Found unknown events

Found unknown events

[wufei]

Hello All,

I am using perf to measure a multi-threaded program on kernel 3.2.23,
the system has 40 cores and there are lots of context switches during
the test. The command is:
  # perf record -e sched:sched_stat_sleep -e sched:sched_switch -e
sched:sched_process_exit -g -r 0 -p 39249 sleep 5
It generates a ~1GB size perf.data. `perf report` got 2 issues:

1. A lot of this warning message:
     Warning: Timestamp below last timeslice flush

2. No useful output, but this:
     Found 32117 unknown events!

     Is this an older tool processing a perf.data file generated by a
more recent tool?

     If that is not the case, consider reporting to
linux-kernel@vger.kernel.org.

If I set a breakpoint perf_session_deliver_event(), I can see that
event->header.type is 44, which is larger than PERF_RECORD_MAX (10).
If my program is in an inactive state and less events generated,
`perf report` looks fine. Are they the knowns bugs and fixed?
Anyone can shed some light?

Thanks,
Wufei.

Re: Found unknown events

[David Ahern]

On 11/4/13, 7:40 AM, wufei wrote:
> Hello All,
>
> I am using perf to measure a multi-threaded program on kernel 3.2.23,
> the system has 40 cores and there are lots of context switches during
> the test. The command is:
>    # perf record -e sched:sched_stat_sleep -e sched:sched_switch -e
> sched:sched_process_exit -g -r 0 -p 39249 sleep 5
> It generates a ~1GB size perf.data. `perf report` got 2 issues:
>
> 1. A lot of this warning message:
>       Warning: Timestamp below last timeslice flush

What's the NUMA make up? multiple sockets?

>
> 2. No useful output, but this:
>       Found 32117 unknown events!

Try increasing the mmap size. e.g., -m 2048. Does it decrease or remove 
this stat?

David

Re: Found unknown events

[wufei]

On Tue, Nov 5, 2013 at 1:04 PM, David Ahern <dsahern@gmail.com> wrote:
> On 11/4/13, 7:40 AM, wufei wrote:
>>
>> Hello All,
>>
>> I am using perf to measure a multi-threaded program on kernel 3.2.23,
>> the system has 40 cores and there are lots of context switches during
>> the test. The command is:
>>    # perf record -e sched:sched_stat_sleep -e sched:sched_switch -e
>> sched:sched_process_exit -g -r 0 -p 39249 sleep 5
>> It generates a ~1GB size perf.data. `perf report` got 2 issues:
>>
>> 1. A lot of this warning message:
>>       Warning: Timestamp below last timeslice flush
>
>
> What's the NUMA make up? multiple sockets?
>
There are 4 nodes, each one with 10 cores.

>
>>
>> 2. No useful output, but this:
>>       Found 32117 unknown events!
>
>
> Try increasing the mmap size. e.g., -m 2048. Does it decrease or remove this
> stat?
>
It helps. There is no such problem again with this option if I only
trace it in 2 seconds,
but if it's 5 seconds, this problem is still there and another error message:
   "failed to mmap file"

Thanks,
Wufei.

> David
>