IO架构图：

内存管理图：

很容易误解的socket buffer：

以下是从man 2 connnect摘抄的:

EINPROGRESS
The socket(2,7,n) is non-blocking and the connection cannot be com-
pleted immediately. It is possible to select(2,7,2 select_tut)(2) or poll(2) for
completion by selecting the socket(2,7,n) for writing. After select(2,7,2 select_tut)
indicates writability, use getsockopt(2) to read(2,n,1 builtins) the SO_ERROR
option at level SOL_SOCKET to determine whether connect com-
pleted successfully (SO_ERROR is zero) or unsuccessfully
(SO_ERROR is one of the usual error(8,n) codes listed here, explain-
ing the reason for the failure).

我们看下erlang的inet_drv是如何处理的.
./erts/emulator/drivers/common/inet_drv.c:

        {
            int error = 0;      /* Has to be initiated, we check it */
            unsigned int sz = sizeof(error); /* even if we get -1 */
            int code = sock_getopt(desc->inet.s, SOL_SOCKET, SO_ERROR,
                                   (void *)&error, &sz);

            if ((code < 0) || error) {
                desc->inet.state = TCP_STATE_BOUND;  /* restore state */
                ret = async_error(INETP(desc), error);
                goto done;
            }
        }

结论: 细节决定品质.

演示如下：
我们通过修改过
extern int sysctl_tcp_fin_timeout;的值来达到目的。是因为这个值是proc导出的 我们好验证是否成功。

root@localhost ~]# cat /proc/sys/net/ipv4/tcp_fin_timeout
15000
[root@localhost ~]# cat test.stp
probe begin
{
        printf("ready go\n");
}

probe kernel.function("do_tcp_setsockopt")
{
        $sysctl_tcp_fin_timeout = $1
        printf("sysctl_tcp_fin_timeout = %d\n", $sysctl_tcp_fin_timeout);
        exit()
}

[root@localhost ~]# stap -g test.stp 18000
ready go

这个时候 stap在运行， 只是还没有触发do_tcp_setsockopt.
现在我们来触发

[root@localhost ~]# erl
Erlang R13B02 (erts-5.7.3) [source] [64-bit] [smp:2:2] [rq:2] [async-threads:0] [hipe] [kernel-poll:false]

Eshell V5.7.3  (abort with ^G)
1> {ok, LSock} = gen_tcp:listen(0, []).
{ok,#Port<0.437>}
2>
2> inet:setopts(LSock, [{nodelay,true}]).
ok
3>

Ok,这时候回头可以看到stap打出来以下：
sysctl_tcp_fin_timeout = 18000

我们来验证下：

root@localhost ~]# cat /proc/sys/net/ipv4/tcp_fin_timeout
18000

OK,成功。

Tips：
1. stap对全局变量的写需要-g guru模式。
2. 全局变量必须在一个单元内的函数里面才可以修改， 而且必须是在内核上下文。

好不容易从32位系统中逃脱，解决了4G内存的问题，而且64位的cpu更多的寄存器，可以带来更好的性能。但是怎么又碰到问题了。64位的系统，64位的指针，意味着更多的数据访问， cpu速度是提高了，但是内存的带宽和访问速度没有大的提高。而且内存的访问速度和cpu的cycle差几个数量级别，所以对于普通的网络程序来讲，大部分是处理信息的变形，也就是说是把内存里面的数据从一个形式变成另外一个形式。 cpu性能的提高对这种程序来讲 影响不是很大。倒是因为数据的量加大了1倍， 导致整体的性能降低了百分几十。实际的硬件系统也没有那么多内存，一般都是16G以下，所以才有人去想在64位系统下，用36位的指针，减少内存的访问。

erlang的系统基本上是个网络程序，所以这个问题就非常突出。 目前没看到otp vm的这方面的打算， 我还是乖乖的用我的32位系统

我这里要说的是他们的工作原理。 这些模块的核心都是根据erlang的trace机制实现的。在模块执行的时候，trace机制会通知那个函数被调用 返回。根据这些信息就可以统计出来函数调用的频度，调用栈等。

但是利用这个机制会有严重的性能损失。因为每个函数调用都要发送一条trace信息，每个trace 信息会引起上下文切换 而且要耗费2-3的时间。这个对大型的系统是不可接受的。

所以知道这些原理以后， 我们在profile大型的系统的时候，我们可以在dbg模块的帮助下， 只收集我们感兴趣的东西，而且严格限定范围，避免对系统造成大的干扰，这样收集出来的东西才有意义。

我之前写过一篇文章http://mryufeng.javaeye.com/blog/113852 如何调试erlang 但是这是土八路的方法， cerl才是现代化工业。

# This is a script to start Erlang/OTP for debugging. PATH is set to
# include this script so if slave nodes are started they will use this
# script as well.
#
# usage: cerl [ OPTIONS ] [ ARGS ]
#
# The OPTIONS are
#
# -rootdir $MYROOTDIR
# Run an installed emulator built from this source
# -debug Run debug compiled emulator
# -gdb Run the debug compiled emulator in emacs and gdb.
# You have to start beam in gdb using “run”.
# -break F Run the debug compiled emulator in emacs and gdb and set break.
# The session is started, i.e. “run” is already don for you.
# -xxgdb FIXME currently disabled
# -purify Run emulator compiled for purify
# -quantify Run emulator compiled for quantify
# -purecov Run emulator compiled for purecov
# -gcov Run emulator compiled for gcov
# -valgrind Run emulator compiled for valgrind
# -lcnt Run emulator compiled for lock counting
# -nox Unset the DISPLAY variable to disable us of X Windows
#

