RFR: Parallel +AlwaysPreTouch should run with max workers

[Roman Kennke]

Am 11.06.2018 um 15:53 schrieb Aleksey Shipilev:
> Current +AlwaysPreTouch is not parallel (it used to be!), because default policy starts with a
> single active GC worker during init. We need to wind up more active workers for this phase. We know
> this affects sensitive workloads on NUMA machines, because it basically interleaves memory, so this
> should improve performance there.
> 
> diff -r e4a301c3f11b src/hotspot/share/gc/shenandoah/shenandoahHeap.cpp
> --- a/src/hotspot/share/gc/shenandoah/shenandoahHeap.cpp	Mon Jun 11 13:54:58 2018 +0200
> +++ b/src/hotspot/share/gc/shenandoah/shenandoahHeap.cpp	Mon Jun 11 15:49:38 2018 +0200
> @@ -280,10 +280,11 @@
>      assert (!AlwaysPreTouch, "Should have been overridden");
> 
>      // For NUMA, it is important to pre-touch the storage under bitmaps with worker threads,
>      // before initialize() below zeroes it with initializing thread. For any given region,
>      // we touch the region and the corresponding bitmaps from the same thread.
> +    ShenandoahWorkerScope scope(workers(), _max_workers);
> 
>      log_info(gc, heap)("Parallel pretouch " SIZE_FORMAT " regions with " SIZE_FORMAT " byte pages",
>                         _num_regions, page_size);
>      ShenandoahPretouchTask cl(bitmap0.base(), bitmap1.base(), _bitmap_size, page_size);
>      _workers->run_task(&cl);
> 
> Testing: tier1_gc_shenandoah, +AlwaysPreTouch with -Xmx100g
> 
> Thanks,
> -Aleksey
> 

Yes. Thank you!