Memory & Binding team snippet

[Kentaro Hara]

Hi

Oilpan:

(haraken, sigbjornf, keishi, yutak) Resolved a bunch of issues Oilpan has introduced. We made great progress:

[Stability]

- sigbjornf@ fixed a bunch of crash reports.

- keishi@ fixed failures of inspector tests.

- As far as I look at the top 150 crash reports from 50.0.2638.0 (Shipped to Win/Mac Dev on Feb 2), all Oilpan-related crashes are already filed and taken care of. None of them is a serious crasher. The stability looks extremely good.

[Memory]

- yutak@ investigated regressions in memory.memory_health_plan and identified that the regressions are caused by the fact that Oilpan does not aggressively discard system pages in free lists. haraken@ created a CL to make Oilpan discard system pages in free lists on low-RAM devices.

[Performance]

- haraken@ investigated regressions in V8 GC metrics and identified that the regressions are caused by the fact that Oilpan's GC can be triggered during prologues/epilogues of V8 GCs, making the V8 GC metrics worse. haraken@ is working on factoring out Oilpan's GC from prologues/epilogues of V8 GCs (CL, CL, CL).

- keishi@ investigated other performance regressions (e.g., frame_times) and noticed that most of the regressions Oilpan has introduced were already recovered (for some reason we don't know; we are lucky anyway :-). The remaining regressions are page_cycler.morejs and a couple of micro benchmarks.

- keishi@ is investigating slowness of layout tests on Debug builds.

TRIM-platform:

(tasak) Experimenting with purging as many things as possible in the renderer to see how much memory we can save from the renderer. georgesak@ improved the TabManager so that memory pressure notifications are delivered to the renderers, so we can use the notifications to trigger the purging. Our goal is to make all sub-systems in the renderer be aware of the pressure notifications and take action when requested by the TabManager.

(bashi) Investigating how to reduce the peak memory consumption on low-RAM devices. bashi@ noticed that redundant copying from SharedBuffer::m_segments to SharedBuffer::m_buffer may contribute to the peak memory increase. bashi@ is investigating how much memory we can save by getting rid of the redundant copying. (Either way, it's worth cleaning up the memory management of SharedBuffers -- working with the loading team.)

(hajimehoshi) Supporting the heap profiler in Oilpan. Did some refactoring in preparation for that.

(haraken) Wrote a document to summarize the current status of various memory-reduction teams in Chromium. (I'll publish the document once it's ready.)

Onion Soup:

(yukishiino) Moving battery_status from content/renderer/ to Blink's platform/. There are a couple of design issues such as whether we should allow using Mojo in modules/, how to write tests for Mojo services (in layout tests in JS or in unittests in C++). Discussing the design.

(yutak) Cleaning up WTF; e.g., remove time functions from platform, introduce std::tuple etc. Oilpan => unique_ptr replacement is blocked by Oilpan's WillBe type removals.

(tasak, jbroman) Supporting core-module componentization in GN.

(haraken) I've learned what the "Onion Soup" means.

An onion has a lot of layers:

​

By (reasonably) breaking the unnecessary layers into a soup, the world becomes a better place to live:

​

V8 bindings:

(haraken) Removed sampling tracing macros from the V8 binding layer (CL). (Now that V8 supports tracing, if we really want to have tracing macros in the binding layer, we should add the macros to V8 instead of adding them to each C++ callback for DOM attribute/method/constant.) This reduced the binary size by 600 KB.

(yukishiino, verwaest) Working on a complicated behavioral bug around SetAccessor/SetNativeDataProperty and [Replaceable] DOM attributes.

(peria, haraken) Oilpan knows a good estimation of the size of DOM objects indirectly retained by V8 wrappers (e.g., size of NodeLists, size of Nodes etc). peria@ is making Oilpan report the size to V8. This enables us to remove hand-written AdjustAmountOfExternalAllocatedMemory from the code base.

(haraken) Moving hasPendingActivity from ActiveDOMObject to ScriptWrappable. Now the CL passes all tests.

(haraken) Considering how to deprecate v8::Object::CreationContext (requested by the V8 team).

-- 

Kentaro Hara, Tokyo, Japan

--

Kentaro Hara, Tokyo, Japan