New issue
Advanced search Search tips
Note: Color blocks (like or ) mean that a user may not be available. Tooltip shows the reason.

Issue 646103 link

Starred by 2 users

Issue metadata

Status: Duplicate
Owner: ----
Closed: Sep 2016
Cc:
Components:
EstimatedDays: ----
NextAction: ----
OS: Linux
Pri: 3
Type: Bug



Sign in to add a comment

CommandExecutor:PutChanged for DisplayCompositor taking 4s

Project Member Reported by boliu@chromium.org, Sep 12 2016

Issue description

Filing for gilbertleung@. This is on this desktop linux machine where tasks on the gpu thread are taking 4s. Attached trace and about:gpu

This is 53.0.2785.101

afaict this is just drawing the browser chrome, not some complex page causing slow down, since things are already hosed before any renderer produced new frames (I think?)

not sure if any other trace category will help here..
 
trace_gilbertleung.json.gz
3.2 MB Download
gilbertleung_chrome___gpu.pdf
425 KB Download

Comment 1 by piman@chromium.org, Sep 12 2016

Can you run 'nvidia-smi -q' and attach the output?
Attached is the output of nvidia-smi -q
nvidia
5.2 KB View Download

Comment 3 by piman@chromium.org, Sep 12 2016

Mergedinto: 623786
Status: Duplicate (was: Untriaged)
As I suspected, this happens because the GPU is out of memory:

    FB Memory Usage
        Total                       : 1023 MiB
        Used                        : 1007 MiB
        Free                        : 16 MiB


This results in trashing, similar to when out of memory and pages get swapped in and out to disk.


Oftentimes, we observed a lot of the GPU memory to be consumed by the desktop compositor (e.g. compiz) - possibly because of leaks there - and there's not a whole lot we can do if that's the case. One way to tell is to open the Chrome task manager, enable the "GPU memory" column (right-click -> check "GPU memory"), and look at the value under "GPU Process". That will be the amount we believe we use.


There are several similar bugs, will duplicate into one of them - feel free to follow the discussion there, and/or post additional information.

I'm seeing about 90,000-10,000K under the GPU RAM column. Is that standard?

Comment 5 by piman@chromium.org, Sep 13 2016

@#4: is that for the GPU process, or for individual tabs?

The data is slightly confusing: only the GPU process allocates GPU memory, and the amount reported for "GPU process" is the actual amount of GPU memory that is allocated (it has a "*" next to it). We further break it down by associating that GPU memory to the clients that caused that allocation, and that's what is reported on the other lines (if you do the math, you'll see that the amount for the GPU process is equal to the sum of amounts for all other processes).

We try to release memory for offscreen tabs, but keep up to 5 in LRU fashion for fast tab switching. The allocation for those is about the tab's dimensions worth of pixels. For a 30" screen at 2560x1600, a fullscreen tab is about 16MB of memory. We also have memory for the visible tabs (1 per browser window), and this can be more than the screen worth because we pre-render for smooth scrolling. So for long pages, it's common to see those use more than a screen worth. There's also some GPU-intensive pages (e.g. Google Maps) as well as pathological cases that need more memory. 90K is fairly high but not unheard of.

Sign in to add a comment