web server in Samba4

[David Collier-Brown]

  I have done a tiny bit of work in this space (;-)) mostly using
Samba with external metrics programs and can contribute code,
if and when it becomes a priority.

  To oversimplify, the resource stats to capture are a superset
of the old BSD per-process data, and example of which is:

#time lwp pid ppid   uid usr% sys% wait% chld% rtime utime stime ttime tftime dftime kftime ltime slptime wtime stoptime ctime size   rss    minpf/s majpf/s swap/s inblk/s outblk/s chario/s msgs sig/s sysc/s vctx/s ictx/s 
 10   1  1121   563 57957 0.00 0.00 0.00 0.00  9.99  0.00  0.00  0.00  0.00   0.00   0.00   0.00   9.99   0.00     0.00   0.00  1112  736    0.00    0.00   0.00   0.00    0.00       59.25  0.00  0.10  1.10   0.10   0.00 
 20   1  1121   563 57957 0.00 0.00 0.00 0.00 10.00  0.00  0.00  0.00  0.00   0.00   0.00   0.00  10.00   0.00     0.00   0.00  1112  824    0.00    0.00   0.00   0.00    0.00       59.20  0.00  0.10  1.10   0.10   0.00 

  Performance stats to capture are
#date      time     latency xfertime responsetime throughput bytes type
05/02/2005 10:45:52 1.001623 2.009927 3.011550 0.000000 0 C
 expected latency=1s, xfertime=2s, responsetime=3s throughput=0

This is pretty easy to capture, and much easier if you do it inside
the application. I usually use net snoops and the (Solaris only) plugin,
If we have a metrics framework or plan, I can build low-cost
collectors (1 stack frame and 4-7 instructions in the common
non-collection case)

  I have a long discussion (a draft Usenix paper) of how to do performance
work with latency and response time used with/against resource usage, 
which is to big for the list: anyone interested send me mail.

-dave


James Peach wrote:
> On Wed, Jun 01, 2005 at 08:01:44PM +1000, Andrew Tridgell wrote:
> 
>>James,
>>
>> > One way we have done this successfully in the past is to write a PCP
>> > agent to export the performance metrics.
>>
>>yes, PCP is nice, and I wrote an agent for Samba2 for PCP a few years
>>back, but I don't want to rely on it for performance monitoring the
>>management interface.
> 
> 
> Fair enough. I re-read my earlier email, and it's a bit ambiguous. I'm
> not really advocating that PCP itself should be the performance metrics
> export mechanism. The export mechanism should be general enough that
> people can write both a PCP agent and an SNMP agent if they want.
> 
> 
>>I suspect we'll end up with a little monitoring framework of some sort
>>builtin, and then have some code to glue this to other monitoring
>>systems. I don't yet have a firm idea what this framework would look
>>like and right now it isn't a high priority for me. Maybe once we've
>>got the first pre-release done we can think about it.
> 
> 
> Hopefully I'll have something more concrete to contribute by that time :)
> 

-- 
David Collier-Brown,      | Always do right. This will gratify
Sun Microsystems, Toronto | some people and astonish the rest
davecb at canada.sun.com     |                      -- Mark Twain
(416) 263-5733 (x65733)   |