这篇文章涉及mesos如何在原生的mesoscontainerizer和docker containerizer上支持gpu的,以及如果自己实现一个mesos之上的framework capos支持gpu调度的实现原理,(capos是hulu内部的资源调度平台 refer to https://www.cnblogs.com/yanghuahui/p/9304302.html)。
mesos slave在启动的时候需要初始化containerizer的resource,包含cpu/mem/gpu等,这对于mesos containerizer和docker containerizer都是通用的
void Slave::initialize() { ... Try<Resources> resources = Containerizer::resources(flags); ... }
然后到了src/slave/containerizer/containerizer.cpp 代码块中, 根据mesos-slave/agent的启动参数flags,调用allocator逻辑
Try<Resources> Containerizer::resources(const Flags& flags) { ... // GPU resource. Try<Resources> gpus = NvidiaGpuAllocator::resources(flags); if (gpus.isError()) { return Error("Failed to obtain GPU resources: " + gpus.error()); } // When adding in the GPU resources, make sure that we filter out // the existing GPU resources (if any) so that we do not double // allocate GPUs. resources = gpus.get() + resources.filter( [](const Resource& resource) { return resource.name() != "gpus"; }); ... }
src/slave/containerizer/mesos/isolators/gpu/allocator.cpp 会用nvidia的管理gpu的命令nvml以及根据启动参数,返回这台机器上gpu的资源,供之后的调度使用。
// To determine the proper number of GPU resources to return, we // need to check both --resources and --nvidia_gpu_devices. // There are two cases to consider: // // (1) --resources includes "gpus" and --nvidia_gpu_devices is set. // The number of GPUs in --resources must equal the number of // GPUs within --nvidia_gpu_resources. // // (2) --resources does not include "gpus" and --nvidia_gpu_devices // is not specified. Here we auto-discover GPUs using the // NVIDIA management Library (NVML). We special case specifying // `gpus:0` explicitly to not perform auto-discovery. // static Try<Resources> enumerateGpuResources(const Flags& flags) { ... }
因为gpu资源是需要绑定gpu卡number的,gpu资源在调度的数据结构中,是一个set<Gpu>, allocator.go提供allocate和deallocate接口的实现
Future<Nothing> allocate(const set<Gpu>& gpus) { set<Gpu> allocation = available & gpus; if (allocation.size() < gpus.size()) { return Failure(stringify(gpus - allocation) + " are not available"); } available = available - allocation; allocated = allocated | allocation; return Nothing(); } Future<Nothing> deallocate(const set<Gpu>& gpus) { set<Gpu> deallocation = allocated & gpus; if (deallocation.size() < gpus.size()) { return Failure(stringify(gpus - deallocation) + " are not allocated"); } allocated = allocated - deallocation; available = available | deallocation; return Nothing(); }
但是封装到上层,供containerizer调用的时候,指定需要allocate的gpu number就可以
Future<set<Gpu>> NvidiaGpuAllocator::allocate(size_t count) { // Need to disambiguate for the compiler. Future<set<Gpu>> (NvidiaGpuAllocatorProcess::*allocate)(size_t) = &NvidiaGpuAllocatorProcess::allocate; return process::dispatch(data->process, allocate, count); }
但是deallocate仍然需要显示指定需要释放哪些gpu
Future<Nothing> NvidiaGpuAllocator::deallocate(const set<Gpu>& gpus) { return process::dispatch( data->process, &NvidiaGpuAllocatorProcess::deallocate, gpus); }
然后如果作业是用docker containerizer,可以看到src/slave/containerizer/docker.cpp中调用gpu的逻辑
Future<Nothing> DockerContainerizerProcess::allocateNvidiaGpus( const ContainerID& containerId, const size_t count) { if (!nvidia.isSome()) { return Failure("Attempted to allocate GPUs" " without Nvidia libraries available"); } if (!containers_.contains(containerId)) { return Failure("Container is already destroyed"); } return nvidia->allocator.allocate(count) .then(defer( self(), &Self::_allocateNvidiaGpus, containerId, lambda::_1)); }
所以之上,就是在slave中启动的时候加载确认gpu资源,然后在启动containerizer的时候,可以利用slave中维护的gpu set资源池,去拿到资源,之后启动作业。
那capos是如何实现的呢,capos是hulu内部的资源调度平台(refer to https://www.cnblogs.com/yanghuahui/p/9304302.html),因为自己实现了mesos的capos containerizer,我们的做法是,在mesos slave注册的时候显示的通过参数或者自动探测的机制,发现gpu资源,然后用--resources=gpu range的形式启动mesos agent,这样offer资源的gpu在capos看来就是一个range,可以类似使用port资源的方式,来调度gpu,在capos containerizer中,根据调度器指定的gpu range,去绑定一个或者多个gpu资源到docker nvidia runtime中。完成gpu调度功能。