公开(公告)号：US20200349078A1

公开(公告)日：2020-11-05

申请号：US16932318

申请日：2020-07-17

Applicant: HUAWEI TECHNOLOGIES CO., LTD.

Inventor： Yongbo CHENG ,  Chenghong HE ,  Tao HE

IPC: G06F12/0817 ,  G06F12/0891 ,  G06F12/0895 ,  G06F9/54

Abstract: A directory processing method and apparatus are provided to resolve a problem that a directory occupies a relatively large quantity of caches in an existing directory processing solution. The method includes: receiving, by a first data node, a first request sent by a second data node; searching for, by the first data node, a matched directory entry in a directory of the first data node based on tag information and index information in a first physical address; creating, when no matched directory entry is found, a first directory entry of the directory based on the first request, where the first directory entry includes the tag information, first indication information, first pointer information, and first status information, the first pointer information is used to indicate that data in the memory address corresponding to the indication bit that is set to valid is read by the second data node.

公开(公告)号：US20150370712A1

公开(公告)日：2015-12-24

申请号：US14746391

申请日：2015-06-22

Applicant: Huawei Technologies Co., Ltd.

Inventor： Yongbo CHENG

IPC: G06F12/08 ,  G06F12/12

CPC classification number: G06F12/128 ,  G06F11/10 ,  G06F11/1048 ,  G06F12/0813 ,  G06F12/0826 ,  G06F12/0828 ,  G06F12/121 ,  G06F2212/2542 ,  G06F2212/62 ,  G06F2212/69

Abstract: A method for processing an error directory of a node in a cache coherence non-uniform memory access (CC-NUMA) system and a node are provided. The method effectively reduces a possibility of a breakdown of the system caused by accumulation of the error bits in the directory memory of the CC-NUMA system. The method comprises: when a quantity of bits of a correctable error of a directory stored in a directory memory of the node is greater than a preset threshold, controlling all processors in the CC-NUMA system to write dirty data in a corresponding cache back to a corresponding main memory, flush the dirty data, and directly flush clean data in the corresponding cache; and controlling the CC-NUMA system to enter a quiescent state, clearing a record stored in the directory memory to zero, and controlling, after the zero clearing is completed, the CC-NUMA system to exit the quiescent state.

Abstract translation: 提供了一种用于处理高速缓存一致性非均匀存储器访问（CC-NUMA）系统和节点中的节点的错误目录的方法。 该方法有效地减少了由CC-NUMA系统的目录存储器中的错误位累积引起的系统故障的可能性。 该方法包括：当存储在节点的目录存储器中的目录的可校正错误的位的数量大于预设阈值时，控制CC-NUMA系统中的所有处理器将相应缓存中的脏数据写回到 相应的主内存，刷新脏数据，并直接刷新相应缓存中的清理数据; 并控制CC-NUMA系统进入静止状态，清除存储在目录存储器中的记录为零，并在清零完成后控制CC-NUMA系统退出静态状态。

公开(公告)号：US20150363315A1

公开(公告)日：2015-12-17

申请号：US14732172

申请日：2015-06-05

Applicant: Huawei Technologies Co., Ltd.

Inventor： Kejia LAN ,  Yongbo CHENG ,  Chenghong HE

IPC: G06F12/08

CPC classification number: G06F12/0831 ,  G06F2212/1016

Abstract: A method for temporarily storing data and a storage device is provided. The method for temporarily storing data is applied to the storage device, and the storage device includes a source agent and a target agent. The method includes: sending, by the source agent, a data obtaining request to the target agent; receiving, by the source agent, target data that is corresponding to the data obtaining request and is returned by the target agent; determining, by the source agent, whether a snooping request that is for the target data and sent by the target agent is received after the data obtaining request is sent and before the target data is received, where the snooping request indicates that the target agent is simultaneously processing an obtaining request from another source agent for the target data; and if the snooping request is received, discarding, by the source agent, the target data

