[Pvfs2-users] Re: Large times recorded in trace data

[Phil Carns]

I think that the seconds and microseconds are in absolute times (as in 
from the gettimeofday() call).

This particular event tracing mechanism is pretty limited, but you can 
see how it breaks down by looking in pvfs2-event.h.  That shows what the 
api and event values mean so that you can see what is disk (trove) and 
what is network (bmi), and also look at specific operations (like bmi 
send, for example).

There is no tracing of the client with this mechanism, it is server side 
only.

This trace information makes no distinction between I/O servers and meta 
servers; it should be showing you data for all of the servers regardless 
of their role.

-Phil

Nirmal Thacker wrote:
> I forgot to add the attachment.
> Its here now
> This is a full trace , with api and operations set to 0xFFFF
> -nirmal
> Nirmal Thacker wrote:
>> Hello,
>>
>> When I trace events in PVFS2, using the pvfs2-set-eventmask and
>> pvfs2-event-mon-example I get a randomly large counter for the seconds
>> and microseconds for each trace.
>>
>> I have all my servers and clients synced using NTP .
>>
>> I have attached a trace of 16 clients to 4 I/O servers, 1 MDS server
>> [separate MDS] opening 2500 files using the mpi-md-test distributed with
>> pvfs2. The entire operation too about 17 seconds.
>>
>> The large seconds value is used more as a counter. As I can see the
>> number of microseconds each operation takes and this increments the
>> seconds counter later eventually. But my question is- why do you have
>> such a large counter- does this number mean something and how is it
>> generated?
>>
>> The total amount of time spent at servers doesnt seem to be more than a
>> second. For example server No. 3 [in the trace file] has all its
>> operations done in 0.165442 microseconds . I obtained this value from
>> the difference in the timing counters of the first and last trace for
>> server no. 3.
>>
>> Am I interpreting this right?
>> If yes, can it be possible for me to obtain a time breakup of time spent
>> over the network, Client, Disk etc?
>>
>> Finally, I can see traces for servers 0,1,2,3. These would be the I/O
>> servers. How can I generate MDS traces or view the times for MDS operations?
>>
>> Thanks
>> Nirmal
>>
>>
>>
>>
>>   
> 
> 
> ------------------------------------------------------------------------
> 
> _______________________________________________
> Pvfs2-users mailing list
> Pvfs2-users at beowulf-underground.org
> http://www.beowulf-underground.org/mailman/listinfo/pvfs2-users