RE: [PATCH 39/52] tools/perf/build: Automatically build in parallel, based on number of CPUs in the syst

RE: [PATCH 39/52] tools/perf/build: Automatically build in parallel, based on number of CPUs in the syst

[Pádraig Brady]

On 10/08/2013 10:02 AM, Ingo Molnar wrote:
> +ifeq ($(JOBS),)
> +  JOBS := $(shell grep -c ^processor /proc/cpuinfo 2>/dev/null)

nproc is probably ubiquitous enough to use now
(available since coreutils 8.1 (end of 2009))

As well as being more concise, it will take
account of offline CPUs etc.

> +  ifeq ($(JOBS),)
> +    JOBS := 1
> +  endif
> +endif

Re: [PATCH 39/52] tools/perf/build: Automatically build in parallel, based on number of CPUs in the syst

[Ingo Molnar]

* Pádraig Brady <P@draigBrady.com> wrote:

> On 10/08/2013 10:02 AM, Ingo Molnar wrote:
> > +ifeq ($(JOBS),)
> > +  JOBS := $(shell grep -c ^processor /proc/cpuinfo 2>/dev/null)
> 
> nproc is probably ubiquitous enough to use now
> (available since coreutils 8.1 (end of 2009))
> 
> As well as being more concise, it will take
> account of offline CPUs etc.

/proc/cpuinfo takes account of offline CPUs as well:

  # grep -c ^processor /proc/cpuinfo 2>/dev/null
  16

  # echo 0 > /sys/devices/system/cpu/cpu11/online
  # grep -c ^processor /proc/cpuinfo 2>/dev/null
  15

  # echo 1 > /sys/devices/system/cpu/cpu11/online
  # grep -c ^processor /proc/cpuinfo 2>/dev/null
  16

But nproc is indeed a better choice:

1)

It is scheduler syscall based and will thus will work in limited 
environments as well, for example when /proc is not mounted.

2)

It will also properly detect affinity-limited environments:

  # taskset 1 nproc
  1

3)

It is also faster than grepping /proc/cpuinfo:

  # perf stat --null --repeat 100 nproc >/dev/null
   Performance counter stats for 'nproc' (100 runs):
       0.000652928 seconds time elapsed                                          ( +-  0.53% )

versus:

  # perf stat --null --repeat 100 grep -c ^processor /proc/cpuinfo >/dev/null
    Performance counter stats for 'grep -c ^processor /proc/cpuinfo' (100 runs):
       0.001037034 seconds time elapsed                                          ( +-  0.32% )

so with 0.652 msecs versus 1.037 msecs it's about 60% faster than grep.

Thanks,

	Ingo