Abstract translation: 提供一种临时存储数据的方法和存储装置。 临时存储数据的方法被应用于存储装置，并且存储装置包括源代理和目标代理。 该方法包括：由源代理向目标代理发送数据获取请求; 由所述源代理接收对应于所述数据获取请求并由所述目标代理返回的目标数据; 由源代理确定在发送数据获取请求之后并且在接收到目标数据之前是否接收到用于目标数据并由目标代理发送的窥探请求，其中，窥探请求指示目标代理是 同时处理来自另一个源代理的针对目标数据的获取请求; 并且如果接收到窥探请求，则由源代理丢弃目标数据

公开(公告)号：US20150113230A1

公开(公告)日：2015-04-23

申请号：US14515940

申请日：2014-10-16

Applicant: Huawei Technologies Co., Ltd.

Inventor： Yongbo CHENG

IPC: G06F12/08

CPC classification number: G06F12/0824 ,  G06F2212/271 ,  G06F2212/622

Abstract: The present invention discloses a directory storage method and a directory storage node controller. The method includes: obtaining, by a node controller NC in a local node, a storage address of a data block in a CPU in the local node, where the data block is read by a remote node; determining first content and second content that are respectively located in a first specific bit and a second specific bit of the storage address; determining, according to the first content and from each preset storage space used for storing a directory, a storage space in which an addressing address matches the first content; and correspondingly storing the second content and the directory in the determined storage space.

Abstract translation: 本发明公开了一种目录存储方法和目录存储节点控制器。 该方法包括：由本地节点中的节点控制器NC获取本地节点中的CPU中的数据块的存储地址，其中数据块被远程节点读取; 确定分别位于存储地址的第一特定位和第二特定位中的第一内容和第二内容; 根据所述第一内容和用于存储目录的每个预设存储空间，确定寻址地址与所述第一内容匹配的存储空间; 并且相应地将所述第二内容和所述目录存储在所确定的存储空间中。

公开(公告)号：US20180189181A1

公开(公告)日：2018-07-05

申请号：US15857726

申请日：2017-12-29

Applicant: HUAWEI TECHNOLOGIES CO., LTD.

Inventor： Yongbo CHENG ,  Kejia LAN ,  Chenghong HE

IPC: G06F12/0831 ,  G06F12/0842

CPC classification number: G06F12/0831 ,  G06F12/0817 ,  G06F12/0842 ,  G06F2212/1044

Abstract: The present disclosure provides a data read method and apparatus. An example method includes receiving a first read request message sent by a second processor at a first processor. If an address of the first data is within a range from an upper bound to a lower bound recorded in a first directory range register of the first processor, the first processor sends a first snoop message to at least one third processor other than the second processor, and simultaneously sends a second read request message to the memory. A first snoop response message sent by each of the at least one third processor according to the first snoop message is received. The first data sent by the memory according to the second data read request message is also received. The first processor sends the first data to the second processor according to the first snoop response message.

公开(公告)号：US20180113808A1

公开(公告)日：2018-04-26

申请号：US15795129

申请日：2017-10-26

Applicant: Huawei Technologies Co., Ltd.

Inventor： Kejia LAN ,  Yongbo CHENG ,  Chenghong HE

IPC: G06F12/0817 ,  G06F12/0891 ,  G06F12/0871 ,  G06F12/0811

CPC classification number: G06F12/0817 ,  G06F12/0811 ,  G06F12/0871 ,  G06F12/0891

Abstract: The present invention disclose a memory access method, including: receiving, by a first processing module, a read request sent by a second processing module for caching; determining, according to a destination directory, one or more target storage spaces in which the target data is cached; sending a listening request to one or more target processing modules, so that the target processing module returns a listening response to the second processing module, where the listening response is used for responding to the expected status; returning a target packet to the second processing module, where the target packet includes the target data and a listening quantity; when a quantity of the listening response received by the second processing module matches the listening quantity, receiving an update packet sent by the second processing module; and updating the destination directory according to the update packet.