GPU Computing · Cluster
Heterogeneous Systems
Host-device orchestration and unified memory.
- 01 Pinned Host Memory and Why Pageable Transfers Cannot Be Asynchronous Why a copy engine cannot be handed an ordinary host pointer, what the driver does instead, and how the staging buffer turns an asynchronous copy into a synchronous one.
- 02 Unified Virtual Addressing and How a Pointer Becomes Usable on a Device Why a single address space spanning host and device memory lets an API infer a pointer location from its value, what that inference depends on, and where it fails.
- 03 Stream Ordering: What a Queue Guarantees and What It Does Not The ordering contract of a command stream, how host-side submission stays asynchronous, and why two streams give no ordering at all without an explicit dependency.
- 04 GPU Page Faults and Demand Migration in Unified Memory How a device MMU fault suspends a warp, what the driver does with the faulting address, and why migration granularity decides whether oversubscription works at all.
- 05 Host-Mapped Memory over PCIe: Zero-Copy Access and Its Latency Wall What happens when a kernel dereferences host memory directly across the bus, why write-combining changes the cost, and the access patterns that make it worth doing.
- 06 System-Scope Atomics and Memory Ordering Across the Bus Why an atomic operation on device memory is not automatically visible to the host, what the scope qualifier changes, and how release and acquire order a flag against its payload.
- 07 Residency Hints and the Arithmetic of Managed-Memory Thrashing How preferred-location and accessed-by advice changes where a managed page lives, why advice is not a guarantee, and how to tell a capacity problem from a policy problem.
- 08 Peer-to-Peer Transfers and the Topology That Decides Whether They Work What peer access actually establishes in the device page tables, why the path between two devices is not the path through the host, and how to read the topology.
- 09 Partitioning Work Across Devices and the Cost of Every Dependency How to split a computation across several devices without drowning it in synchronization, and why the cost of a dependency is set by the slowest participant.
- 10 Page Table Mirroring Between CPU and Device How the kernel lets a device driver snapshot a process page table, why the snapshot needs an invalidation callback, and what breaks when the two disagree.