[Lustre-devel] Lustre RPC visualization

[Michael Kluge <Michael.Kluge@tu-dresden.de>]

Eric,

Vampir does scale up to this point. Almost, if I remember well, but 
thing might have changed recently). If it does not, well' make it ;) We 
have visualized real trace files with up to 20.000 processes and 
artificial trace files with many more. Heat maps can be emulated. Once 
we have had something like a color-coded counter timeline that procduces 
a heat map. Right now it can be done by on the timeline by placing 
artificial enter/leave events for function called "5%", "10%", and so 
on. I have done this already in the past, so it should not be too difficult.


Michael

Am 16.05.2010 15:12, schrieb Eric Barton:
> Excellent :)
>
> How do you think measurements taken from 1000 servers with 100,000
> clients can be visualised?  We've used heat maps to visualise
> 10s-100s of concurrent measurements (y) over time (x) but I wonder
> if that will scale.  Does vampire support heat maps?
>
>      Cheers,
>                Eric
>
>> -----Original Message-----
>> From: Michael Kluge [mailto:Michael.Kluge at tu-dresden.de]
>> Sent: 16 May 2010 10:30 AM
>> To: di.wang
>> Cc: Eric Barton; Andreas Dilger; Robert Read; Galen M. Shipman; lustre-devel
>> Subject: Re: [Lustre-devel] Lustre RPC visualization
>>
>> Hi WangDi,
>>
>> the first version works. Screenshot is attached. I have a couple of
>> counter realized: RPC's in flight and RPC's completed in total on the
>> client, RPC's enqueued, RPC's in processing and RPC'c completed in total
>> on the server. All these counter can be broken down by the type of RPC
>> (op code). The picture has not yet the lines that show each single RPC,
>> I still have to do counter like "avg. time to complete an RPC over the
>> last second" and there are some more TODO's. Like the timer
>> synchronization. (In the screenshot the first and the last counter show
>> total values while the one in the middle shows a rate.)
>>
>> What I like to have is a complete set of traces from a small cluster
>> (<100 nodes) including the servers. Would that be possible?
>>
>> Is one of you in Hamburg May, 31-June, 3 for ISC'2010? I'll be there and
>> like to talk about what would be useful for the next steps.
>>
>>
>> Regards, Michael
>>
>> Am 03.05.2010 21:52, schrieb di.wang:
>>> Michael Kluge wrote:
>>>>>> One more question: RPC 1334380768266400 (in the log WangDi sent me)
>>>>>> has on the client side only a "Sending RPC" message, thus missing the
>>>>>> "Completed RPC". The server has all three (received,start work, done
>>>>>> work). Has this RPC vanished on the way back to the client? There is
>>>>>> no further indication what happend. The last timestamp in the client
>>>>>> log is:
>>>>>> 1272565368.228628
>>>>>> and the server says it finished the processing of the request at:
>>>>>> 1272565281.379471
>>>>>> So the client log has been recorded long enough to contain the
>>>>>> "Completed RPC" message for this RPC if it arrived ever ...
>>>>> Logically, yes. But in some cases, some debug logs might be abandoned
>>>>> for some reasons(actually, it happens not rarely), and probably you need
>>>>> maintain an average time from server "Handled RPC" to client "Completed
>>>>> RPC", then you just guess the client "Completed RPC" time in this case.
>>>>
>>>> Oh my gosh ;) I don't want to start speculations about the helpfulness
>>>> of incomplete debug logs. Anyway, what can get lost? Any kind of
>>>> message on the servers and clients? I think I'd like to know what
>>>> cases have to be handled while I try to track individual RPC's on
>>>> their way.
>>> Any records can get lost here. Unfortunately, there are not any messages
>>> indicate the missing happened. :(
>>> (Usually, I would check the time stamp in the log, i.e. no records for a
>>> "long" time, for example several seconds, but this is not the accurate
>>> way).
>>>
>>> I guess you can just ignore these uncompleted records in your first
>>> step? Let's see how these incomplete log will
>>> impact the profiling result, then we will decide how to deal with this?
>>>
>>> Thanks
>>> Wangdi
>>>>
>>>> Regards, Michael
>>>> _______________________________________________
>>>> Lustre-devel mailing list
>>>> Lustre-devel at lists.lustre.org
>>>> http://lists.lustre.org/mailman/listinfo/lustre-devel
>>>
>>>
>>
>>
>> --
>> Michael Kluge, M.Sc.
>>
>> Technische Universit?t Dresden
>> Center for Information Services and
>> High Performance Computing (ZIH)
>> D-01062 Dresden
>> Germany
>>
>> Contact:
>> Willersbau, Room WIL A 208
>> Phone:  (+49) 351 463-34217
>> Fax:    (+49) 351 463-37773
>> e-mail: michael.kluge at tu-dresden.de
>> WWW:    http://www.tu-dresden.de/zih
>
>