Dataguard_network_tuning_tvd_layout

盖国强

381

8页

1次

2019-06-21

5墨值下载

info@trivadis.com . www.trivadis.com . Info-Tel. 0800 87 482 347 . Datum 29.04.2009 . Seite 1 / 8

Data Guard in SYNC mode – does my DB suffer

from network weakness ?

Yann Neuhaus – Senior Consultant – 16.oct.2007

Reviewed Yann Neuhaus – Principal Consultant – 28.apr.2009

(issues with v$system_event metrics and improved interpretation)

Data Guard network tuning is always a big and nice challenge which should be taken seriously.

For instance, my Grid Control console is often polluted by warnings about wait events of the

network event class :

Is my Data Guard setup really suffering from poor performance?

Before any tuning action each Data Guard administrator should know and read both white papers

listed in the metalink note 387174.1 For Oracle 10.2 this note points to:

MAA_WP_10gR2_DataGuardNetworkBestPractices.pdf

Without dwelling on the details of this Oracle document, find below the main settings required

for an ideal network configuration for Data Guard:

First of all calculate the Bandwidth Delay Product (BDP). For this purpose try to identify the

Round Trip Time using ping:

ping ttchp02

PING ttchp02.ttc.trivadis.com: (172.16.65.7): 56 data bytes

64 bytes from 172.16.65.7: icmp_seq=0 ttl=255 time=0 ms

64 bytes from 172.16.65.7: icmp_seq=1 ttl=255 time=0 ms

64 bytes from 172.16.65.7: icmp_seq=2 ttl=255 time=0 ms

On AIX, the ping command shows that the elapsed time is equal to 0 ms, therefore the estimated

RTT is 1 ms (we take the pessimistic approach).

info@trivadis.com . www.trivadis.com . Info-Tel. 0800 87 482 347 . Datum 29.04.2009 . Seite 2 / 8

On a Linux machine fractions of milliseconds could be displayed:

ping 172.17.38.8

PING 172.17.38.8 (172.17.38.8) 56(84) bytes of data.

64 bytes from 172.17.38.8: icmp_seq=4 ttl=64 time=0.339 ms

64 bytes from 172.17.38.8: icmp_seq=5 ttl=64 time=0.424 ms

RTT * Bandwidth =

0.001 (1 ms) * 1'000'000'000 (1Gb/s) / 8 (to get bytes) =

125'000

It is recommended to set the OracleNet TCP parameters to three times the BDP: 375'000

In sqlnet.ora on the primary and standby server :

RECV_BUF_SIZE=375000

SEND_BUF_SIZE=375000

Make also sure that the parameter TCP.NODELAY (per default set to yes) has not been explicitly

set to “no”. Either do not set the parameter or set it to “yes” :

TCP.NODELAY=yes

This parameter will have the effect that each change vector(s) package will be directly transmitted

to the standby site without waiting for another packet. The Naggle algorithm is therefore disabled.

The goal is to privilege the primary database performance even if the network utilization won’t be

optimal.

More info about the Naggle algorithm and OracleNet (Although this is an old publication the

fundaments are correctly explained) :

http://www.miracleas.dk/tools/Miracle_4_nagle.pdf

On AIX, the platform used during these tests some tuning are requested.

Set the TCP queue length :

no -o ipqmaxlen=100

Set the send and receive space as documented under :

http://publib16.boulder.ibm.com/pseries/en_US/aixbman/prftungd/netperf3.htm#i79137

tcp_sendspace = 131072

tcp_recvspace = 65536

Once these best practices have been configured we can investigate why the Grid Control

warnings still occur. The database is up since about 3 days (72 hours) :

SYS@STAGPRD1 > select sysdate from dual;

SYSDATE

--------------------

25-SEP-2007 16:07:15

SYS@STAGPRD1 > select startup_time from v$instance;

STARTUP_TIME

--------------------

22-SEP-2007 19:28:40

First of all try to identify the wait classes which generated the most important wait times:

set lines 130

select * from V$SYSTEM_WAIT_CLASS

info@trivadis.com . www.trivadis.com . Info-Tel. 0800 87 482 347 . Datum 29.04.2009 . Seite 2 / 8

On a Linux machine fractions of milliseconds could be displayed:

ping 172.17.38.8

PING 172.17.38.8 (172.17.38.8) 56(84) bytes of data.

64 bytes from 172.17.38.8: icmp_seq=4 ttl=64 time=0.339 ms

64 bytes from 172.17.38.8: icmp_seq=5 ttl=64 time=0.424 ms

RTT * Bandwidth =

0.001 (1 ms) * 1'000'000'000 (1Gb/s) / 8 (to get bytes) =

125'000

It is recommended to set the OracleNet TCP parameters to three times the BDP: 375'000

In sqlnet.ora on the primary and standby server :

RECV_BUF_SIZE=375000

SEND_BUF_SIZE=375000

Make also sure that the parameter TCP.NODELAY (per default set to yes) has not been explicitly

set to “no”. Either do not set the parameter or set it to “yes” :

TCP.NODELAY=yes

This parameter will have the effect that each change vector(s) package will be directly transmitted

to the standby site without waiting for another packet. The Naggle algorithm is therefore disabled.

The goal is to privilege the primary database performance even if the network utilization won’t be

optimal.

More info about the Naggle algorithm and OracleNet (Although this is an old publication the

fundaments are correctly explained) :

http://www.miracleas.dk/tools/Miracle_4_nagle.pdf

On AIX, the platform used during these tests some tuning are requested.

Set the TCP queue length :

no -o ipqmaxlen=100

Set the send and receive space as documented under :

http://publib16.boulder.ibm.com/pseries/en_US/aixbman/prftungd/netperf3.htm#i79137

tcp_sendspace = 131072

tcp_recvspace = 65536

Once these best practices have been configured we can investigate why the Grid Control

warnings still occur. The database is up since about 3 days (72 hours) :

SYS@STAGPRD1 > select sysdate from dual;

SYSDATE

--------------------

25-SEP-2007 16:07:15

SYS@STAGPRD1 > select startup_time from v$instance;

STARTUP_TIME

--------------------

22-SEP-2007 19:28:40

First of all try to identify the wait classes which generated the most important wait times:

set lines 130

select * from V$SYSTEM_WAIT_CLASS

of 8

5墨值下载

关注

评论