Skip to content

stanford-mast/Silo

Repository files navigation

Silo is a multi-platform topology-aware memory management library for x86-based systems containing one or more NUMA nodes. It features a simple API for allocating and deallocating memory in a NUMA-aware manner, including in ways not automatically supported by the underlying system. For example, Silo keeps track of all buffers and arrays it allocates to simplify the deallocation procedure.

In addition to simple buffer allocation (in the style of most existing malloc-type functions) Silo supports multi-node array allocation. A multi-node array is a virtually-contiguous buffer physically backed by memory on different NUMA nodes. This mode of allocation is useful for the purpose of obtaining NUMA awareness without adding an explicit software layer of indirection, relying on hardware address translation instead.

Silo is implemented using C++.

Requirements

To build and link with Silo, the following are required.

  • 64-bit x86-compatible processor with support for AVX instructions

    Silo has been tested with Intel processors of the Sandy Bridge generation and newer. Non-Intel processors may also be compatible if they support the required instructions. Do not attempt to use it on older-generation processors, at the risk of encountering "Illegal Instruction" errors.

  • Windows 10 Pro or Ubuntu 14.04

    64-bit versions of these operating systems are required. Other Linux distributions are likely compatible, but they have not been tested.

  • Topo

    Topo provides some helpers and wrappers around the hwloc library.

  • libnuma (Linux only)

    Silo makes use of functions in this library for certain memory allocation operations. It is likely available as a package. No equivalent library is required on Windows; all needed functions are built into the operating system.

  • hwloc and its dependencies

    Silo makes use of this library to obtain system topology information in a platform-independent manner. It has been tested with version 1.11.4. Some distributions of Linux may make this library available as a package. Otherwise, and in all cases on Windows, it must be downloaded and installed manually.

Building

On all platforms, Silo compiles to a static library.

The Windows build system is based on Visual Studio 2015 Community Edition. Compilation is known to work from the graphical interface, but command-line build is also likely possible.

To build on Linux, just type make from within the repository directory.

Linking and Using

Projects that make use of Silo should include the top-level silo.h header file and nothing else.

Assuming a Linux-based C-language project that uses Silo and consists of a single source file called "main.c", the following command would build and link with Silo.

g++ main.c -lsilo -ltopo -lhwloc -lnuma -lpciaccess -lxml2

Getting Started

Documentation is available and can be built using Doxygen. It covers both the external API in silo.h and Silo's internals, the latter being of interest only to those who wish to modify the implementation of Silo.

On Linux, type make docs to compile the documentation. On Windows, run the Doxygen tool using the repository directory as the working directory (it may be necessary to create the output directory manually first).

The remainder of this section is designed to facilitate a quick understanding of the Silo API. It explains key concepts and defines key terms, providing pointers to relevant API functions and data structures. Source code examples are provided immediately following.

Concepts and Terminology

Silo supports allocating two types of buffers: a simple buffer and a multi-node array. Both are contiguous blocks of memory in virtual address space, but they differ in how they are specified to Silo and how they map to physical memory. In all cases, Silo allocates memory in blocks no smaller than the system's allocation granularity, whose value in Bytes may be obtained by calling siloGetAllocationUnitSize().

A simple buffer, allocated using siloSimpleBufferAlloc(), is a block of memory local to a single NUMA node, similar to what would be allocated by a call to malloc(). Silo supports binding the backing physical memory to a specific NUMA node but otherwise behaves identically to any standard memory allocation function.

A multi-node array, allocated using siloMultinodeArrayAlloc(), is a virtually-contiguous block of memory backed by different NUMA nodes at controlled boundaries. This type of allocation is specified piece-wise, whereby each piece defines the size of a block and the NUMA node that should back it physically. The total size of the array is the sum of the sizes of each piece, and each piece may be physically backed by memory on any NUMA node in the system. There is no defined limit on the number of pieces that can be specified.

All memory allocated via Silo is to be freed by calling siloFree() and passing only a pointer to the buffer originally returned from one of Silo's memory allocation functions. Silo internally handles all required book-keeping so that the operating system can properly free all allocated memory.

Examples

Two examples are provided: one showing how to allocate a simple buffer and one showing how to allocate a multi-node array.

Example 1: Simple Buffer Allocation

This example allocates and initializes simple 1GB buffer on the first NUMA node in the system.

#include <silo.h>
#include <string.h>

int main(int argc, char* argv[])
{
    // Allocate the buffer.
    // Size is the first parameter, NUMA node index is the second.
    void* allocatedBuffer = siloSimpleBufferAlloc(1024llu * 1024llu * 1024llu, 0);
    
    // Use the buffer.
    if (NULL != allocatedBuffer)
    {
        // Initialize the array.
        memset(allocatedBuffer, 0, 1024llu * 1024llu * 1024llu);
        
        // Code goes here...
    }
    
    // Free the buffer.
    siloFree(allocatedBuffer);
    
    return 0;
}

Example 2: Multi-Node Array Allocation

This example allocates and initializes a multi-node array, 1GB per NUMA node.

#include <malloc.h>
#include <silo.h>
#include <string.h>
#include <topo.h>

int main(int argc, char* argv[])
{
    const size_t numNumaNodes = (size_t)topoGetSystemNUMANodeCount();
    
    // Create one piece specification per NUMA node in the system.
    SSiloMemorySpec* specs = (SSiloMemorySpec*)malloc(sizeof(SSiloMemorySpec) * numNumaNodes);
    if (NULL == specs)
        return 1;
    
    // Initialize the specifications.
    for (size_t i = 0; i < numNumaNodes; ++i)
    {
        specs[i].size = 1024llu * 1024llu * 1024llu;
        specs[i].numaNode = i;
    }
    
    // Allocate the multi-node array.
    void* allocatedBuffer = siloMultinodeArrayAlloc(numNumaNodes, specs);
    
    // Use the buffer.
    if (NULL != allocatedBuffer)
    {
        // Initialize the array.
        memset(allocatedBuffer, 0, numNumaNodes * 1024llu * 1024llu * 1024llu);
        
        // Code goes here...
    }
    
    // Free the buffer.
    siloFree(allocatedBuffer);
    
    return 0;
}

Copyright

Silo is licensed under BSD 3-clause (see "LICENSE" in the top-level source code directory).

Copyright (c) 2016-2017 Stanford University, Department of Electrical Engineering. Authored by Samuel Grossman.

About

Multi-platform topology-aware memory management library

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published