> On Feb. 11, 2015, 1:23 a.m., Timothy Chen wrote: > > So after looking at this review, I'm not sure we really need to add > > anything to Mesos, as getting graphic compute units and memory could be a > > seperate process/script that just injects the resources via the slave > > flags. On the framework side it could simply look up the resources by name, > > since the methods you added is only relevant if frameworks links with Mesos > > itself > > chester kuo wrote: > Use script/process to inject isn't a good idea , since it only treat the > gpgpu or computing device as sepearte one , but some protocol need host > /device side to work together , which is mean we need a cpu and other > resources, and still takes host side resources if run. let's why i add into > mesos framework. > > Ya, this is first draft to have fundamental support for generic gpgpu > support, i'm planning to put more focus on the framework/execator and back to > extend more feautres into mesoe/slave in the future. > > chester kuo wrote: > BTW, One more thing, the computing device could be CPU, GPGPU, APU, > FPGA..etc , if computing device selecte a CPU as computing unit , the slave > need to adjust its resource offer automatically. > > Timothy Chen wrote: > Hi Chester, I don't really understand your first statement of host / > device side working together. When you provide additional resources through > the slave flags (--resources="gcompute_units(*):40,gmemory(*):1gb"), in the > resource offer from the slave it will include the slave's CPU, Memory, Ports > and also gcompute_units and gmemory that the framework can choose to use. > > Are you saying that you want to replace cpu/memory with gpu resources > when they're available, instead of just adding to the existing offers?
Hi Tim I'm so sorry didn't provide a clear pictures here , but here is JIRA i created (https://issues.apache.org/jira/browse/MESOS-2262) to support Heterogeneous computing unit. This is first draft (POC) i created to make sure the Heterogeneous computing env did work with Mesos, which is why i said need to more work to finish this task as Mesos code change often and like to provide feature/patches as early as possible. For some Heterogeneous computing env , when you launch a task , you need to not only a device kernel code(ex: run on GPGPU) but also host side (ex: CPU) programming environment ready, so we need a full pictures of slave resources, besides , the Heterogeneous device can be a CPU core/device and may need to exported if supported on some programming model, so we may need to adjust host's resources offer dynamicly. This patches don't provide those features yet, i belive this is confusing parts , an applogy for this. Chester - chester ----------------------------------------------------------- This is an automatically generated e-mail. To reply, visit: https://reviews.apache.org/r/30736/#review71897 ----------------------------------------------------------- On Feb. 6, 2015, 5:33 p.m., chester kuo wrote: > > ----------------------------------------------------------- > This is an automatically generated e-mail. To reply, visit: > https://reviews.apache.org/r/30736/ > ----------------------------------------------------------- > > (Updated Feb. 6, 2015, 5:33 p.m.) > > > Review request for mesos, Adam B and Timothy Chen. > > > Bugs: MESOS-2262 > https://issues.apache.org/jira/browse/MESOS-2262 > > > Repository: mesos > > > Description > ------- > > Extending Mesos to support Heterogeneous resource such as GPGPU/FPGA..etc as > computing resources in the data-center, OpenCL will be first target to add > into Mesos. > > This is first draft, some slave resouce discover not ready but i plan to work > on a framework first so we can have people to test it early. > > > Diffs > ----- > > configure.ac acc685caa47717f0fb61a7c18572d7b763ec707b > include/mesos/resources.hpp 3b57568c10233a0c692787de6464f21af5eaadf4 > src/Makefile.am 93537d17d3c7604a8532ee1453e405630c481ddc > src/common/resources.cpp 98371f6873482d0cdbefeb279b58ae6cc680a88f > src/examples/test_opencl_executor.cpp PRE-CREATION > src/examples/test_opencl_framework.cpp PRE-CREATION > src/slave/containerizer/containerizer.cpp > cf2ece80252350c25db38116e29cf28a3164662d > src/tests/master_allocator_tests.cpp > 1eebefd2e423e4bb89d76ed7b7d8acc9d1bb7760 > src/tests/master_tests.cpp 62ba35b9c3f999c59a95bffb01b8b82cc543a34f > > Diff: https://reviews.apache.org/r/30736/diff/ > > > Testing > ------- > > make check > & > I also create a simple frame test (test-opencl-framework so it can have a > quick smoke test) > > > Thanks, > > chester kuo > >
