Re: heads up/RFC: 'perf trace' using ordered_events

[David Ahern <dsahern@gmail.com>]

On 3/10/15 12:06 AM, Namhyung Kim wrote:
> Hi Arnaldo,
>
> On Mon, Mar 09, 2015 at 10:21:35AM -0300, Arnaldo Carvalho de Melo wrote:
>> For trace I need to take advantage of the fact that each mmap is ordered
>> already and then just sort by the timestamp in the mmap head, etc.
>>
>> In retrospect, the perf.data file should have kept that ordering, i.e.
>> have one file per mmap, that would be saved in parallel, without any of
>> those PERF_RECORD_FINISHED_ROUND records.
>>
>> But I have to experiment with that, leaving the existing code around to
>> deal with older files.
>
> It seems like what you said is almost same as my multi-thread work.
> It saves data files per mmap and then merges them with an index
> table so that they can be processed in parallel.

I think you and Jiri both have worked on saving a file per mmap. Where 
does that stand? I would like to try it out.

With the 1024 cpu systems I am seeing 5GB files in 1 second runs and 
perf is not handling it well. The perf-script/perf-report (stdio) will 
'hang' for 45 minutes munging through the file. I have to connect gdb 
from time to time to verify it is making progress (file_offset is 
increasing). I believe what happens is that there is 'no round' -- it 
has to process all mmaps (1024 cpus and 6 or 7 events) through the 
ordered events queue before it can push out results. I need to look at 
that once I figure out the task scheduler problem.

David