[
https://issues.apache.org/jira/browse/ARROW-5285?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
shengjun.li updated ARROW-5285:
-------------------------------
Description:
cpp/CMakeLists.txt
option(ARROW_CUDA "Build the Arrow CUDA extensions (requires CUDA toolkit)"
ON)
option(ARROW_PLASMA "Build the plasma object store along with Arrow" ON)
In the plasma client, GpuProcessHandle is never released although GPU object is
deleted.
Thus, cuIpcCloseMemHandle is never called.
When I repeatly creat and delete gpu memory, the following error may occur.
IOError: Cuda Driver API call in
/home/zilliz/arrow/cpp/src/arrow/gpu/cuda_context.cc at line 155 failed with
code 208: cuIpcOpenMemHandle(&data, *handle, CU_IPC_MEM_LAZY_ENABLE_PEER_ACCESS)
Note: CUDA_ERROR_ALREADY_MAPPED = 208
was:
When GPU memory created, cuIpcOpenMemHandle is called.
But when GPU memory distorted, cuIpcCloseMemHandle is never called.
cpp/CMakeLists.txt
option(ARROW_CUDA "Build the Arrow CUDA extensions (requires CUDA toolkit)"
ON)
option(ARROW_PLASMA "Build the plasma object store along with Arrow" ON)
Repeatly creat and delete gpu memory, the following error may occur.
error: 5 IOError: Cuda Driver API call in
/home/zilliz/arrow/cpp/src/arrow/gpu/cuda_context.cc at line 155 failed with
code 208: cuIpcOpenMemHandle(&data, *handle, CU_IPC_MEM_LAZY_ENABLE_PEER_ACCESS)
Note: CUDA_ERROR_ALREADY_MAPPED = 208
Summary: [C++][Plasma] GpuProcessHandle is not released when GPU object
deleted (was: [C++][Plasma] never call cuIpcCloseMemHandle before GPU memory
distorted)
> [C++][Plasma] GpuProcessHandle is not released when GPU object deleted
> ----------------------------------------------------------------------
>
> Key: ARROW-5285
> URL: https://issues.apache.org/jira/browse/ARROW-5285
> Project: Apache Arrow
> Issue Type: Bug
> Components: C++
> Affects Versions: 0.13.0
> Reporter: shengjun.li
> Priority: Major
> Labels: pull-request-available
> Fix For: 0.14.0
>
> Time Spent: 10m
> Remaining Estimate: 0h
>
> cpp/CMakeLists.txt
> option(ARROW_CUDA "Build the Arrow CUDA extensions (requires CUDA toolkit)"
> ON)
> option(ARROW_PLASMA "Build the plasma object store along with Arrow" ON)
> In the plasma client, GpuProcessHandle is never released although GPU object
> is deleted.
> Thus, cuIpcCloseMemHandle is never called.
> When I repeatly creat and delete gpu memory, the following error may occur.
> IOError: Cuda Driver API call in
> /home/zilliz/arrow/cpp/src/arrow/gpu/cuda_context.cc at line 155 failed with
> code 208: cuIpcOpenMemHandle(&data, *handle,
> CU_IPC_MEM_LAZY_ENABLE_PEER_ACCESS)
> Note: CUDA_ERROR_ALREADY_MAPPED = 208
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
