3 ms·

I've seen internal numbers from $MEGACORP that suggest large-scale C++ applications spend 15-20% of their time doing memory management. Incidentally, the memory

by titzer 5y ago

I've seen internal numbers from $MEGACORP that suggest large-scale C++ applications spend 15-20% of their time doing memory management. Incidentally, the memory fragmentation on those same applications can end up as high as 40%. In GC'd languages fragmentation is typically closer to 10%.

jeffbee 5y ago
This paper gives 10% for allocation in Google's fleet, and that was before significant efficiency improvements in tcmalloc. https://static.googleusercontent.com/media/research.google.com/en//pubs/archive/44271.pdf https://static.googleusercontent.com/media/research.google.c...
- titzer 5y ago
  Thanks for the link. I am going by ballparks I've heard internally, from less than 3 years ago (so, with newer allocators), and unfortunately can't provide a reference. It would be great if there was more public information from large-scale studies.
- ahh 5y ago
  https://research.google/pubs/pub50370/ https://research.google/pubs/pub50370/ gives more up to date information, as does the slide deck here: https://docs.google.com/presentation/d/1taP_gqeRE1XGB2td7f32izso0n2KD6o8QhDDvNDHdVk/edit?usp=sharing https://docs.google.com/presentation/d/1taP_gqeRE1XGB2td7f32... (We get down to 4% on the graph on slide 3.) It's far below 10% now (thanks to the hard work of a lot of people), and Google actually gotten to the point where we chose to spend _more_ time in malloc to make overall efficiency higher.
ncmncm 5y ago
You can spend literally as much time as you like on memory management. Google has chosen 20%. They have also, by banning exceptions, chosen to spend a very substantial part of their time on checking for error status, and on two-phase initialization and cleanup. I most typically choose 0% post start-up, with 0% fragmentation. For a different problem domain, I might choose differently.
- titzer 5y ago
  So you're running programs without dynamic memory allocation? That makes sense for a certain domain, like microcontrollers and embedded systems, but most programs do end up dynamically allocating memory, which leads to non-zeroes for both of those things.
  - ncmncm 5y ago
    My programs mmap thousands of hugepages at startup, then run for years. Well, not all of them.

by titzer 5y ago

jeffbee 5y ago

This paper gives 10% for allocation in Google's fleet, and that was before significant efficiency improvements in tcmalloc. https://static.googleusercontent.com/media/research.google.com/en//pubs/archive/44271.pdf https://static.googleusercontent.com/media/research.google.c...

titzer 5y ago
Thanks for the link. I am going by ballparks I've heard internally, from less than 3 years ago (so, with newer allocators), and unfortunately can't provide a reference. It would be great if there was more public information from large-scale studies.
ahh 5y ago
https://research.google/pubs/pub50370/ https://research.google/pubs/pub50370/ gives more up to date information, as does the slide deck here: https://docs.google.com/presentation/d/1taP_gqeRE1XGB2td7f32izso0n2KD6o8QhDDvNDHdVk/edit?usp=sharing https://docs.google.com/presentation/d/1taP_gqeRE1XGB2td7f32... (We get down to 4% on the graph on slide 3.) It's far below 10% now (thanks to the hard work of a lot of people), and Google actually gotten to the point where we chose to spend _more_ time in malloc to make overall efficiency higher.

ncmncm 5y ago

You can spend literally as much time as you like on memory management. Google has chosen 20%. They have also, by banning exceptions, chosen to spend a very substantial part of their time on checking for error status, and on two-phase initialization and cleanup. I most typically choose 0% post start-up, with 0% fragmentation. For a different problem domain, I might choose differently.

titzer 5y ago
So you're running programs without dynamic memory allocation? That makes sense for a certain domain, like microcontrollers and embedded systems, but most programs do end up dynamically allocating memory, which leads to non-zeroes for both of those things.
- ncmncm 5y ago
  My programs mmap thousands of hugepages at startup, then run for years. Well, not all of them.