Data_Namespace Namespace Reference

Namespaces
	anonymous_namespace{DataMgr.cpp}
	detail

Classes
class	AbstractBuffer
	An AbstractBuffer is a unit of data management for a data manager. More...
class	TypedThrustAllocator
	a Templated version of Data_Namespace::ThrustAllocator that can be used as a custom allocator in thrust device containers such as thrust::device_vector. Note that this class derives from thrust::mr::allocator in order to meet the requirements of an Allocator More...
struct	MemoryData
struct	MemoryInfo
class	ProcMeminfoParser
	Parse /proc/meminfo into key/value pairs. More...
class	ProcBuddyinfoParser
	Parse /proc/buddyinfo into a few fragmentation-related data. More...
class	DataMgr

Typedefs
using	DataMgrAllocationPolicy = typename thrust::detail::allocator_aware_execution_policy< thrust::device_execution_policy >::execute_with_allocator_type< ThrustAllocator >::type
	This is a custom device execution policy to allow a custom allocator backed by a DataMgr instance for any temp allocations done in thrust calls. This execution policy is required when using thrust functions that do host <-> device transfers internally, such as thrust::equal_range(). Without the use of this custom policy, thrust::equal_range() will crash. For more clarity, if you were to use the thrust::device(allocator) as the execution policy for thrust::equal_range, where allocator is a ThrustAllocator instance, thrust will crash because that policy by default uses a cuda default stream wrapper class (execute_on_stream_base): More...
template<typename T >
using	ThrustAllocatorDeviceVector = thrust::device_vector< T, TypedThrustAllocator< T >>
using	CpuTierSizeVector = std::vector< size_t >

Enumerations
enum	CpuTier { DRAM = 0, PMEM = 1 }
enum	MemoryLevel { DISK_LEVEL = 0, CPU_LEVEL = 1, GPU_LEVEL = 2 }

Functions
std::ostream &	operator<< (std::ostream &os, const DataMgr::SystemMemoryUsage &mem_info)

Variables
constexpr size_t	numCpuTiers = 2

Typedef Documentation

using Data_Namespace::CpuTierSizeVector = typedef std::vector<size_t>

Definition at line 28 of file TieredCpuBufferMgr.h.

using Data_Namespace::DataMgrAllocationPolicy = typedef typename thrust::detail::allocator_aware_execution_policy< thrust::device_execution_policy>::execute_with_allocator_type<ThrustAllocator>::type

This is a custom device execution policy to allow a custom allocator backed by a DataMgr instance for any temp allocations done in thrust calls. This execution policy is required when using thrust functions that do host <-> device transfers internally, such as thrust::equal_range(). Without the use of this custom policy, thrust::equal_range() will crash. For more clarity, if you were to use the thrust::device(allocator) as the execution policy for thrust::equal_range, where allocator is a ThrustAllocator instance, thrust will crash because that policy by default uses a cuda default stream wrapper class (execute_on_stream_base):

https://github.com/thrust/thrust/blob/cuda-10.2/thrust/system/cuda/detail/par.h#L74

The get_stream friend function in that wrapper class attempts to get a stream from a wrapper that is not defined, and therefore segfaults.

So this custom device policy is a way around that problem. It ensures that execute_on_stream_base is not used as a default stream policy and instead uses the default stream.

If streams are ever intended to be used in the future for any of this thrust work, this will need to be adjusted, and likely care required in getting thrust::equal_range to work.

Note the above crash scenario was noticed while using cuda 10.2 on 06/03/2020.

Definition at line 51 of file DataMgrAllocationPolicy.h.

template<typename T >

using Data_Namespace::ThrustAllocatorDeviceVector = typedef thrust::device_vector<T, TypedThrustAllocator<T>>

Definition at line 164 of file TypedThrustAllocator.h.

Enumeration Type Documentation

enum Data_Namespace::CpuTier

Enumerator
DRAM
PMEM

Definition at line 27 of file TieredCpuBufferMgr.h.

{ DRAM = 0, PMEM = 1 };

enum Data_Namespace::MemoryLevel

Enumerator
DISK_LEVEL
CPU_LEVEL
GPU_LEVEL

Definition at line 21 of file MemoryLevel.h.

{ DISK_LEVEL = 0, CPU_LEVEL = 1, GPU_LEVEL = 2 };

Function Documentation

std::ostream & Data_Namespace::operator<<	(	std::ostream &	os,
		const DataMgr::SystemMemoryUsage &	mem_info
	)

Definition at line 712 of file DataMgr.cpp.

References Data_Namespace::DataMgr::SystemMemoryUsage::avail_pages, Data_Namespace::DataMgr::SystemMemoryUsage::frag, Data_Namespace::DataMgr::SystemMemoryUsage::free, Data_Namespace::DataMgr::SystemMemoryUsage::high_blocks, Data_Namespace::DataMgr::SystemMemoryUsage::regular, Data_Namespace::DataMgr::SystemMemoryUsage::resident, Data_Namespace::DataMgr::SystemMemoryUsage::shared, Data_Namespace::DataMgr::SystemMemoryUsage::total, and Data_Namespace::DataMgr::SystemMemoryUsage::vtotal.

                                                                                 {
  os << "jsonlog ";
  os << "{";
  os << " \"name\": \"CPU Memory Info\",";
  os << " \"TotalMB\": " << mem_info.total / (1024. * 1024.) << ",";
  os << " \"FreeMB\": " << mem_info.free / (1024. * 1024.) << ",";
  os << " \"ProcessMB\": " << mem_info.resident / (1024. * 1024.) << ",";
  os << " \"VirtualMB\": " << mem_info.vtotal / (1024. * 1024.) << ",";
  os << " \"ProcessPlusSwapMB\": " << mem_info.regular / (1024. * 1024.) << ",";
  os << " \"ProcessSharedMB\": " << mem_info.shared / (1024. * 1024.) << ",";
  os << " \"FragmentationPercent\": " << mem_info.frag;
  os << ", \"BuddyinfoHighBlocks\": " << mem_info.high_blocks;
  os << ", \"BuddyinfoAvailPages\": " << mem_info.avail_pages;
  os << " }";
  return os;
}

Variable Documentation

constexpr size_t Data_Namespace::numCpuTiers = 2

Definition at line 26 of file TieredCpuBufferMgr.h.

Referenced by Data_Namespace::DataMgr::populateMgrs(), and Buffer_Namespace::TieredCpuBufferMgr::TieredCpuBufferMgr().