Distributed Symmetric Multiprocessing Computing Architecture

• US 20140143368A1
• Filed: 12/09/2013
• Published: 05/22/2014
• Est. Priority Date: 11/13/2009
• Status: Abandoned Application

• Current Assignee: Richard S. Anderson
• Original Assignee: Richard S. Anderson
• Inventors: Anderson, Richard S.
• Application Number:
  US14/100,415
• Publication Number:
  US 20140143368A1
• Time in Patent Office: Days
• US Class Current: 709/213
• CPC Class Codes: G06F 12/084 with a shared cache
• CPC Class Codes: G06F 15/17331 Distributed shared memory [DSM], e.g. remote direct memory access [RDMA]

Abstract

Example embodiments of the present invention includes systems and methods for implementing a scalable symmetric multiprocessing (shared memory) computer architecture using a network of homogeneous multi-core servers. The level of processor and memory performance achieved is suitable for running applications that currently require cache coherent shared memory mainframes and supercomputers. The architecture combines new operating system extensions with a high-speed network that supports remote direct memory access to achieve an effective global distributed shared memory. A distributed thread model allows a process running in a head node to fork threads in other (worker) nodes that run in the same global address space. Thread synchronization is supported by a distributed mutex implementation. A transactional memory model allows a multi-threaded program to maintain global memory page consistency across the distributed architecture. A distributed file access implementation supports non-contentious file I/O for threads. These and other functions provide a symmetric multiprocessing programming model consistent with standards such as Portable Operating System Interface for Unix (POSIX).

Claims

1. A method for the implementation of simultaneous, non-contentious file access by multiple threads or processes using distributed buffered file input/output on an underlying network of symmetric multiprocessing (SMP) nodes, comprising:

performing all direct file reads and writes from the head node of the network of SMP nodes;

maintaining a pool of buffers in the head node to store pending write buffers;

extending buffered file read/write access to threads running in the head node or non-head nodes through the use of messages exchanged between the head node and other nodes; and

providing non-blocking file writes that are guaranteed atomic up to the maximum buffer size, and that eliminate the need for locking.

2. The method of claim 1, further including a distributed message queue hosted on the head node that can receive messages from any node in the network requesting file operations, allowing for multiple simultaneous blocking atomic read file operations up to the maximum buffer size as well as supporting non-blocking file write operations.

3. A method for providing memory coherency in a cluster of SMP nodes supporting distributed symmetric multiprocessing (DSMP), the method comprising:

support for hardware memory cache coherency for memory accesses within a node, but not for memory accesses between nodes;

providing a mechanism for explicitly locking a global memory page for exclusive access by an executing application program thread;

providing a mechanism for explicitly forcing an immediate update (synchronization) of a global memory page with a local copy of the page by an executing application program thread; and

memory page locking and synchronization executed explicitly by an application program thread via system calls within the application program or a library or programming language construct used by the application program;

but not automatically by the operating system or underlying hardware;

thereby forming an application program controlled memory page coherency model.

4. The method of claim 3, further allowing the assembly of a scalable symmetric multiprocessing (shared memory) supercomputer that will support thousands of nodes with tens of thousands of processor sockets by eliminating the need for memory cache line snoops between nodes through the use of distributed symmetric multiprocessing (DSMP) as described in the parent patent.

7. The method of claim 3, further allowing the option for a subset of nodes to be designated as compute only nodes that contain a local memory partition only, and do not contain a global memory partition, but are able to request pages from the global memory of other nodes, and thereby support running DSMP application program threads.

5. A method of providing one or more instances of a distributed symmetric multiprocessing (DSMP) computing environment within a cluster of SMP compute nodes, the method comprising:

designating one or more compute nodes in the cluster to be a head node, with the remaining nodes designated as worker nodes;

assembling one or more distributed symmetric multiprocessing (DSMP) instances as described in the parent patent, by selecting a single head node and a subset of the available worker nodes for each instance, with the number of worker nodes determined by the requirements of the particular DSMP instance.

6. The method of claim 5, further allowing the option for a subset of worker nodes to be designated as memory only nodes that are by convention restricted from running DSMP application program threads on the node, but do contain a global memory partition, and service memory page requests from other nodes.