Re: some questions about systemtap

Re: some questions about systemtap

[Peter Teoh]

kernel path tracing is what systemtap is good at, for for TCP -
evolution of stack trace indicate something of the kernel path
traversed:

 0xc0612dc7 : tcp_poll+0x1/0x133 [kernel]
 0xc04a2578 : do_sys_poll+0x210/0x3c3 [kernel]
 0xc04a287a : sys_poll+0x45/0x8f [kernel]
 0xc0403b6b : ia32_sysenter_target+0x7b/0x138 [kernel]
--->tcp_poll

 0xc0620c8e : tcp_v4_init_sock+0x1/0x14b [kernel]
 0xc05e3a93 : __sock_create+0x11d/0x19d [kernel]
 0xc05e3b5a : sock_create+0x29/0x2e [kernel]
 0xc05e3d63 : sys_socket+0x31/0x5f [kernel]
 0xc05e46c8 : sys_socketcall+0x73/0x192 [kernel]
 0xc0403b6b : ia32_sysenter_target+0x7b/0x138 [kernel]
--->tcp_v4_init_sock

 0xc0620185 : tcp_init_xmit_timers+0x1/0x1f [kernel]
 0xc062da65 : inet_create+0x24c/0x26f [kernel]
 0xc05e3a93 : __sock_create+0x11d/0x19d [kernel]
 0xc05e3b5a : sock_create+0x29/0x2e [kernel]
 0xc05e3d63 : sys_socket+0x31/0x5f [kernel]
 0xc05e46c8 : sys_socketcall+0x73/0x192 [kernel]
 0xc0403b6b : ia32_sysenter_target+0x7b/0x138 [kernel]
--->tcp_init_xmit_timers

 0xc0620cc6 : tcp_v4_init_sock+0x39/0x14b [kernel]
 0xc062da65 : inet_create+0x24c/0x26f [kernel]
 0xc05e3a93 : __sock_create+0x11d/0x19d [kernel]
 0xc05e3b5a : sock_create+0x29/0x2e [kernel]
 0xc05e3d63 : sys_socket+0x31/0x5f [kernel]
 0xc05e46c8 : sys_socketcall+0x73/0x192 [kernel]
 0xc0403b6b : ia32_sysenter_target+0x7b/0x138 [kernel]
--->tcp_prequeue_init

 0xc06237a3 : tcp_v4_connect+0x1/0x4c2 [kernel]
 0xc05e4034 : sys_connect+0x65/0x82 [kernel]
 0xc05e46e8 : sys_socketcall+0x93/0x192 [kernel]
 0xc0403b6b : ia32_sysenter_target+0x7b/0x138 [kernel]
--->tcp_v4_connect

 0xc0613a50 : tcp_set_state+0x1/0xe1 [kernel]
 0xc062d489 : inet_stream_connect+0x8f/0x1f9 [kernel]
 0xc05e4034 : sys_connect+0x65/0x82 [kernel]
 0xc05e46e8 : sys_socketcall+0x93/0x192 [kernel]
 0xc0403b6b : ia32_sysenter_target+0x7b/0x138 [kernel]
--->tcp_set_state
 0xc061de43 : tcp_connect+0x1/0x3a0 [kernel]
 0xc062d489 : inet_stream_connect+0x8f/0x1f9 [kernel]
 0xc05e4034 : sys_connect+0x65/0x82 [kernel]
 0xc05e46e8 : sys_socketcall+0x93/0x192 [kernel]
 0xc0403b6b : ia32_sysenter_target+0x7b/0x138 [kernel]
--->tcp_connect

 0xc061de51 : tcp_connect+0xf/0x3a0 [kernel]
 0xc0623c0c : tcp_v4_connect+0x46a/0x4c2 [kernel]
 0xc062d489 : inet_stream_connect+0x8f/0x1f9 [kernel]
 0xc05e4034 : sys_connect+0x65/0x82 [kernel]
 0xc05e46e8 : sys_socketcall+0x93/0x192 [kernel]
 0xc0403b6b : ia32_sysenter_target+0x7b/0x138 [kernel]
--->tcp_connect_init

 0xc0620c1d : tcp_v4_md5_lookup+0x1/0x3a [kernel]
 0xc0623c0c : tcp_v4_connect+0x46a/0x4c2 [kernel]
 0xc062d489 : inet_stream_connect+0x8f/0x1f9 [kernel]
 0xc05e4034 : sys_connect+0x65/0x82 [kernel]
 0xc05e46e8 : sys_socketcall+0x93/0x192 [kernel]
 0xc0403b6b : ia32_sysenter_target+0x7b/0x138 [kernel]
--->tcp_v4_md5_lookup

