Skip to content

disable pinned host memory for pod #318

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

Merged
AtlantaPepsi merged 2 commits into from
Jun 1, 2026
Merged

disable pinned host memory for pod #318

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
9da7db7
disable pinned host memory for pod
AtlantaPepsi Jun 1, 2026
8368f0c
update message
AtlantaPepsi Jun 1, 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
34 changes: 31 additions & 3 deletions src/header/TransferBench.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1429,8 +1429,6 @@ 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;
return ERR_NONE;
Expand Down Expand Up @@ -1529,7 +1527,6 @@ 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;
desc.flags = hipMemAccessFlagsProtReadWrite;
Expand Down Expand Up @@ -2584,6 +2581,37 @@ namespace {
i, t.exeDevice.exeRank});
break;
}

// Pod (cross-rank) transfers with a GPU executor are exchanged via CUDA/HIP fabric handles,
// which, for current version, only support device backed allocations.
// Reject host memory allocations up front.
if (IsGpuExeType(t.exeDevice.exeType)) {
bool hasRemoteCpuMem = false;
MemDevice offender = {};
char const* role = nullptr;
for (auto const& src : t.srcs) {
if (src.memRank != t.exeDevice.exeRank && IsCpuMemType(src.memType)) {
hasRemoteCpuMem = true; offender = src; role = "SRC"; break;
}
}
if (!hasRemoteCpuMem) {
for (auto const& dst : t.dsts) {
if (dst.memRank != t.exeDevice.exeRank && IsCpuMemType(dst.memType)) {
hasRemoteCpuMem = true; offender = dst; role = "DST"; break;
}
}
}
if (hasRemoteCpuMem) {
errors.push_back({ERR_FATAL,
"Transfer %d: Cross-rank GPU executor (R%d%c%d) cannot access remote host memory "
"(%s on rank %d is %s). Fabric-handle sharing only supports GPU memory for 1.67; use a NIC "
"executor (e.g. R%dN..) for cross-rank transfers involving host memory.",
i, t.exeDevice.exeRank, ExeTypeStr[t.exeDevice.exeType], t.exeDevice.exeIndex,
role, offender.memRank, GetMemTypeName(offender.memType), t.exeDevice.exeRank});
Comment thread
AtlantaPepsi marked this conversation as resolved.
hasFatalError = true;
break;
}
}
}

// Check subexecutors
Expand Down
Loading