Skip to content

revert pinned hostmem for pod comm #292

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
AtlantaPepsi wants to merge 2 commits into
base: candidate
Choose a base branch
from
+33 −10
Open

revert pinned hostmem for pod comm #292

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
19b0ae9
revert pinned hostmem for pod comm
AtlantaPepsi May 8, 2026
634e46e
minor fix
AtlantaPepsi May 14, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
43 changes: 33 additions & 10 deletions src/header/TransferBench.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -615,6 +615,7 @@ namespace TransferBench
#define hipMemGenericAllocationHandle_t CUmemGenericAllocationHandle
#define hipMemAccessDesc CUmemAccessDesc
#define hipMemFabricHandle_t CUmemFabricHandle
#define hipMemLocation CUmemLocation

// Enumerations
#define hipDeviceAttributeClockRate cudaDevAttrClockRate
Expand All @@ -628,6 +629,7 @@ namespace TransferBench
#define hipMemcpyHostToDevice cudaMemcpyHostToDevice
#define hipSuccess cudaSuccess
#define hipMemLocationTypeDevice CU_MEM_LOCATION_TYPE_DEVICE
#define hipMemLocationTypeHostNuma CU_MEM_LOCATION_TYPE_HOST_NUMA
#define hipMemAllocationTypePinned CU_MEM_ALLOCATION_TYPE_PINNED
#define hipMemHandleTypeFabric CU_MEM_HANDLE_TYPE_FABRIC
#define hipMemAllocationGranularityRecommended CU_MEM_ALLOC_GRANULARITY_RECOMMENDED
Expand Down Expand Up @@ -1411,6 +1413,25 @@ namespace {
}

#ifdef POD_COMM_ENABLED
static ErrResult GetMemLocation(MemDevice const& memDevice, hipMemLocation& location)
{
if (IsCpuMemType(memDevice.memType)) {
location.type = hipMemLocationTypeHostNuma;
} else if (IsGpuMemType(memDevice.memType) && memDevice.memType != MEM_MANAGED) {
location.type = hipMemLocationTypeDevice;
} else {
return {ERR_FATAL, "Unsupported memory location"};
}

// Determine location id
if (memDevice.memType == MEM_CPU_CLOSEST) {
location.id = GetClosestCpuNumaToGpu(memDevice.memIndex);
} else {
location.id = memDevice.memIndex;
}
Comment thread
AtlantaPepsi marked this conversation as resolved.
return ERR_NONE;
}

static ErrResult GetMemAllocationProp(MemDevice const& memDevice, hipMemAllocationProp& prop)
{

Expand All @@ -1428,10 +1449,7 @@ namespace {
}

prop.requestedHandleTypes = hipMemHandleTypeFabric;
// at this point shouldn't have any memtype other than device
// ERR_CHECK(GetMemLocation(memDevice, prop.location));
prop.location.type = hipMemLocationTypeDevice;
prop.location.id = memDevice.memIndex;
ERR_CHECK(GetMemLocation(memDevice, prop.location));
return ERR_NONE;
}
#endif
Expand Down Expand Up @@ -1519,19 +1537,20 @@ namespace {

// Specify memory access descriptor to enable local read/write
hipMemAccessDesc desc;
// ERR_CHECK(GetMemLocation(memDevice, desc.location));
desc.location.type = hipMemLocationTypeDevice;
desc.location.id = memDevice.memIndex;
ERR_CHECK(GetMemLocation(memDevice, desc.location));
desc.flags = hipMemAccessFlagsProtReadWrite;

// Set access flags for virtual address range
ERR_CHECK(hipMemSetAccess((gpu_device_ptr)*memPtr, roundedUpBytes, &desc, 1));

// Clear the memory
if (IsCpuMemType(memType)) {
// Note: CheckPages() / move_pages() is intentionally NOT called here.
// For fabric-exportable HOST_NUMA memory the VA is owned by the driver
// (not a normal anonymous mmap VMA), so move_pages() returns
// -EFAULT/-EINVAL and would falsely trip a fatal error. NUMA placement
// should be already enforced by the prop.location passed to hipMemCreate().
memset(*memPtr, 0, roundedUpBytes);
// Check that the allocated pages are actually on the correct NUMA node
ERR_CHECK(CheckPages((char*)*memPtr, roundedUpBytes, deviceIdx));
} else if (IsGpuMemType(memType)) {
ERR_CHECK(hipSetDevice(memDevice.memIndex));
ERR_CHECK(hipMemset(*memPtr, 0, numBytes));
Expand Down Expand Up @@ -4075,7 +4094,9 @@ static bool IsConfiguredGid(union ibv_gid const& gid)
hipError_t exportErr = hipSuccess;
const char* exportStep = "hipSetDevice";
if (memDevice.memRank == GetRank()) {
exportErr = hipSetDevice(memDevice.memIndex);
if (IsGpuMemType(memDevice.memType)) {
exportErr = hipSetDevice(memDevice.memIndex);
}
Comment on lines +4097 to +4099
Copy link
Copy Markdown
Collaborator

@gilbertlee-amd gilbertlee-amd May 16, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Why is this hipSetDevice even necessary? The runtime should already have the ability to determine where the memory handle is. It also seems strange that you would not need it for CPU memory types but GPU memory types.

Based on the documentation, cuMemExportToSharableHandle doesn't require the current device context either, and won't ever return CUDA_ERROR_INVALID_CONTEXT

if (exportErr == hipSuccess) {
exportStep = "hipMemExportToShareableHandle";
exportErr = hipMemExportToShareableHandle(&fabricHandle, *memHandle, hipMemHandleTypeFabric, 0);
Expand Down Expand Up @@ -8064,6 +8085,7 @@ static bool IsConfiguredGid(union ibv_gid const& gid)
#undef hipMemGenericAllocationHandle_t
#undef hipMemAccessDesc
#undef hipMemFabricHandle_t
#undef hipMemLocation

// Enumerations
#undef hipDeviceAttributeClockRate
Expand All @@ -8077,6 +8099,7 @@ static bool IsConfiguredGid(union ibv_gid const& gid)
#undef hipMemcpyHostToDevice
#undef hipSuccess
#undef hipMemLocationTypeDevice
#undef hipMemLocationTypeHostNuma
#undef hipMemAllocationTypePinned
//#undef hipMemAllocationTypeUncached
#undef hipMemHandleTypeFabric
Expand Down