Bug 187919 – Revert back to using phys_footprint to calculate isUnderMemoryPressure()

• Status: RESOLVED
• Resolution: FIXED
• Component: bmalloc

[Saam Barati]

Currently, we rely on bmalloc's physicalFootprint calculation. However, bmalloc may have a small or mediumish heap, but the process may be using a lot of dirty memory from elsewhere. It could still be helpful in preventing jetsam if bmalloc freed stuff.

[Saam Barati]

Created attachment 345630 [details]
patch

[Simon Fraser (smfr)]

Comment on attachment 345630 [details]
patch

View in context: https://bugs.webkit.org/attachment.cgi?id=345630&action=review

> Source/bmalloc/bmalloc/Scavenger.cpp:136
> +        task_vm_info_data_t vmInfo;
> +        mach_msg_type_number_t vmSize = TASK_VM_INFO_COUNT;
> +        if (KERN_SUCCESS == task_info(mach_task_self(), TASK_VM_INFO, (task_info_t)(&vmInfo), &vmSize))
> +            return static_cast<double>(vmInfo.phys_footprint);
> +        return 0.0;

Can this share code with memoryStatus()?

> Source/bmalloc/bmalloc/Scavenger.cpp:140
> +    if (osFootprint / hardwareMemoryCapacity > 0.9)
> +        return true;

I don't know how to reason whether that 90% threshold is reasonable. If feels bad to not scavenge if we're using 800MB, because 800MB is huge.

[Saam Barati]

Comment on attachment 345630 [details]
patch

View in context: https://bugs.webkit.org/attachment.cgi?id=345630&action=review

>> Source/bmalloc/bmalloc/Scavenger.cpp:136
>> +        return 0.0;
> 
> Can this share code with memoryStatus()?

Yeah I can make it shared

>> Source/bmalloc/bmalloc/Scavenger.cpp:140
>> +        return true;
> 
> I don't know how to reason whether that 90% threshold is reasonable. If feels bad to not scavenge if we're using 800MB, because 800MB is huge.

This is just a heuristic. But specifically in this case, 800/840 > .9

[Joseph Pecoraro]

Comment on attachment 345630 [details]
patch

This looks good to me.

[Geoffrey Garen]

The logic of this code is getting pretty complicated, with multiple authors and multiple constraints we're trying to solve for.

How do we know if we're making things better or worse?

[Saam Barati]

(In reply to Geoffrey Garen from comment #5)
> The logic of this code is getting pretty complicated, with multiple authors
> and multiple constraints we're trying to solve for.
> 
> How do we know if we're making things better or worse?

When I implemented partial scavenging I manually verified we use less memory on average on the popular JS benchmarks.

For this change, I don’t know of a way to validate it besides just thinking about certain heap shapes that might lead bmalloc to not return memory to the OS.

Reasoning about this change, the only way I see it making things worse is by overscavenging when we’re close to the process’s memory limit. If that happens we’re in a world where we’re burning more CPU than necessary. However, it still seems better under that situation to spend CPU trying to free memory instead of allowing the process to jetsam

[Radar WebKit Bug Importer]

<rdar://problem/42552888>

[Saam Barati]

Created attachment 345731 [details]
patch

[WebKit Commit Bot]

Comment on attachment 345731 [details]
patch

Clearing flags on attachment: 345731

Committed r234185: <https://trac.webkit.org/changeset/234185>

[WebKit Commit Bot]

All reviewed patches have been landed.  Closing bug.