Thank you for your email, Michael! We will have both implementations available so that experts here can discuss and review the pros and cons of each way :)
The performance gains of enabling SO_REUSEPORT is significant, especially on the multi-core systems. We would like to have it enabled in OpenJDK so that users can take advantage of it. Thanks, Yingqi (Lucy) Lu From: Michael McMahon [mailto:michael.x.mcma...@oracle.com] Sent: Thursday, November 19, 2015 9:12 AM To: Lu, Yingqi; Kharbas, Kishor; net-dev@openjdk.java.net Subject: Re: Patch for adding SO_REUSEPORT socket option Hi Yingqi, I think we should consider making it a standard option if there is a reasonable prospect that all the major platforms will eventually support it. Even if a particular platform does not support it initially (such as Windows) the API allows for this possibility by throwing UnsupportedOperationException on the set and get methods. The jdk.net API is probably more intended for options that are less likely to be widely supported. - Michael On 19/11/15 16:40, Lu, Yingqi wrote: Hi Michael, Thank you very much for your email. The version 1 of the patch is already uploaded at http://cr.openjdk.java.net/~mcberg/jdk/6432031/webrev.01/<http://cr.openjdk.java.net/%7Emcberg/jdk/6432031/webrev.01/>. The link is also available from https://bugs.openjdk.java.net/browse/JDK-6432031. We just completed the version 2 which implements the SO_REUSEPORT as an extended socket option (jdk.net.ExtendedSocketOptions) instead of a standard one (Following Alan Bateman's suggestion). The patch will be uploaded very soon hopefully today. If we have issues uploading it, we will let you know. Once we are done uploading, we will notify this community. Thanks, Yingqi Lu (Lucy) From: net-dev [mailto:net-dev-boun...@openjdk.java.net] On Behalf Of Michael McMahon Sent: Thursday, November 19, 2015 6:11 AM To: Kharbas, Kishor; net-dev@openjdk.java.net<mailto:net-dev@openjdk.java.net> Subject: Re: Patch for adding SO_REUSEPORT socket option Hi Kishor Did you mean to include the patch with this message? If you send it, I can convert it to a webrev and put it on the cr.openjdk server for review Thanks Michael On 16/11/15 18:49, Kharbas, Kishor wrote: Hello all, I request the community to review a patch for adding SO_REUSEPORT support. There is already an existing JBS opened at https://bugs.openjdk.java.net/browse/JDK-6432031 Details : SO_REUSEPORT removes 1:1 assignment between listen socket and IP:PORT pair and enable multiple sockets listening to the same address and port. This improves the scalability and parallelism of network traffic handling. It is enabled for both TCP and UDP sockets (at least for Linux). For more details, please refer to https://lwn.net/Articles/542629/. Many applications, especially Linux or BSD based webservers such as Apache httpd and Nginx are already supporting it now. Ruby and Python have it supported as well. Other Java applications such as Netty webserver have it supported via JNI function since JDK has not supported it yet. By enabling the SO_REUSEPORT feature itself, up to 4X throughput and latency improvement have been observed from various applications. Specific to Java application with this patch, we modified Apache Hadoop Distributed File System (HDFS) source code to take advantage of this feature. We observed up to 1.93x performance improvements. The feature is supported since Linux Kernel 3.9. It is also supported in BSDs, Solaris and Mac OS. Windows does not have it. In the current patch, we only enable the feature on Linux platform since we do not have BSD, Solaris and Mac OS for testing. Whether the feature is supported or not on the running kernel is determined at the run time. P.S. Based on Alan Baleman's comment on JBS, we are in meanwhile working on adding this option to 'java.net.ExtendedSocketOption'. Regards, Kishor Kharbas
