> Is all code loaded into memory as one lump on B2G, or is it paged in as > needed? There may be quite a bit of code in there that's not commonly needed.
I don't know how it works, to be honest. I can't imagine that library pages are loaded only on demand with no pre-fetching, since that would be slow. And I imagine once a code page is in memory it's not going out. Our devices don't have swap, and I've never observed them dropping clean code pages (which don't have to go into a swap file) to make room for data. -Justin On Mon, Dec 3, 2012 at 8:04 PM, Norbert Lindenberg <mozillali...@lindenbergsoftware.com> wrote: > > On Dec 3, 2012, at 16:15 , Justin Lebar wrote: > >> On Mon, Dec 3, 2012 at 5:39 PM, Norbert Lindenberg >> <mozillali...@lindenbergsoftware.com> wrote: >>> >>> Well, the first question is what size increase would be acceptable given >>> the benefits that ICU provides. >> >>> I have currently trimmed it to 9.7 MB for the data library and 3.1 MB for >>> two code libraries. >> >> Ignoring download size for a moment, I want to consider the memory >> usage of this at runtime. > > Memory usage on small devices certainly warrants some investigation and > discussion. Unfortunately, I don't have real data yet. > >> Does all of this data need to be loaded into memory? > > Most of the data is locale data for several hundred locales, separated by > locale and functionality, so as long as applications don't request a specific > locale/functionality combination, it doesn't need to be loaded. Note though > that the size of the data isn't uniform per locale - e.g., Chinese collation > data is huge. > >> If so, 13mb of code + data is likely unacceptable to B2G. That's >> roughly 10% of all memory we have available to gecko on our devices. >> I would lobby hard against turning this on for B2G. > > Understood. Now, how does B2G support internationalization in the absence of > ICU? > >> If we can avoid loading most of that data into memory, the situation >> is much better. But even 3mb of code is dicey; we consider a 3mb >> memory win to be substantial and worthy of a large amount of effort to >> obtain. > > Is all code loaded into memory as one lump on B2G, or is it paged in as > needed? There may be quite a bit of code in there that's not commonly needed. > The ICU documentation suggests static linking as a way to reduce code size - > I haven't tried yet how much that would help. > >> I'd imagine that the Fennec folks working on project 256mb [1] would >> have similar reactions. >> >> On Windows desktop, our median memory usage is ~500mb, and the 5th >> percentile is ~175mb, so an extra 13mb, while not great, might be >> acceptable. 3mb wouldn't be a big deal. > > Good to know. > > Norbert > _______________________________________________ dev-platform mailing list dev-platform@lists.mozilla.org https://lists.mozilla.org/listinfo/dev-platform
