On Monday, 3 June 2013 at 18:07:57 UTC, Diggory wrote:
Great talk! Would love to see the improvements to phobos
suggested.
An idea for the virtual address space problem on 32-bit:
- Assuming each stack has a marker page at the end to prevent
overflow, by simply exchanging the stack memory with a separate
block of memory you can reduce the number of marker pages
saving up to one page per fiber
- Could use some fast compression method to save a not recently
used stack in memory
- As a last resort can save stack to a file and read it in
again when it is required. Since events are queued the event
loop can easily peek ahead in the queue and start loading in a
stack early so that it is ready by the time that event gets to
the front.
Thanks!
One of the main advantages of fibers over threads is the low
overhead for switching - basically you save registers and change
ESP. By comparison, switching to another thread requires an OS
kernel system call. Copying entire stacks might negate this
performance benefit. It's worth noting that it looks like the
average stack size will grow for D programs in the future, as the
push is made to minimize heap allocations throughout Phobos and
use the stack more.
