On Tue, Apr 17, 2018 at 2:08 PM, Thierry Goubier <thierry.goub...@gmail.com> wrote:
> 2018-04-17 14:03 GMT+02:00 Clément Bera <bera.clem...@gmail.com>: > > Hi Pavel, > > > > I'm looking at PosixSharedMemory again since I have to write a Master > > student proposal and I think this could be a good topic. I'm not really > > expert on SharedMemory, so I'm going to share what I have in mind, we can > > discuss about it, you could also co-supervise the student indirectly > (though > > officially the supervisor has to be VUB staff). > > > > Implementation-wise, I was thinking what could be interesting are: > > - Making PosixSharedMemory compatible with TaskIt so from you image you > can > > create a SharedMemory buffer, spawn another image+VM and attach it to the > > shared memory to have multiple threads working on the shared section. > > - Add the #at:if:put: primitive, which write into the shared memory using > > compare and swap instruction for efficient thread-safe access. > > - Add on SharedMemory all the primitives to read/write native types to > the > > buffer (int64, double, etc.) with CAS and non CAS instructions. > > - Maybe add APIs to read/write objects through Fuel to pass them by copy, > > though this looks difficult in some cases. > > - Implement a lock system or a semaphore system on top of the CAS > > - implement lock-free and lock-full algorithm using CAS and non CAS > > instructions (I think a first try would be parallelSort on a 1Gb buffer > of > > int32 with 4 native threads) > > > > What do you think ? Do you have ideas ? Are you interested ? Do you thing > > having a student on this would be nice ? > > > > The master thesis proposal has to include a research question. I am not > sure > > what other languages do regarding shared memory.It's not clear so far > what > > the research question is. > > I can forward that to a researcher here working on distributed shared > memory. > > Research question can be: > - heterogeneity (x86 + pi at the same time) > - load balancing between competing images on heterogeneous hardware > ... > - Object migration (pointer forwarding). Probably not, state of the > art is very advanced on that which means a costly implementation to > reach parity. > > SharedMemory on heterogenous hardware ? Do you mean you need to physically plug the memory into a Raspberry Pie and an x86 computer ? Or you mean exporting the RAM of one hardware as NFS to the others ? I was just thinking sharing the memory between multiple pairs of image+VM on the same machine to be able to run some multi-threaded algorithm on the shared buffer. I know it's not much but we need to start somewhere. The student would have 6 months, including 1 month to write the thesis so it cannot be too heavy. Something that we can re-use with a minor research contribution would be nice. Thierry > > > > > Best, > > > > On Mon, Feb 12, 2018 at 10:33 AM, Pavel Krivanek < > pavel.kriva...@gmail.com> > > wrote: > >> > >> Hi, > >> > >> among other less interesting things, I spent some time on existing > >> PosixSharedMemory project. It is a UFFI binding for the LibC methods > >> that provide support for the memory allocation between several > >> separate processes. I significantly improved the performance by > >> implementing the block access. Writing of 10MB byte array takes about > >> 1 millisecond, reading of it from other image took me about 4 > >> milliseconds. While serialization with Fuel is very fast, it opens > >> interesting possibilities. > >> To have a shared memory without synchronization tools is not very > >> useful so I wrote a basic UFFI interface for the POSIX named > >> semaphores. They are quite easy to use and work nicely with Pharo. The > >> VM can all wait on the semaphore or it can check the status of it > >> periodically in an image thread. It has two small disadvantages. It > >> requires to dynamically link the next library (pthread) and they must > >> be cleaned manually. I plan to look at System V alternative in future. > >> Now we should write a nice framework for inter-image communication on > >> top of it or/and adopt Seamless for it ;-) > >> > >> Cheers, > >> -- Pavel > >> > > > > > > > > -- > > Clément Béra > > https://clementbera.github.io/ > > https://clementbera.wordpress.com/ > > -- Clément Béra https://clementbera.github.io/ https://clementbera.wordpress.com/
