公开(公告)号：US20160077997A1

公开(公告)日：2016-03-17

申请号：US14881157

申请日：2015-10-13

Applicant: Cray Inc.

Inventor： Edwin L. Froese ,  Eric P. Lundberg ,  Igor Gorodetsky ,  Howard Pritchard ,  Charles Giefer ,  Robert L. Alverson ,  Duncan Roweth

IPC: G06F15/167 ,  H04L29/08

CPC classification number: H04L67/10 ,  G06F8/458 ,  G06F9/52 ,  G06F9/524 ,  G06F9/54 ,  G06F15/167 ,  H04L43/00 ,  H04L43/02 ,  H04L45/02 ,  H04L45/04 ,  H04L67/2842

Abstract: An improved method for the prevention of deadlock in a massively parallel processor (MPP) system wherein, prior to a process sending messages to another process running on a remote processor, the process allocates space in a deadlock-avoidance FIFO. The allocated space provides a “landing zone” for requests that the software process (the application software) will subsequently issue using a remote-memory-access function. In some embodiments, the deadlock-avoidance (DLA) function provides two different deadlock-avoidance schemes: controlled discard and persistent reservation. In some embodiments, the software process determines which scheme will be used at the time the space is allocated.

Abstract translation: 一种用于在大规模并行处理器（MPP）系统中防止死锁的改进方法，其中在将消息发送到在远程处理器上运行的另一进程之前，该进程在死锁避免FIFO中分配空间。 分配的空间为软件进程（应用软件）随后将使用远程存储器访问功能发出的请求提供“着陆区域”。 在一些实施例中，死锁避免（DLA）功能提供两种不同的死锁避免方案：受控丢弃和永久保留。 在一些实施例中，软件过程确定在分配空间时将使用哪种方案。

公开(公告)号：US09160607B1

公开(公告)日：2015-10-13

申请号：US13798074

申请日：2013-03-12

Applicant: Cray Inc.

Inventor： Edwin L. Froese ,  Eric P. Lundberg ,  Igor Gorodetsky ,  Howard Pritchard ,  Charles Giefer ,  Robert L. Alverson ,  Duncan Roweth

IPC: G06F15/167 ,  H04L29/08 ,  G06F9/45 ,  G06F9/52 ,  G06F9/54 ,  G06F15/00 ,  G06F9/46

CPC classification number: H04L67/10 ,  G06F8/458 ,  G06F9/52 ,  G06F9/524 ,  G06F9/54 ,  G06F15/167 ,  H04L12/2602 ,  H04L43/00 ,  H04L43/02 ,  H04L45/02 ,  H04L45/04 ,  H04L67/2842

Abstract: An improved method for the prevention of deadlock in a massively parallel processor (MPP) system wherein, prior to a process sending messages to another process running on a remote processor, the process allocates space in a deadlock-avoidance FIFO. The allocated space provides a “landing zone” for requests that the software process (the application software) will subsequently issue using a remote-memory-access function. In some embodiments, the deadlock-avoidance (DLA) function provides two different deadlock-avoidance schemes: controlled discard and persistent reservation. In some embodiments, the software process determines which scheme will be used at the time the space is allocated.

Abstract translation: 一种用于在大规模并行处理器（MPP）系统中防止死锁的改进方法，其中在将消息发送到在远程处理器上运行的另一进程之前，该进程在死锁避免FIFO中分配空间。 分配的空间为软件进程（应用软件）随后将使用远程存储器访问功能发出的请求提供“着陆区域”。 在一些实施例中，死锁避免（DLA）功能提供两种不同的死锁避免方案：受控丢弃和永久保留。 在一些实施例中，软件过程确定在分配空间时将使用哪种方案。

公开(公告)号：US10129329B2

公开(公告)日：2018-11-13

申请号：US14881157

申请日：2015-10-13

Applicant: Cray Inc.

Inventor： Edwin L. Froese ,  Eric P. Lundberg ,  Igor Gorodetsky ,  Howard Pritchard ,  Charles Giefer ,  Robert L. Alverson ,  Duncan Roweth

IPC: G06F15/167 ,  H04L29/08 ,  G06F8/41 ,  G06F9/52 ,  G06F9/54 ,  H04L12/26 ,  H04L12/751 ,  H04L12/715

Abstract: An improved method for the prevention of deadlock in a massively parallel processor (MPP) system wherein, prior to a process sending messages to another process running on a remote processor, the process allocates space in a deadlock-avoidance FIFO. The allocated space provides a “landing zone” for requests that the software process (the application software) will subsequently issue using a remote-memory-access function. In some embodiments, the deadlock-avoidance (DLA) function provides two different deadlock-avoidance schemes: controlled discard and persistent reservation. In some embodiments, the software process determines which scheme will be used at the time the space is allocated.

公开(公告)号：US09294551B1

公开(公告)日：2016-03-22

申请号：US13798082

申请日：2013-03-13

Applicant: Cray Inc.

Inventor： Edwin L. Froese ,  Robert Baird ,  Steven L. Scott

IPC: G06F15/16 ,  H04L29/08 ,  H04L12/715 ,  H04L12/751 ,  H04L12/26

CPC classification number: H04L67/10 ,  G06F8/458 ,  G06F9/52 ,  G06F9/524 ,  G06F9/54 ,  G06F15/167 ,  H04L12/2602 ,  H04L43/00 ,  H04L43/02 ,  H04L45/02 ,  H04L45/04 ,  H04L67/2842

Abstract: A data-processing system and method for performing collective operations. Some embodiments provide a plurality of leaf software processes, a plurality of collective engines (CEs), and a network operatively coupled to the plurality of CEs, wherein collective operations messages are sent between CEs. Each of the plurality of hierarchies includes a root, leaf CEs, and one or more intermediate levels of CEs between the root and the leaf CEs. Each CE except the root is configured to have a parent CE, and each non-leaf CE in the hierarchy that is not the root CE and not one of the leaf CEs has one or more child CEs. Data is sent from software processes to one or more of the plurality of CEs, and data is received to one or more software processes from one or more of the plurality of collective engines. The root CE outputs a final result.

Abstract translation: 用于执行集体操作的数据处理系统和方法。 一些实施例提供多个叶软件过程，多个集体引擎（CE）和可操作地耦合到多个CE的网络，其中在CE之间发送集合操作消息。 多个层次中的每一个包括根，叶CE，以及​​根和叶CE之间的CE的一个或多个中间层。 除了root之外的每个CE被配置为具有父CE，并且层次结构中不是根CE而不是一个叶CE的每个非叶CE都具有一个或多个子CE。 数据从软件进程发送到多个CE中的一个或多个，并且数据从多个集合引擎中的一个或多个接收到一个或多个软件进程。 根CE输出最终结果。