Skip to content

[7.12][ML] Reduce input stream buffer from 8KB to 2KB (#1881) #1884

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Conversation

dimitris-athanasiou
Copy link
Contributor

When we flush the input stream the java side writes
enough spaces to fill the input stream buffer.
However, in the case of data frame analytics, this may
cause the job to freeze. The reason is that java
writes the data and flushes in the same thread that
goes on to then restore the state. However, when
c++ reads in the end-of-data control message, it stops
reading from the stream and goes on to perform the analysis.
If the 8KB of spaces do not fit in the OS buffer for the
names pipe, the java side blocks. It never proceeds with
restoring the state and this causes a job that is
being restarted and has state to freeze.

This commit reduces the buffer to 2KB. This means the
buffer is smaller than the buffer of all supported OS.
Note that it is 4KB on Windows. I have performed
a test to assess performace impact. In 10 runs of
a job that processes ~1M docs with a total input size
of 62MB, the result was the same both with the buffer
being 8KB and 2KB. The result supports that there is
no significant impact on performance.

Also note an alternative solution would have been to
use an additional thread to keep reading the input stream.
However, this solution seems to be much less complicated.

Backport of #1881

When we flush the input stream the java side writes
enough spaces to fill the input stream buffer.
However, in the case of data frame analytics, this may
cause the job to freeze. The reason is that java
writes the data and flushes in the same thread that
goes on to then restore the state. However, when
c++ reads in the end-of-data control message, it stops
reading from the stream and goes on to perform the analysis.
If the 8KB of spaces do not fit in the OS buffer for the
names pipe, the java side blocks. It never proceeds with
restoring the state and this causes a job that is
being restarted and has state to freeze.

This commit reduces the buffer to 2KB. This means the
buffer is smaller than the buffer of all supported OS.
Note that it is 4KB on Windows. I have performed
a test to assess performace impact. In 10 runs of
a job that processes ~1M docs with a total input size
of 62MB, the result was the same both with the buffer
being 8KB and 2KB. The result supports that there is
no significant impact on performance.

Also note an alternative solution would have been to
use an additional thread to keep reading the input stream.
However, this solution seems to be much less complicated.

Backport of elastic#1881
@dimitris-athanasiou
Copy link
Contributor Author

retest

@dimitris-athanasiou dimitris-athanasiou merged commit 27dee13 into elastic:7.12 Apr 29, 2021
@dimitris-athanasiou dimitris-athanasiou deleted the reduce-input-buffer-size-to-2kb-7_12 branch April 29, 2021 08:05
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Projects
None yet
Development

Successfully merging this pull request may close these issues.

1 participant