Optimize receive buffer #1281
Optimize receive buffer #1281
Conversation
jalfd
force-pushed
the
optimize-receive-buffer
branch
from
fe01fcf to
df50d34
Compare
Previously, we would compact the buffer (moving unread data to the start of the buffer) as follows: - after processing a message, if there are zero unread bytes, just reset the indices for first and last unread byte to zero - else, if at least 1/8th of the buffer is used, copy remaining data to the beginning of the buffer The second option is never actually necessary, as before inserting new data into the array, we already check if there's enough free space, and compact the buffer first if necessary. So we've been doing a lot of copies that weren't actually needed. Let's not do that any more.
When compacting the receive buffer, we should only copy the bytes between _rQi and _rQlen (the index of the first unread byte, and the next write position). Previously, we copied everything for _rQi up untill the end of the buffer.
jalfd
force-pushed
the
optimize-receive-buffer
branch
from
df50d34 to
08567b0
Compare
There was a problem hiding this comment.
Apologies for the delay.
Very nice! I have to commend you for good code, helpful comments, and great commit messages. Thank you for your work.
I've looked through the commits and tested the code in IE 11 on Windows 10 with a 4k monitor. And I can see a slight improvement, very nice.
|
Thanks for the kind words. :)
FYI, I should mention that because noVNC currently downscales the canvas to compensate for your DPI scaling, you won't see the full benefit unless you set the scaling factor in Windows to 100%. If you do that, you get a canvas matching the size of your actual desktop resolution, and then the difference in performance becomes very noticeable. That's how we became aware of the problem. |
|
Yeah, by default the scaling factor in Windows was 200%, before testing this I changed this to 100%. I guess the client machine I tested on was a bit too powerful to notice a big difference. I never did any profiling etc. |
Fix two inefficiencies in how and when we compact the websocket receive buffer.
In most browsers, the impact of this is negligible, but on a high-res monitor in IE11, these issues could freeze the buffer for over half a second when a large FBU is received.
With these changes, compactions effectively disappear from IE11's profiler graph.