[jira] [Comment Edited] (WAGON-537) Maven transfer speed of large artifacts is slow due to unsuitable buffer strategy

["Olaf Otto (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/WAGON-537?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16739361#comment-16739361 ] 

Olaf Otto edited comment on WAGON-537 at 1/10/19 12:34 PM:
-----------------------------------------------------------

I have began testing using docker on win 10 (with hyper-v). For testing, I have run a local docker container via
{code:java}
docker run -i -t -v [local maven installation dir]:/opt/maven -v [local dir with test POM]:/opt/test -v [local .m2 dir]:/root/.m2  openjdk:latest /bin/bash {code}
Then, I have executed a CURL download of a 4GB test file from a remote nexus repo as a reference point. Subsequently, I executed a maven artifact download using maven 3.6.0 with and without the patch. Here are the results:

*Reference download via CURL:* 11 MB/s
 *Download with this patch:* 11 MB /s
 *Download without patch:* < 1 MB /s

Moreover, the unpatched version caused massive CPU usage due to the millions of invocations of fireTransferProgress. Thus, when using docker for windows, one can see a significant improvement when the remote artifact repo is relatively slow.

I will make another test on a mac to see how this played out on a bare metal *nix.


was (Author: o.otto):
I have began testing using docker on win 10 (with hyper-v). For testing, I have run a local docker container via

 {code}docker run -i -t -v [local maven installation dir]:/opt/maven -v [local dir with test POM]:/opt/test -v [local .m2 dir]:/root/.m2  openjdk:latest /bin/bash {code}

Then, I have executed a CURL download of a 4GB test file from a remote nexus repo as a reference point. Subsequently, I executed a maven artifact download using maven 3.6.0 with and without the patch. Here are the results:

*Reference download via CURL:* 11 MB/s
*Download with this patch:* 11 MB /s
*Download without patch:* < 1 MB /s

Moreover, the unpatched version caused massive CPU usage due to the millions of invocations of fireTransferProgress. Thus, when using docker for windows, one can see a significant improvement when the remote artifact repo is relatively slow.

I will make another test on a mac to see how this played out on a bare metal *nix.


> Maven transfer speed of large artifacts is slow due to unsuitable buffer strategy
> ---------------------------------------------------------------------------------
>
>                 Key: WAGON-537
>                 URL: https://issues.apache.org/jira/browse/WAGON-537
>             Project: Maven Wagon
>          Issue Type: Improvement
>          Components: wagon-http, wagon-provider-api
>    Affects Versions: 3.2.0
>         Environment: Windows 10, JDK 1.8, Nexus  Artifact store > 100MB/s network connection.
>            Reporter: Olaf Otto
>            Assignee: Michael Osipov
>            Priority: Major
>              Labels: perfomance
>             Fix For: 3.3.0, 3.3.1
>
>         Attachments: wagon-issue.png
>
>
> We are using maven for build process automation with docker. This sometimes involves uploading and downloading artifacts with a few gigabytes in size. Here, maven's transfer speed is consistently and reproducibly slow. For instance, an artifact with 7,5 GB in size took almost two hours to transfer in spite of a 100 MB/s connection with respective reproducible download speed from the remote nexus artifact repository when using a browser to download. The same is true when uploding such an artifact.
> I have investigated the issue using JProfiler. The result shows an issue in AbstractWagon's transfer( Resource resource, InputStream input, OutputStream output, int requestType, long maxSize ) method used for remote artifacts and the same issue in AbstractHttpClientWagon#writeTo(OutputStream).
> Here, the input stream is read in a loop using a 4 Kb buffer. Whenever data is received, the received data is pushed to downstream listeners via fireTransferProgress. These listeners (or rather consumers) perform expensive tasks.
> Now, the underlying InputStream implementation used in transfer will return calls to read(buffer, offset, length) as soon as *some* data is available. That is, fireTransferProgress may well be invoked with an average number of bytes less than half the buffer capacity (this varies with the underlying network and hardware architecture). Consequently, fireTransferProgress is invoked *millions of times* for large files. As this is a blocking operation, the time spent in fireTransferProgress dominates and drastically slows down the transfers by at least one order of magnitude. 
> !wagon-issue.png! 
> In our case, we found download speed reduced from a theoretical optimum of ~80 seconds to to more than 3200 seconds.
> From an architectural perspective, I would not want to make the consumers / listeners invoked via fireTransferProgress aware of their potential impact on download speed, but rather refactor the transfer method such that it uses a buffer strategy reducing the the number of fireTransferProgress invocations. This should be done with regard to the expected file size of the transfer, such that fireTransferProgress is invoked often enough but not to frequent.



--
This message was sent by Atlassian JIRA
(v7.6.3#76005)