As of Chrome 48, MemoryInfra supports heap profiling. Chrome will track all live allocations (calls to new or malloc without a subsequent call to delete or free) along with sufficient metadata to identify the code that made the allocation.
By default, MemoryInfra traces will not contain heap dumps. Heap profiling must be enabled via chrome://memory-internals or about://flags.
memlog
.save dump
. This is stored as a MemoryInfra trace.addr2line-pdb
from the chromium repository. For subsequent commands, add the flag --addr2line-executable=<path_to_addr2lin-pdb>
./third_party/catapult/tracing/bin/symbolize_trace --is-local-build <path_to_trace>
./third_party/catapult/tracing/bin/symbolize_trace <path_to_trace>
. This will request authentication with google cloud storage to obtain symbol files [googlers only].--use-breakpad-symbols
.--only-symbolize-chrome-symbols
.On arm64 and x86-64, you can build chrome normally and follow steps above to obtain heap dumps.
To obtain native heap dumps on arm32, you will need a custom build of Chrome with the GN arguments enable_profiling = true
, arm_use_thumb = false
, is_component_build = false
and symbol_level=1
. All other steps are the same.
Alternatively, if you want to use an official build of Chrome, use is_official_build = true
for arm32. If you want to use a released build, profiling only works on Dev and Canary on arm, and all channels on x86-64. In this case, you also need to fetch symbols manually and pass to the symbolize_trace script above.
For the most part, the setting enable-heap-profiling
in chrome://flags
has a similar effect to the various memlog
flags.
Select a heavy memory dump indicated by a purple dot.
In the analysis view, cells marked with a triple bar icon (☰) contain heap dumps. Select such a cell.
Scroll down all the way to Heap Details.
To navigate allocations, select a frame in the right-side pane and press Enter/Return. To pop up the stack, press Backspace/Delete.
Run python ./third_party/catapult/experimental/tracing/bin/diff_heap_profiler.py <path_to_trace>
This produces a directory output
, which contains a JSON file.
Load the contents of the JSON file in any JSON viewer, e.g. jsonviewer.
The JSON files shows allocations segmented by stacktrace, sorted by largest first.
The heap details view contains a tree that represents the heap. The size of the root node corresponds to the selected allocator cell.
The heap can be broken down in two ways: by backtrace (marked with an ƒ), and by type (marked with a Ⓣ). When tracing is enabled, Chrome records trace events, most of which appear in the flame chart in timeline view. At every point in time these trace events form a pseudo stack, and a vertical slice through the flame chart is like a backtrace. This corresponds to the ƒ nodes in the heap details view. Hence enabling more tracing categories will give a more detailed breakdown of the heap.
The other way to break down the heap is by object type. At the moment this is only supported for PartitionAlloc.
To keep the trace log small, uninteresting information is omitted from heap dumps. The long tail of small nodes is not dumped, but grouped in an <other>
node instead. Note that although these small nodes are insignificant on their own, together they can be responsible for a significant portion of the heap. The <other>
node is large in that case.
In the trace below, ParseAuthorStyleSheet
is called at some point.
The pseudo stack of trace events corresponds to the tree of ƒ nodes below. Of the 23.5 MiB of memory allocated with PartitionAlloc, 1.9 MiB was allocated inside ParseAuthorStyleSheet
, either directly, or at a deeper level (like CSSParserImpl::parseStyleSheet
).
By expanding ParseAuthorStyleSheet
, we can see which types were allocated there. Of the 1.9 MiB, 371 KiB was spent on ImmutableStylePropertySet
s, and 238 KiB was spent on StringImpl
s.
It is also possible to break down by type first, and then by backtrace. Below we see that of the 23.5 MiB allocated with PartitionAlloc, 1 MiB is spent on Node
s, and about half of the memory spent on nodes was allocated in HTMLDocumentParser
.
Heap dump diffs are fully supported by trace viewer. Select a heavy memory dump (a purple dot), then with the control key select a heavy memory dump earlier in time. Below is a diff of theverge.com before and in the middle of loading ads. We can see that 4 MiB were allocated when parsing the documents in all those iframes, almost a megabyte of which was due to JavaScript. (Note that this is memory allocated by PartitionAlloc alone, the total renderer memory increase was around 72 MiB.)