[
https://issues.apache.org/jira/browse/HADOOP-14999?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Genmao Yu updated HADOOP-14999:
-------------------------------
Description:
This mechanism is designed for uploading file in parallel and asynchronously:
- improve the performance of uploading file to OSS server. Firstly, this
mechanism splits result to multiple small blocks and upload them in parallel.
Then, getting result and uploading blocks are asynchronous.
- avoid buffering too large result into local disk. To cite an extreme
example, there is a task which will output 100GB or even larger, we may need to
output this 100GB to local disk and then upload it. Sometimes, it is
inefficient and limited to disk space.
This patch reuse {{SemaphoredDelegatingExecutor}} as executor service and
depends on HADOOP-15039.
Attached {{asynchronous_file_uploading.pdf}} illustrated the difference between
previous {{AliyunOSSOutputStream}} and {{AliyunOSSBlockOutputStream}}, i.e.
this asynchronous multi-part based uploading mechanism.
1. {{AliyunOSSOutputStream}}: we need to output the whole result to local disk
before we can upload it to OSS. This will poses two problems：
- if the output file is too large, it will run out of the local disk.
- if the output file is too large, task will wait long time to upload result
to OSS before finish, wasting much compute resource.
2. {{AliyunOSSBlockOutputStream}}: we cut the task output into small blocks,
i.e. some small local file, and each block will be packaged into a uploading
task. These tasks will be submitted into {{SemaphoredDelegatingExecutor}}.
{{SemaphoredDelegatingExecutor}} will upload this blocks in parallel, this will
improve performance greatly.
3. Each task will retry 3 times to upload block to Aliyun OSS. If one of those
tasks failed, the whole file uploading will failed, and we will abort current
uploading.
was:
This mechanism is designed for uploading file in parallel and asynchronously:
- improve the performance of uploading file to OSS server. Firstly, this
mechanism splits result to multiple small blocks and upload them in parallel.
Then, getting result and uploading blocks are asynchronous.
- avoid buffering too large result into local disk. To cite an extreme example,
there is a task which will output 100GB or even larger, we may need to output
this 100GB to local disk and then upload it. Sometimes, it is inefficient and
limited to disk space.
This patch reuse {{SemaphoredDelegatingExecutor}} as executor service and
depends on HADOOP-15039.
Attached {{asynchronous_file_uploading.pdf}} illustrated the difference between
previous {{AliyunOSSOutputStream}} and {{AliyunOSSBlockOutputStream}}, i.e.
this asynchronous multi-part based uploading mechanism.
1. {{AliyunOSSOutputStream}}: we need to output the whole result to local disk
before we can upload it to OSS. This will poses two problems：
- if the output file is too large, it will run out of the local disk.
- if the output file is too large, task will wait long time to upload
result to OSS before finish, wasting much compute resource.
2. {{AliyunOSSBlockOutputStream}}: we cut the task output into small blocks,
i.e. some small local file, and each block will be packaged into a uploading
task. These tasks will be submitted into {{SemaphoredDelegatingExecutor}}.
{{SemaphoredDelegatingExecutor}} will upload this blocks in parallel, this will
improve performance greatly.
> AliyunOSS: provide one asynchronous multi-part based uploading mechanism
> ------------------------------------------------------------------------
>
> Key: HADOOP-14999
> URL: https://issues.apache.org/jira/browse/HADOOP-14999
> Project: Hadoop Common
> Issue Type: Sub-task
> Components: fs/oss
> Affects Versions: 3.0.0-beta1
> Reporter: Genmao Yu
> Assignee: Genmao Yu
> Priority: Major
> Attachments: HADOOP-14999.001.patch, HADOOP-14999.002.patch,
> HADOOP-14999.003.patch, HADOOP-14999.004.patch, HADOOP-14999.005.patch,
> HADOOP-14999.006.patch, HADOOP-14999.007.patch,
> asynchronous_file_uploading.pdf
>
>
> This mechanism is designed for uploading file in parallel and asynchronously:
> - improve the performance of uploading file to OSS server. Firstly, this
> mechanism splits result to multiple small blocks and upload them in parallel.
> Then, getting result and uploading blocks are asynchronous.
> - avoid buffering too large result into local disk. To cite an extreme
> example, there is a task which will output 100GB or even larger, we may need
> to output this 100GB to local disk and then upload it. Sometimes, it is
> inefficient and limited to disk space.
> This patch reuse {{SemaphoredDelegatingExecutor}} as executor service and
> depends on HADOOP-15039.
> Attached {{asynchronous_file_uploading.pdf}} illustrated the difference
> between previous {{AliyunOSSOutputStream}} and
> {{AliyunOSSBlockOutputStream}}, i.e. this asynchronous multi-part based
> uploading mechanism.
> 1. {{AliyunOSSOutputStream}}: we need to output the whole result to local
> disk before we can upload it to OSS. This will poses two problems：
> - if the output file is too large, it will run out of the local disk.
> - if the output file is too large, task will wait long time to upload result
> to OSS before finish, wasting much compute resource.
> 2. {{AliyunOSSBlockOutputStream}}: we cut the task output into small blocks,
> i.e. some small local file, and each block will be packaged into a uploading
> task. These tasks will be submitted into {{SemaphoredDelegatingExecutor}}.
> {{SemaphoredDelegatingExecutor}} will upload this blocks in parallel, this
> will improve performance greatly.
> 3. Each task will retry 3 times to upload block to Aliyun OSS. If one of
> those tasks failed, the whole file uploading will failed, and we will abort
> current uploading.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
---------------------------------------------------------------------
To unsubscribe, e-mail: common-issues-unsubscr...@hadoop.apache.org
For additional commands, e-mail: common-issues-h...@hadoop.apache.org
