Costless huge virtual memory? /dev/same, /dev/null?
From: Samuel Thibault
Date: Mon Feb 29 2016 - 11:28:43 EST
I'm wondering whether we could introduce a /dev/same device to allow
costless huge virtual memory.
The use case is the simulation of the execution of a big irregular HPC
application, to provision memory usage, cpu time, etc. We know how much
time each computation loop takes, and it's easy to replace them with a
mere accounting. We'd however like to avoid having to revamp the rest
of the code, which does allocation/memcpys/etc., by just replacing
the allocation calls with virtual allocations, i.e. allocations which
return addresses of buffers that one can read/write, but the values you
read are not necessarily what you wrote, i.e. the data is not actually
properly stored (since we don't do the actual computations that's not a
The way we currently do this is by some folding: we map the same normal
file several times contiguously to form the virtual allocation. By using
a small 1MiB file, this limits memory consumption to 1MiB plus the page
table (and fits the dumb data in a typical cache). This however creates
one VMA per file mapping, we get limited by the 65535 VMA limit, and
VMA lookup becomes slow.
The way I could see is to have a /dev/same device: when you open it, it
allocates one page. When you mmap it, it maps the same page over the
whole resulting single VMA.
This is a quite specific use case, but it seems to be easy to implement,
and it seems to me that it could be integrated mainline. Actually I was
thinking that /dev/null itself could be providing that service?
(currently it returns ENODEV)
What do people think? Is there perhaps another solution to achieve this
that I didn't think about?