> On May 23, 2015, at 4:46 AM, Joel Rees <joel.r...@gmail.com> wrote: > > >> Warner suggested that emulated hardware could be faster than the native > hardware being used. If emulated hardware were twice as fast, would that > justify using it in addition to native builds? >> > > If only such theoretical speed advantage were real, it might be worth > talking about. Maybe. On the other hand, the openbsd project never used > emulation on the 68k platforms, where the clockspeed difference was very > real.
The speed advantage is real. To build the same set of packages natively for amd64 it takes 12 or so hours. To build them using user-level emulation of the native binaries takes about 30 hours, though some of the binaries are native producing target binaries. To build 100% native on something like a Wandboard takes more like 20-30 days, since swapping hurts a lot... The advantage for this is two fold: memory and core count. While you can do something about the core count by running lots of boards in parallel, that takes a more complex build system.You can’t do much about memory, however. When you have a machine with 256GB of RAM, clang runs much better than it could ever run on an arm board with 2GB of RAM. The native binaries help, and the user-level emulation copes with those silly programs that expect to run binaries in the target environment to produce their output. But hey, I’m not trying to tell anybody what to do. I’m just giving you data from what a sister project has been able to do lately. Warner
signature.asc
Description: Message signed with OpenPGP using GPGMail
