king3171

原帖由 Yong Huang 于 2009-7-30 03:40 发表


Thanks for this update. I looked at the IBM document, then searched Metalink for "socket buffer overflows" and found many bugs. But all are incomplete or duplicate bugs. I think if there's really a bug acknowledged by Oracle, it most likely is port specific (limited to AIX only). We run Oracle 10gR2 on Linux (Red Hat 4 and 5.3, with kernel 2.6). I guess I'm lucky.

On my Linux boxes, netstat -s does not have an entry for "socket buffer overflows". The closest is "packets pruned from receive queue because of socket buffer overrun", but it's under the heading of TcpExt. Under Udp, I only have

Udp:
    3212588471 packets received
    2753777 packets to unknown port received.
    4070 packet receive errors
    3129298945 packets sent

I'm not sure if Linux simply does not record the socket buffer overrun or overflow statistics.

Yong Huang

      确实这个问题大多出现在AIX系统上，这和RAC的内部通讯协议有关，AIX和HPUX平台上是使用UDP，所以RAC环境下在AIX和HPUX平台上都可以看到UDP 的socket buffer overflows，而solaris和linux印象中使用的不是UDP，所以看不到UDP 的socket buffer overflows，具体使用什么协议，有兴趣可以去查证。当我的9I RAC FOR AIX遇到 High 'global cache cr request' wait events in a statspack report.时，我在ITPUB和GOOGLE上搜索到很多同样问题的帖子，都是在 AIX系统上发生的，但都没有给出解决方案，有些人甚至无奈的直接关掉RAC中一个实例单实例来跑业务，这样就可以避免缓存融合，我也是无奈的情况下摸索了很久才发现是UDP通信 的socket buffer overflows导致的，最后引申到db_file_multiblock_read_count参数。

[ 本帖最后由 king3171 于 2009-7-31 14:19 编辑 ]

Yong Huang

> solaris和linux印象中使用的不是UDP，所以看不到UDP 的socket buffer overflows

On all OS'es I know, RAC interconnect uses UDP by default. Just login as sys as sysdba, run "oradebug ipc" and check trace file.

Besides, "socket buffer overflows" is an OS level statistic. It exists or does not exist no matter whether you have installed Oracle.

> 我也是无奈的情况下摸索了很久才发现是UDP通信 的socket buffer overflows导致的，最后引申到
> db_file_multiblock_read_count参数。

I have more confidence now to say that the problem of high db_file_multiblock_read_count causing high 'global cache cr request' is limited to AIX. I don't see more evidence from your research that indicates it's not port-specific.

Yong Huang

wangkxxe

学习啊，看的有点累来

kewin

On one of linux platform: db_file_multiblock_read_count default value is 16:
SQL> show parameter multi

NAME                                 TYPE                              VALUE
------------------------------------ --------------------------------- ------------------------------
db_file_multiblock_read_count        integer                           16

kewin

to Yong Huang
And on linix rac env:
What value for db_file_multiblock_read_count is fit?

Yong Huang

原帖由 kewin 于 2009-8-2 23:31 发表
to Yong Huang
And on linix rac env:
What value for db_file_multiblock_read_count is fit?

Always tell us Oracle version.

In 10gR2, leave it alone.

Yong Huang