guetzli
guetzli copied to clipboard
google
Not sure this is an issue but a request
I did a comparison to run guetzli on a GCP Ubuntu server (n1-standard-8 (8 vCPUs, 30 GB memory) or on my laptop. The compress time is almost the same on both. And the CPU on the server is used only 12% measured from top. Parallel process seems the only way to utilize the server's processing power to speed up. Any suggestion to speed up on the powerful server? From this particular server type, each image can only takes about 12% CPU regardless of the images size from 100k to 4M. And memory usage is also low, I would assume the usage is about the same size as the images. The processing speed is main concern. If anyone can provide some configuration which can speed up the processing, I would greatly appreciate it
Thank you very much for your help!
Please see the screenshot above for more detail. CPU can be used with batching process but memory does not seem to be. Not sure whether that can be one of area to speed up the process. Currently it is fairly slow.
I also noticed this. Essentially it looks like Guetzli is single-threaded.
I believe Guetzli tries a number of encodes, scores them on their visual fidelity against the original, and picks the best one. It seems to me that a good multi-threaded version ought to try each encode (and potentially each Butteraugli scoring task) on its own thread. Or have a jobs flag like GNU make has, eg -j 8 for 8 threads, -j 4 for 4 threads.
I think it would increase memory usage by a lot (just a guess), but I'm not sure. Not a problem for machines with lots of RAM, though, and there already a memory-limiting flag.
