Commit d76457f8 authored by Swann Perarnau's avatar Swann Perarnau

[doc] small readme that actual match the code

Things have changed. Reflect the new code.
parent 15cd651b
Argonne's Memory Library
========================
AML: Building Blocks for Memory Management
=========================================
This is a prototype library to manage the hierarchical memory present in recent
and future HPC systems, including NVM and stacked memory.
AML is a library to manage byte-addressable memory devices. The library
is designed as a collection of building blocks, so that users can create custom
memory management policies for allocation and placement of data across devices.
This library defines "memory nodes" as hardware locations where data can
reside. Not all of those memory nodes can provide virtual memory access to the
data they contain, and as such, allocations are identified only by a UID.
Memory nodes are defined by a public data structure that also contains
information on the granularity of data accesses to a node.
This library is still in the prototyping phase. APIs might break often.
The library then provides functions to "pull" data into a virtual memory
location. That memory location should be seen as temporary: the user should ask
the library to "push" out the data afterwards.
# General Architecture
The architecture of the library relies on two principles: type-specific
initialization functions, and generic interfaces to each building block.
The type-specific initialization functions include:
- `_DECL` macros to declare a typed building block on the stack.
- `_create_` functions to allocate a building block and return its pointer.
- `_init_` functions to initialize stack-allocated building blocks.
Generic interfaces all take a pointer to a building block.
# Low-Level Building Blocks
Low-level building blocks provide the basic mechanisms required to implement
any high-level memory management policy. This include:
- *arenas:* allocation policies inside a memory region
- *areas:* actual memory reservation on devices
- *bindings:* mapping pages/tiles of memory unto multiple devices
- *tilings:* chunking data
# High-Level Building Blocks
High-level building blocks use the low-level ones to provide fancier memory
management facilities, including:
- *dmas:* moving data across areas
- *scratchpads:* using an area as an explicitly managed cache of another one
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment