- 论坛徽章:
- 2
|
原帖由 flickxie 于 2006-12-30 21:40 发表
蜘蛛回复,荣幸之至啊.
忘了说版本了,是10的,具体好像是什么什么33的..记不清了,明天再去机房看看...机房上不了网..郁闷啊..
蜘蛛能不能解释一下这两句是什么意思??谢谢了.
修复ce网卡的bug的!
Document Audience: SPECTRUM
Document ID: 79189
Title: Sun[TM] Cluster 3.x: ce driver, kernel parameter settings
Update Date: Mon Jul 25 00:00:00 MDT 2005
Products: Clustering
Technical Areas: Clustering
--------------------------------------------------------------------------------
--------------------------------------------------------------------------------
Description: Top
The software driver for Sun[TM] GigaSwift Ethernet hardware - is also known as the Cassini ethernet(ce) driver. Sun[TM] Cluster software/hardware has some requirements for the driver's kernel parameter settings, when the device is used as cluster private transport.
Document Body: Top
Changing private transport ethernet driver parameters, can have great impact on performance. Planned changes, should always be tested in a laboratory or test environment first, before implementing in a production environment. Monitor the cluster closely, when rolling changes into production.
Under Sun Cluster software, version 3.0 only, it's required to put the following kernel parameter into the /etc/system file, for the ce driver, when a ce device is being used for public interfaces:
Edit the "/etc/system" file, inserting the following line:
set ce:ce_reclaim_pending=1
For all versions of Sun Cluster 3.x though, when ce interfaces are used for private transport, the following additional parameter settings are recommended in the /etc/system file:
set ce:ce_taskq_disable=1
set ce:ce_ring_size=1024
set ce:ce_comp_ring_size=4096
In situations where the system uses CE interfaces for both public and private (cluster interconnect) networks, parameters may be set on a per CE instance basis - i.e. on an individual basis. To do so, instead of modifying /etc/system, one might choose to modify the following files:
for Solaris 9 x86:
/kernel/drv/ce.conf
for SPARC:
/platform/sun4u/kernel/drv/ce.conf
The per instance method requires the "parent" and "unit-address" properties associated with the instance(s) being configured. These properties can be found by looking at the entries associated with the respective instance(s) within the /etc/path_to_inst file.
For example, the /etc/path_to_inst file may contain an entry associated with a ce0 interface as follows:
"/pci@1f,0/pci@1,1/network@1" 0 "ce"
----------------- - -
| | |
| | |
| | instance
| |
parent unit-address
Based on the above, the following entry may now be inserted into ce.conf
file to set the <ce_reclaim_pending> parameter:
...
name="ce" parent="/pci@1f,0/pci@1,1" unit-address = "1" ce_reclaim_pending = 1;
...
Some background on these parameters
ce_taskq_disable: The ce driver sends packets to the IP layer, in interrupt context. This can cause a problem on busy systems, and prevent other threads from running. This is very disruptive for clusters where we run all processes and threads with timeouts.
ce_ring_size: this parameter controls the number of 8K pages in the Receive buffer ring. This value must be a power of 2. The maximum value is 8192 buffers of 8K each.
ce_comp_ring_size: this parameter controls the number of packets that may be accepted in the hardware. (The size of each Receive completion descriptor ring. It also is power of 2.)
ce_ring_size and ce_comp_ring_size should be changed when ce_taskq_disable is set to 1 (which means taskq is disabled), to increase ring size to keep up with load.
ce_reclaim_pending: for better performance, it can be set as high as 32. When it is set to 1, it will reclaim for each packet processed, which means it will "count" every packet, instead of doing it in one shot when the reclaim threshold is reached.
Since Sun Cluster 3.0 software Public Network Management(PNM) relies on kernel network statistics to detect interface failure, it is important to keep statistical counts as up-to-date as possible. In Sun Cluster 3.1 software, and later, IPMP is used, which does not rely on kstat, so ce_reclaim_pending can be set high. |
|