Re: [PATCH 0/6] perf core: Read from overwrite ring buffer

[Wangnan (F)]

On 2016/1/22 11:21, Alexei Starovoitov wrote:
> On Fri, Jan 22, 2016 at 10:21:19AM +0800, Wangnan (F) wrote:
> > On 2016/1/21 14:51, Wangnan (F) wrote:
> > > On 2016/1/20 10:20, Alexei Starovoitov wrote:
> > > > On Wed, Jan 20, 2016 at 09:37:42AM +0800, Wangnan (F) wrote:
> > > > > On 2016/1/20 1:42, Alexei Starovoitov wrote:
> > > > > > On Tue, Jan 19, 2016 at 11:16:44AM +0000, Wang Nan wrote:
> > > > > > > This patchset introduces two methods to support reading from
> > > > > > > overwrite.
> > > > > > >
> > > > > > >   1) Tailsize: write the size of an event at the end of it
> > > > > > >   2) Backward writing: write the ring buffer from the end of it to
> > > > > > > the
> > > > > > >      beginning.
> > > > > > what happend with your other idea of moving the whole header to the
> > > > > > end?
> > > > > > That felt better than either of these options.
> > > > > I'll try it today. However, putting all of the three together is
> > > > > not as easy as this patchset.
> > > > I'm missing something. Why all three in one set?
> > > Can't implement all three in one, but implement two of them make
> > > benchmarking simpler :)
> > >
> > > Here comes some numbers.
> > >
> > > I attach a target program at the end of this mail. It calls
> > > close(-1) for 3000000 times, and use gettimeofday to check
> > > how many us it takes.
> > >
> > > Following cases are tested:
> > >
> > >
> > > BASE    : ./a.out
> > > RAWPERF : ./perf record -o /dev/null -e raw_syscalls:* ./a.out
> > > WRTBKWRD: ./perf record -o /dev/null -e raw_syscalls:* ./a.out
> > > TAILSIZE: ./perf record --no-has-write-backward -o /dev/null -e
> > > raw_syscalls:*/overwrite/ ./a.out
> > > RAWOVWRT: ./perf record --no-has-write-backward --no-has-tailsize -o
> > > /dev/null -e raw_syscalls:*/overwrite/ ./a.out
> > >
> > > With this script:
> > >
> > > func() {
> > >         for x in `seq 1 100` ; do $1; done | tee data_$2
> > > }
> > >
> > > func ./a.out base
> > > func "./perf record -o /dev/null -e raw_syscalls:* ./a.out" rawperf
> > > func "./perf record -o /dev/null -e raw_syscalls:*/overwrite/ ./a.out"
> > > wrtbkwrd
> > > func "./perf record -o /dev/null --no-has-write-backward -e
> > > raw_syscalls:*/overwrite/ ./a.out" tailsize
> > > func "./perf record -o /dev/null --no-has-write-backward --no-has-tailsize
> > > -o /dev/null -e raw_syscalls:*/overwrite/ ./a.out" rawovwrt
> > >
> > > Result:
> > >
> > >             MEAN           STDVAR
> > > BASE    :  879870.81      11913.13
> > > RAWPERF : 2603854.7      706658.4
> > > WRTBKWRD: 2313301.220      6727.957
> > > TAILSIZE: 2383051.860      5248.061
> > > RAWOVWRT: 2315273.180      5221.025
> > Add a number: I tested original perf overwrite ring buffer in pure v4.4
> > on the same machine:
> >
> >                      MEAN          STDVAR
> > RAWOVWRT(original): 2323970.45    5103.39
> >
> > So I think backward writing method doesn't add extra overhead into
> > fastpath.
> >
> > I will send this patchset again with several bugs fixed. After that
> > I'll start working on tail-header if it is still required.
> interesting.
> did I read the numbers correctly that 'write backwards' method
> is actually the fastest? even faster than no-overwrite?

Yes. But notice STDVAR, we can't say 'WRTBKWRD' outperform 'RAWOVWRT'. 
However,
at least 'WRTBKWRD' should be as fast as 'RAWOVWRT'.

> nice. I guess it makes snese that overwrite is faster.

In no-overwrite case perf itself wakes up many times to collect data,
I guess it is the source of high stdvar.

> I guess than moving the header to the end will have the same
> performance in this benchmark, since RAWOVWRT is the same as well.
Yes.

Do you want to test it by yourself? The code is ready.

Thank you.