dtrace discuss - Oct 2005 - Functions cpu_ticks_<> from provider sysinfo desappeared

Hi All,

 On my version of S10 3/05 118822-11, I haven''t functions
 cpu_ticks_kernel, cpu_ticks_user cpu_ticks_idle cpu_ticks_wait
 from the provider sysinfo any more. same issue with last kernel
 version 118822-20. Is it way to find again these functions ?

Thanks,
Antoine

On 10/26/05 07:14, Antoine Percher wrote:
>  On my version of S10 3/05 118822-11, I haven''t functions
>  cpu_ticks_kernel, cpu_ticks_user cpu_ticks_idle cpu_ticks_wait
>  from the provider sysinfo any more. same issue with last kernel
>  version 118822-20. Is it way to find again these functions ?
Those (sdt) probes disappeared when the old sampling based usr/sys/io/wait
mechanism was ripped out and microstate accounting enabled by default
(and modified to accurately collect usr/sys/io/wait times).  There''s
no getting them back.  Depending on what you want to do you could
add probes in the microstate state transition code instead.

Gavin

Hi Antoine,

As the other replies have noted, these probes have been removed.
There''s
a bug open to correct the documentation:

  6206973 well-documented cpu_ticks_* probes don''t exist

If you don''t have an alternate way of gathering the information you
need,
post the problem you''re trying to solve and we''ll try to help.

Adam

On Wed, Oct 26, 2005 at 08:14:20AM +0200, Antoine Percher
wrote:
> Hi All,
> 
>  On my version of S10 3/05 118822-11, I haven''t functions
>  cpu_ticks_kernel, cpu_ticks_user cpu_ticks_idle cpu_ticks_wait
>  from the provider sysinfo any more. same issue with last kernel
>  version 118822-20. Is it way to find again these functions ?
> 
> Thanks,
> Antoine
> _______________________________________________
> dtrace-discuss mailing list
> dtrace-discuss at opensolaris.org
-- 
Adam Leventhal, Solaris Kernel Development       http://blogs.sun.com/ahl

Hi,

I am trying to calculate the following times on an NFS server.

For each NFS client I need

- time spent in CPU 
- time spent in Memory (memory stalls)
- time spent Waiting for I/O.

Memory time is perhaps hard, so I  was hoping to get the other two times as
follows.

sysinfo:::cpu_ticks_*
/execname == "nfsd"/
{
    @nfsCPUTicksIp[probename, self->ip] = count();
}

Assume that self->ip is known (IP address of the client). After I finished
writing this, I realized that cpu_ticks_* probes are no longer supported.  Is
there an alternate way of getting this information?

Alternatively, what probes exist that will help me separate the CPU busy time
with the CPU wait time?

Thanks,
Piyush
 
 
This message posted from opensolaris.org

Could you work with   vtimestamp between svc_getreq:entry  &
svc_getreq:return. At   entry time you  won''t  have self->ip
known but by the time you hit the return,  you should be able
to have set self->ip.


-r

Piyush Shivam writes:
 > Hi,
 > 
 > I am trying to calculate the following times on an NFS server.
 > 
 > For each NFS client I need
 > 
 > - time spent in CPU 
 > - time spent in Memory (memory stalls)
 > - time spent Waiting for I/O.
 > 
 > Memory time is perhaps hard, so I  was hoping to get the other two times
as follows.
 > 
 > sysinfo:::cpu_ticks_*
 > /execname == "nfsd"/
 > {
 >     @nfsCPUTicksIp[probename, self->ip] = count();
 > }
 > 
 > Assume that self->ip is known (IP address of the client). After I
 > finished writing this, I realized that cpu_ticks_* probes are no
 > longer supported.  Is there an alternate way of getting this
 > information? 
 > 
 > Alternatively, what probes exist that will help me separate the CPU
 > busy time with the CPU wait time? 
 > 
 > Thanks,
 > Piyush
 >  
 >  
 > This message posted from opensolaris.org
 > _______________________________________________
 > dtrace-discuss mailing list
 > dtrace-discuss at opensolaris.org

Hi Roch,

Thanks for your reply. I am trying to avoid getting time statistics from the fbt
probes as of now due to tail-call optimization. I am writing a general purpose
observability tool for NFS (v4) and hence the hesitatio to use fbt probes.

I am also trying to get CPU busy and wait time statistics at the granularity of
NFS ops, clients, users, and directories.  I can get the total time spent in
each NFS request/operation using Sam Falkner''s nfs4trace provider, but
I am trying to zoom in further.

-Piyush
 
 
This message posted from opensolaris.org

Piyush Shivam writes:

 > Hi Roch,
 > 
 > Thanks for your reply. I am trying to avoid getting time statistics from 
 > the fbt probes as of now due to tail-call optimization. I am writing a 
 > general purpose observability tool for NFS (v4) and hence the hesitatio 
 > to use fbt probes.

I don''t see svc_getreq() using tail call. Does it ?

-r

That is true. However, if the code changes in future the tool will report
incorrect statistics. Hence the hesitation to use fbt probes.

-Piyush
 
 
This message posted from opensolaris.org