 0xc0620c27 : tcp_v4_md5_lookup+0xb/0x3a [kernel]
 0xc061de77 : tcp_connect+0x35/0x3a0 [kernel]
 0xc0623c0c : tcp_v4_connect+0x46a/0x4c2 [kernel]
 0xc062d489 : inet_stream_connect+0x8f/0x1f9 [kernel]
 0xc05e4034 : sys_connect+0x65/0x82 [kernel]
 0xc05e46e8 : sys_socketcall+0x93/0x192 [kernel]
 0xc0403b6b : ia32_sysenter_target+0x7b/0x138 [kernel]
--->tcp_v4_md5_do_lookup

but for time measurement...u need another tool, possibly some network
benchmarking tool.

2008/11/25 zshan <zshan@cse.buaa.edu.cn>:
>   Hello,
>
>   I and my mentor Mr hao （cc-ed in the mail ）want to research on is the
> network I/O, which including dececting the network I/O path and the time
> each process takes durning one network I/O request .
>
>   We wonder whether it is feasible ?
>
>   Any ideas ?
>
>   thank you ~
>
> 2008-11-25
> ________________________________
> Regrads
> zengshan
>
>
>



-- 
Regards,
Peter Teoh

Ernest Hemingway - "Never mistake motion for action."

Re: some questions about systemtap

[Arnaldo Carvalho de Melo]

Em Tue, Nov 25, 2008 at 04:20:39PM +0800, Peter Teoh escreveu:
> kernel path tracing is what systemtap is good at, for for TCP -
> evolution of stack trace indicate something of the kernel path
> traversed:

<SNIP>
 
> but for time measurement...u need another tool, possibly some network
> benchmarking tool.
> 
> 2008/11/25 zshan <zshan@cse.buaa.edu.cn>:
> >   Hello,
> >
> >   I and my mentor Mr hao （cc-ed in the mail ）want to research on is the
> > network I/O, which including dececting the network I/O path and the time
> > each process takes durning one network I/O request .
> >
> >   We wonder whether it is feasible ?
> >
> >   Any ideas ?

Something like this? For the TCP receive path as seen from functions
that have pointers to struct sock instances:

http://oops.ghostprotocols.net:81/acme/dwarves/callgraphs/sock/0xf61bf500/

More on
http://oops.ghostprotocols.net:81/acme/dwarves/callgraphs/

For dccp, that shares a lot of infrastructure with TCP:

http://oops.ghostprotocols.net:81/ostra/dccp/tx/
http://oops.ghostprotocols.net:81/ostra/dccp/rx/

It generates a systemtap script to insert the hooks at function
entry/exit and uses some postprocessing tools to generate the HTML + CSS
callgraphs, read about it at:

http://git.kernel.org/?p=linux/kernel/git/acme/pahole.git;a=blob_plain;f=README.ctracer

- Arnaldo

Re: some questions about systemtap

[Frank Ch. Eigler]

"Peter Teoh" <htmldeveloper@gmail.com> writes:

> [...]
> but for time measurement...u need another tool, possibly some network
> benchmarking tool. [...]

Why do you think so?  A systemtap script can easily take timestamps at
various points, and perform individual or aggregate computations on them.

- FChE

Re: some questions about systemtap

[Peter Teoh]

On Tue, Nov 25, 2008 at 11:05 PM, Frank Ch. Eigler <fche@redhat.com> wrote:
>
> "Peter Teoh" <htmldeveloper@gmail.com> writes:
>
>> [...]
>> but for time measurement...u need another tool, possibly some network
>> benchmarking tool. [...]
>
> Why do you think so?  A systemtap script can easily take timestamps at
> various points, and perform individual or aggregate computations on them.
>

Of course, stap's command line has the benchmark timestamp option, and
within the stp script you can quote your own timestamp as well....but
perhaps what I meant to say was that if u used some special tools,
everything is pre-written for you.....involving all the different ways
of looking/generating the statistics.   If every other people is also
using that particular tool, then your statistics can easily be
compared with others, without the effort of regenerating the
statistics.

But then there may be other caveats I am not aware of......comments?

> - FChE
>



-- 
Regards,
Peter Teoh

Ernest Hemingway - "Never mistake motion for action."