Option to filter out small allocations in statistics to avoid getting extra stacks

Evgeniy Ivanov i at eivanov.com
Wed Mar 19 03:45:35 PDT 2014

Hi Jason,

I agree, but sometimes you need to hunt big allocations and according
our experience (high frequency and algorithmic trading) cutting helps
a lot.

On Mon, Mar 17, 2014 at 10:38 PM, Jason Evans <jasone at canonware.com> wrote:
> On Mar 17, 2014, at 1:13 AM, Evgeniy Ivanov <i at eivanov.com> wrote:
> We use opt.prof_accum to profile memory allocations. Sometimes performance
> degradation is too high because getting stacks is a heavy operation. In
> resulting backtraces we see some relatively small allocations we are not
> interested in. With DTrace our usual case is to cut allocations smaller than
> 16 KB. I would like to add option to cut allocations smaller, than specified
> one, e.g. "opt.prof_cut" (ssize_t). Will it be accepted upstream?
> The way to reduce backtracing overhead is to increase the sample interval,
> e.g. MALLOC_CONF=lg_prof_sample:20 doubles the default from an average of
> one sample per 2^19 bytes (512 KiB) to 2^20 bytes (1 MiB).  If you
> systematically ignore small allocations you bias the sample, which
> invalidates the math that allows sample-based profiling to converge on
> reality.
> Jason


More information about the jemalloc-discuss mailing list