精通polling参数调优的进来帮帮忙吧

antijp

Before you modify the value, use sysctl -a |grep mbuf to check the mbuf utilization. According to TCP/IP Illustrated Vol.2, bigger packets uses not only mbuf, but also mbufclusters.

antijp

原帖由 雨丝风片 于 2006-5-25 08:53 发表



改这个里面的：
/usr/src/sys/sys/param.h

I think modify the source file is finding trouble.
Simply recompile the kernel with this

1. 
   
2. options MSIZE=512

复制代码

in your kernel configuration file.

antijp

原帖由 雨丝风片 于 2006-5-25 09:08 发表


楼主测试的数据包大小是确定的，改这个值就是想尽量避免使用cluster，看看从对纯mbuf的处理和对cluster的处理流程上是否能找到一些线索。但我感觉应该不只是处理流程的区别，很可能和内存的分配、缓存、复用也 ...

According to the statistics posted out, the mbufclusters must have been used. But I doubt that this statistics was not gathered when running the benchmark, so perhaps it's not very useful.

[ 本帖最后由 antijp 于 2006-5-25 09:46 编辑 ]

antijp

原帖由 雨丝风片 于 2006-5-25 09:47 发表


从楼主的数据长度即可判断出测试的时候肯定会使用cluster了：

当然，由于mbuf头结构的修改，目前的门限数值比208还要小。现在就是想先看看它不用cluster的时候的测试性能如何。

Less than 208? I've traced the code. sizeof(m_hdr) == 24, sizeof(pkthdr) == 24, without m_ext, the maximum size of the data can be stored is 208. These number is when you are running on i386.
I'm running amd64, so the calculation may be incorrect, and under amd64 sizeof(m_hdr)==sizeof(pkthdr)==40, it's so depressing...:em12:

antijp

Now I'm building 6.1 on another i386 machine, I'll check the size later.

antijp

I've checked the two sizes, sizeof(pkthdr) == sizeof(m_hdr) == 24, and sizeof(m_ext) also equals to 24.
MINCLSIZE is 209. This design removes a indirect pointer and would make performance better.

antijp

对了，楼主试试看把kern.polling.idle_poll置1看看，另外同时最好不要运行其他的应用