FreeBSD, GPGPU and OpenCL/CUDA
cbergstrom at pathscale.com
Wed Aug 18 21:26:01 UTC 2010
emorras at xroff.net wrote:
> Hi Christopher,
> "C. Bergström" <cbergstrom at pathscale.com> escribió:
>> PathScale and CAPS recently announced HMPP as a new manycore GPGPU
>> open standard and there's a chance you'll see it ported to FBSD. In
>> addition to this you could see other open standards working well on
>> FBSD, but that depends on market demand and feedback.
> Do we need with this suite a nvidia/ati driver that executes the
> CUDA/OpenCL/Stream code? If yes, we'll have the same problem.
No, PathScale has a full NVIDIA replacement. From front-end programming
model to kernel driver. (I'm happy to give more information, but don't
want to spam the list)
>>> I havn't looked deeper into AMDs offerings, but I guess since it's
>>> silent around OpenCL and AMD-based GPGPU, even with Linux there
>>> isn't much.
>>> I'm not very close to the GPGPU scene, we even start thinking about
>>> porting and developing some mathematical stuff into libraries and
>>> thought about OpenCL.
>>> Unluckily, in my team I'm the only one utilizing FreeBSD.
>>> Maybe someone out here has solved some problems and could email me.
>>> Even AMD seems to be a white spot in the subject of GPGPU and
>>> FreeBSD for me, maybe someone could shed some light on this.
>> What's blocking this from being available now
>> a) someone porting the kernel driver over or
>> b) us getting funding to do it.
> Perhaps the FreeBSD Foundation can open a new project for it. Today it
> can be seen like a "lost-time-addon" for FreeBSD, but not only
> Maths/Physics/Chemistry can use GPGPU, it can be used by databases,
> compilers, servers, and more in a nearer future. For example, all
> algorithms to filter image and video (Scanner, PET, Astronomy, video
> de/compression, etc) are being ported to gpgpu and i can't use FreeBSD
> for this.
>> When we started working on the driver months ago one of the main
>> goals was to allow greater portability. Details for any interested
>> developers is available any time. In general I'd like to see more
>> open source OS diversity in the HPC industry and happy to help where
>> I can.
> FreeBSD has better OpenMP capabilities by its network connections,
> i'll like to use it in the next HPC era.
MPI is typically dependent on the network not OpenMP. OpenMP 3.0 can be
made more scalable if there's tasks built-into the kernel that can be
cleanly exposed to userland. (Like OpenSolaris + libtask from Moinak is
a good example)
Anyway... imho FreeBSD has a number of issues before it can be suitable
1) Better vendor support for 3rd party and open source tools (Allinea,
Totalview, undodb.. compilers, optimized math libs, profilers etc)
2) HPC ready compiler.. (Sorry guys, but LLVM is just not production
ready for this task and is missing Fortran)
3) IB network drivers
4) Hardware vendor to deliver a complete solution + support (iXsystems?)
More information about the freebsd-questions