要使用cerl 我们最好准备个调试版本的erlang。R13B 修正了些编译错误,可以编译出debug版本的系统：./configure && make TYPE＝debug && make

这样就生成了个beam.debug的运行期。

我们要调试的时候 就可以在otp的binx目录下运行 cerl -debug -gdb -break main

这时候cerl自动会加载 emacs 启动 gud, 整个过程都是自动的。但是这个脚本有点小问题， gud模型下没有把源码和当前对应的调试对应起来。可以通过以下方式修正：

yu-fengdemacbook-2:bin yufeng$ diff cerl cerl2
284c284
<     exec $EMACS --eval "(progn (gdb \"gdb $EMU\") $gdbcmd)"
---
>     exec $EMACS --eval "(progn (gdb \"gdb --annotate=3  $EMU\") $gdbcmd)"

具体的操作和界面可以参照这篇文章：

http://www.nabble.com/printing-of-Eterm%27s-from-gdb-td19240493.html

在调试的时候 我们会希望查看 eterm的值，但是由于eterm的格式非常复杂， gdb的print什么的无法满足我们的需求。 otp开发团队于是开发出了一套方法来减轻我们的负担：

1. erts/etc/unix/etp-commands 这是gdb的脚本 包含了几十个etp方法，而且文档非常详细。

2. 源码里面的 pp， ps等函数， 这些函数是专门为gdb调试开发的。 可以用gdb的 call pp(xxx)来调用。

有了这些工具 调试和研究erts变成了一件很快乐的事情！

One of the things I dislike about Erlang is that it severely impairs bragging opportunities. Yesterday I wrote a module that allows writing literal XML in the source and have it parsed into Erlang structures at compile time—sort of like E4X minus the manipulation goodies at runtime (at least for now).

You write:

Doc = '<greeting>Hello!</greeting>',
io:format("~p~n", [Doc]).

And it prints…

{xmlElement,greeting,
            greeting,
            [],
            {xmlNamespace,[],[]},
            [],
            1,
            [],
            [{xmlText,[{greeting,1}],1,[],"Hello!",text}],
            [],
            "/tmp",
            undeclared}

In most languages I’m familiar with, this would have granted the author instant Yacc-demigod status. With Erlang… it was less than 40 LOC. Hardly something you’d wear at a party.

Anyway, this code owes everything to Philip’s writings. It also uses parse_transform/2, and “programmers are strongly advised not to engage in parse transformations and no support is offered for problems encountered”. So unless you, like me, are still at the kid-in-a-candy-shop stage of Erlang experience, think twice before using this in production, ok?

The code is here http://repo.hyperstruct.net/inline_xml/

[root@localhost ~]# erl -s hello
Erlang R13B02 (erts-5.7.3) [source] [64-bit] [smp:2:2] [rq:2] [async-threads:0] [hipe] [kernel-poll:false]

hello world
[root@localhost ~]# cat hello.erl

-module(hello).
-export([start/0]).

start()->
    io:format("hello world~n",[]).

[root@localhost ~]# erlc +debug_info hello.erl
[root@localhost ~]# erl
Erlang R13B02 (erts-5.7.3) [source] [64-bit] [smp:2:2] [rq:2] [async-threads:0] [hipe] [kernel-poll:false]

Eshell V5.7.3  (abort with ^G)
1> rp(beam_lib:chunks(hello, [abstract_code])).
{ok,{hello,[{abstract_code,{raw_abstract_v1,[{attribute,1,
                                                        file,
                                                        {"./hello.erl",1}},
                                             {attribute,1,module,hello},
                                             {attribute,2,export,[{start,0}]},
                                             {function,4,start,0,
                                                       [{clause,4,[],[],
                                                                [{call,5,
                                                                       {remote,5,{atom,5,io},{atom,5,format}},
                                                                       [{string,5,"hello world~n"},{nil,5}]}]}]},
                                             {eof,6}]}}]}}
ok
2>

对着文档开始好好分析吧。 Have fun!

Coming Open Source releases

OTP R13B02 September 23

OTP R13B03 November 25

OTP R13B04 Jan-Feb 2010

OTP R13B05 ???

OTP R14B May-June 2010

OTP R14B01

OTP R14B02

Plans for later releases

Multicore performance improvements

– Delayed deallocation, let the right scheduler do it (R13B03)

– Improved handling of process table

– Separate allocators per scheduler

– Use NUMA info for grouping of schedulers

– Separate poll sets per scheduler (IO)

– Support Scheduler bindings, cpu_topology on Windows as well.

– Optimize Erlang applications in Erlang/OTP

– Fine grained parallelism, language and library functions.

– Better and more benchmarks

New way to build the documentation

– Using XSLTPROC to produce html, man and XSL-FO

– Using Apache FOP to produce PDF

Easier to interface C libraries and to make your own ”BIFs”

– Dynamically linked in BIF’s (for C-code , easier to write and more efficient

than drivers)

Support for validation in xmerl_sax_parser

BIFs for search in binaries (EEP-?)

红色的部分 好期待哦。。。