Re: [PATCH] perf record: mmap output file - v2

[David Ahern <dsahern@gmail.com>]

On 10/15/13 1:44 AM, Ingo Molnar wrote:
>
> * Namhyung Kim <namhyung@kernel.org> wrote:
>
>> [SNIP]
>>> +/* mmap file big chunks at a time */
>>> +#define MMAP_OUTPUT_SIZE   (64*1024*1024)
>>
>> Why did you choose 64MB for the size?  Did you also test other sizes?
>
> Btw., should this value go up if the ring buffer (mmap_pages) is larger
> than 64MB?
>

I made mmap_size a variable:
+	size_t			mmap_size;      /* size of mmap segments */

with the above initial value. I was planning to make it in an option and 
just forgot to complete it.

Why 64M? mmap / munmap are also system calls and I was looking to trade 
off huge file size jumps versus the frequency of adjusting the maps. 64M 
Was just a nice round number between 1 and 100. 8, 16 are too small. 128 
seems to big for a default. That left only 32 and 64. 64M seems the 
better trade off of the two.

Making it a user knob would help with smaller deployments. Could also 
have mmap_size = 0 mean turn it off (use write over mmap).

Perhaps something that adjust automatically would be useful too. e.g., 
For the case that motivates the change I have 16 cpus each with a 4M 
buffer (1024 mmap pages). Should we generically set the size:

mmap_size = ncpus_online * mmap_pages * page_size?

Do that only for system wide profiling?

David