Re: [PATCH 3/3] perf record: mmap output file

[Ingo Molnar <mingo@kernel.org>]

* David Ahern <dsahern@gmail.com> wrote:

> When recording raw_syscalls for the entire system, e.g.,
>     perf record -e raw_syscalls:*,sched:sched_switch -a -- sleep 1
> 
> you end up with a negative feedback loop as perf itself calls
> write() fairly often. This patch handles the problem by mmap'ing the
> file in chunks of 64M at a time and copies events from the event buffers
> to the file avoiding write system calls.
> 
> Before (with write syscall):
> 
> perf record -o /tmp/perf.data -e raw_syscalls:*,sched:sched_switch -a -- sleep 1
> [ perf record: Woken up 0 times to write data ]
> [ perf record: Captured and wrote 81.843 MB /tmp/perf.data (~3575786 samples) ]
> 
> After (using mmap):
> 
> perf record -o /tmp/perf.data -e raw_syscalls:*,sched:sched_switch -a -- sleep 1
> [ perf record: Woken up 31 times to write data ]
> [ perf record: Captured and wrote 8.203 MB /tmp/perf.data (~358388 samples) ]
> 
> In addition to perf-trace benefits using mmap lowers the overhead of
> perf-record. For example,
> 
>   perf stat -i -- perf record -g -o /tmp/perf.data openssl speed aes
> 
> showsi a drop in time, CPU cycles, and instructions all drop by more than a
> factor of 3. Jiri also ran a test that showed a big improvement.

Here are some thoughts on how 'perf record' tracing performance could be 
further improved:

1)

The use of non-temporal stores (MOVNTQ) to copy the ring-buffer into the 
file buffer makes sure the CPU cache is not trashed by the copying - which 
is the largest 'collateral damage' copying does.

glibc does not appear to expose non-temporal instructions so it's going to 
be architecture dependent - but we could build the copy_user_nocache() 
function from the kernel proper (or copy it - we could even simplify it: 
knowing that only large and page aligned buffers are going to be copied 
with it).

See how tools/perf/bench/mem-mem* does that to be able to measure the 
kernel's memcpy() and memset() function performance.

2)

Yet another method would be to avoid the copies altogether via the splice 
system-call - see:

	git grep splice kernel/trace/

To make splice low-overhead we'd have to introduce a mode to not mmap the 
data part of the perf ring-buffer and splice the data straight from the 
perf fd into a temporary pipe and over from the pipe into the target file 
(or socket).

OTOH non-temporal stores are incredibly simple and memory bandwidth is 
plenty on modern systems so I'd certainly try that route first.

Thanks,

	Ingo