799641 - Include io info in profiles

Status: RESOLVED FIXED

(Core :: Gecko Profiler, defect)

Description

[(dormant account)]

Windows, Linux and possibly others provide various ways to get amount of io done by process. Would be good to poll that during profiling and show an icon if a threshold is exceeded. See http://mxr.mozilla.org/mozilla-central/source/toolkit/components/telemetry/TelemetryPing.js#192 for an example of doing it via Windows API.

On linux you can count page-faults(getrusage) and read amount of io via some /proc file(lame I know).

Comment 1

[Mike Hommey [:glandium]]

We've discussed ways to include extra data in the profiler data with Benoit during the snappy work week, and there are a bunch of data points i'm planning to add, page faults being one of them.

As said on irc, for b2g/android, the /proc info is useless because they are block i/o, and yaffs (the filesystem most commonly used on android devices) doesn't do block i/o. There are much better APIs in linux to get more interesting i/o info, but they require accounting flags that are usually not enabled on android phones.

Comment 2

[Benoit Girard (:BenWa)]

I've been contemplating having a dynamic library that is crafted to interpose interesting function using LD_PRELOAD. We could have the extension force firefox to restart with this library preloaded. We could use it to intercept interesting IO function.

Comment 3

[Mike Hommey [:glandium]]

You can't intercept mmap induced i/o, and intercepting read/write doesn't give you any clue as to whether the I/O is actually hurting or not (if you open()/read()/close() the same thing twice in a row, the second time doesn't hit storage)

Comment 4

[(dormant account)]

(In reply to Mike Hommey [:glandium] from comment #3)
> You can't intercept mmap induced i/o, and intercepting read/write doesn't
> give you any clue as to whether the I/O is actually hurting or not (if you
> open()/read()/close() the same thing twice in a row, the second time doesn't
> hit storage)

we can track mmap costs on moz side(it's restricted to zip stuff and some js loading stuff). Timing io calls would still be handy

Comment 5

[Gerald Squelart (he/him) (not at Mozilla since 2022-09-15)]

We now capture a lot of IO operations, through both low-level interceptions, and higher-level labels&markers (e.g., DLL loads).
So I'll mark this bug fixed.

If more is needed, please file more focused bugs blocking bug 1527023.