I recently landed two patches [1], [2] that should make detecting system level memory regressions a bit easier. The system memory reporter now breaks out kgsl (system level graphics memory allocations) and ION (kernel/device memory allocations) into per-process chunks. ION and kgsl details are primarily of importance for Flame devices right now, but in the future I expect more devices to use these features.

This means that when you load a memory report generated by |get_about_memory.py| you'll find new entries under the 'System' entry similar to the following:

11.96 MB (100.0%) -- ion-memory
├───5.63 MB (47.04%) ── Homescreen (pid=2080) [20]
├───3.16 MB (26.46%) ── b2g (pid=1921) [2]
├───1.58 MB (13.23%) ── fd900000.qcom,mdss_mdp (pid=1)
├───1.58 MB (13.23%) ── mdss_fb0 (pid=1927)
└───0.00 MB (00.03%) ── adsprpc-smd (pid=1)

13.14 MB (100.0%) -- kgsl-memory
└──13.14 MB (100.0%) -- b2g (pid=1921)
   ├───3.89 MB (29.60%) ── egl_image [10]
   ├───3.16 MB (24.07%) ── egl_surface [2]
   ├───2.54 MB (19.32%) ── gl [16]
   ├───1.66 MB (12.63%) ── texture [12]
   ├───0.94 MB (07.13%) ── command [15]
   ├───0.70 MB (05.35%) ── any(0) [26]
   └───0.25 MB (01.90%) ++ (4 tiny)

In the above example we can see that the device was using ~12MB total of ION allocations. It further breaks out that Homescreen is using 5.63MB of ION across 20 allocations and b2g is using 3.16MB across 2 allocations.

On the kgsl side you can see the device is using a total of ~13MB of kgsl and a breakdown of how graphics memory is being used by the b2g process.

These values can be particularly useful for evaluating the efficacy of attempts to reduce memory usage. I'm nominating the bugs for 2.0+ in hopes of improving our ability to detect memory regressions across versions of B2G.

-e

[1] - https://hg.mozilla.org/mozilla-central/rev/0b1d53fd456b
[2] - https://hg.mozilla.org/integration/mozilla-inbound/rev/eeccb95d2e83
_______________________________________________
dev-b2g mailing list
[email protected]
https://lists.mozilla.org/listinfo/dev-b2g

Reply via email to