See design doc: https://docs.google.com/document/d/1zcfWSsgwTY8WYYlxd0G8QTR6_b2j38uMfjjXBOuMZnM/edit#

for more details.

hjd is doing some experiments here correlating the full computation from /proc/pid/smaps with the fast approximation from /proc/pid/{statm,status} as described in the doc #1.

Primiano and I experimented. Still finalizing the doc but it seems like the fast calculation is a pretty good approximation. https://docs.google.com/document/d/1gF3JVf10J0xr2u13gNKWW63tMTMBrGG_aLzgS0XfdhE/edit# 

The following revision refers to this bug:
  https://chromium.googlesource.com/chromium/src.git/+/a28aa41324c5572e67d57c610fccd22132df3de4

commit a28aa41324c5572e67d57c610fccd22132df3de4
Author: hjd <hjd@chromium.org>
Date: Wed May 03 11:32:50 2017

memory-infra: Plumb platform_private_footprint to MemoryDumpManager

BUG= 707019 

Review-Url: https://codereview.chromium.org/2852473002
Cr-Commit-Position: refs/heads/master@{#468941}

[modify] https://crrev.com/a28aa41324c5572e67d57c610fccd22132df3de4/base/trace_event/memory_dump_manager.cc
[modify] https://crrev.com/a28aa41324c5572e67d57c610fccd22132df3de4/base/trace_event/memory_dump_request_args.h
[modify] https://crrev.com/a28aa41324c5572e67d57c610fccd22132df3de4/base/trace_event/process_memory_totals.h
[modify] https://crrev.com/a28aa41324c5572e67d57c610fccd22132df3de4/services/resource_coordinator/public/cpp/memory/memory_instrumentation.typemap
[modify] https://crrev.com/a28aa41324c5572e67d57c610fccd22132df3de4/services/resource_coordinator/public/cpp/memory/memory_instrumentation_struct_traits.cc
[modify] https://crrev.com/a28aa41324c5572e67d57c610fccd22132df3de4/services/resource_coordinator/public/cpp/memory/memory_instrumentation_struct_traits.h
[modify] https://crrev.com/a28aa41324c5572e67d57c610fccd22132df3de4/services/resource_coordinator/public/interfaces/memory/memory_instrumentation.mojom

The following revision refers to this bug:
  https://chromium.googlesource.com/chromium/src.git/+/90e0440fbcc073544c8645595268e641ad8e51a5

commit 90e0440fbcc073544c8645595268e641ad8e51a5
Author: hjd <hjd@chromium.org>
Date: Mon May 08 13:00:25 2017

Fill in PlatformPrivateFootprint on Linux 1/2

This CL populates the rss_anon_bytes field of the PlatformPrivateFootprint struct on Linux.

The struct contains platform specific numbers that will be used by the
memory-infra service to compute the private memory footprint of the
process.

BUG= 707019 

Review-Url: https://codereview.chromium.org/2839733004
Cr-Commit-Position: refs/heads/master@{#469968}

[modify] https://crrev.com/90e0440fbcc073544c8645595268e641ad8e51a5/base/trace_event/process_memory_totals.h
[modify] https://crrev.com/90e0440fbcc073544c8645595268e641ad8e51a5/components/tracing/common/process_metrics_memory_dump_provider.cc
[modify] https://crrev.com/90e0440fbcc073544c8645595268e641ad8e51a5/components/tracing/common/process_metrics_memory_dump_provider.h

The following revision refers to this bug:
  https://chromium.googlesource.com/chromium/src.git/+/bae6b3bcfa9dccd18b061beafcb17d0dbd39a7c9

commit bae6b3bcfa9dccd18b061beafcb17d0dbd39a7c9
Author: hjd <hjd@chromium.org>
Date: Wed May 17 04:18:05 2017

Fill in PlatformPrivateFootprint on Linux 2/2

This CL populates the vm_swap_bytes field of the PlatformPrivateFootprint
struct on Linux.

The struct contains platform specific numbers that will be used by the
memory-infra service to compute the private memory footprint of the
process.

BUG= 707019 

Review-Url: https://codereview.chromium.org/2866723002
Cr-Commit-Position: refs/heads/master@{#472321}

[modify] https://crrev.com/bae6b3bcfa9dccd18b061beafcb17d0dbd39a7c9/base/process/process_metrics.h
[modify] https://crrev.com/bae6b3bcfa9dccd18b061beafcb17d0dbd39a7c9/base/process/process_metrics_linux.cc
[modify] https://crrev.com/bae6b3bcfa9dccd18b061beafcb17d0dbd39a7c9/components/tracing/common/process_metrics_memory_dump_provider.cc

Are we always using fast method? I think slow method undercounts dalvik regions (/dev/ashmem/dalvik-), which are essentially anonymous. See implementation of MemMap::MapAnonymous(use_ashmem=true) in art/runtime/mem_map.cc.

Treating "/dev/ashmem/dalvik-" similarly to regions NOT starting with "/" in the script from the document reduces difference between fast and slow methods to 856K (from 2608K).

We use the fast method to compute PMF:

https://cs.chromium.org/chromium/src/services/resource_coordinator/public/cpp/memory_instrumentation/os_metrics_linux.cc?l=201