[prev in list] [next in list] [prev in thread] [next in thread] 

List:       openjdk-nio-dev
Subject:    Re: RFR: 8266369: (se) Add wepoll based Selector [v8]
From:       Brian Burkhalter <bpb () openjdk ! java ! net>
Date:       2021-05-07 22:54:53
Message-ID: kgns-AGpDXomiXppor1grgCQhxG-pC1aonRYUaxEsoE=.731a04d2-2347-4f70-b872-8926053dfcdb () github ! com
[Download RAW message or body]

On Fri, 7 May 2021 07:31:17 GMT, Alan Bateman <alanb@openjdk.org> wrote:

> > This PR proposes a new Selector implementation for the Windows platform based on \
> > the Bert Belder's "wepoll" library. The wepoll library provides a scalable \
> > polling interface to the Ancillary Function Driver for Winsock. The interface \
> > that it exposes is close to the Linux epoll interface. The main benefit is that \
> > the cost of selection operations are a function of the number of sockets that are \
> > ready for I/O rather than the number of channels registered with the Selector. \
> > The existing Selector on Windows uses the legacy "select" which scales horribly, \
> > partly because it's O(N) on the number of registered channels and partly because \
> > it is limited to 1024 sockets and so requires a thread per 1024 sockets. 
> > The patch brings the wepoll sources into the repo, adds a small shim, and the \
> > Selector implementation itself. The new implementation needs to handle OOB data \
> > in the same way as the old implementation so the method to discard OOB moves to \
> > Net. 
> > I've added a microbenchmark to demonstrate the scaling when there is 1, 10, ... \
> > 10k connections and one socket ready for I/O. Windows Server 2019 was used for \
> > the data here. I've seen improvements with almost everything that I've tried. The \
> > only exception is a select after a wakeup where the old implementation skips the \
> > selection operation (arguably a long standing bug in the old implementation). 
> > Existing select based Selector where the selection operation is almost 1ms when \
> > there are 10,000 channels registered. 
> > Benchmark                (nchannels)  Mode  Cnt       Score       Error  Units
> > SelectOne.testSelectOne            1  avgt   20    1587.429  ±     4.053  ns/op
> > SelectOne.testSelectOne           10  avgt   20    2515.324  ±     8.141  ns/op
> > SelectOne.testSelectOne          100  avgt   20   10903.445  ±    65.508  ns/op
> > SelectOne.testSelectOne         1000  avgt   20  101651.369  ±   673.781  ns/op
> > SelectOne.testSelectOne        10000  avgt   20  932032.385  ± 59004.246  ns/op
> > 
> > 
> > New wepoll based Selector:
> > 
> > Benchmark                (nchannels)  Mode  Cnt     Score    Error  Units
> > SelectOne.testSelectOne            1  avgt   20  1428.914  ±  6.648  ns/op
> > SelectOne.testSelectOne           10  avgt   20  1429.806  ±  5.537  ns/op
> > SelectOne.testSelectOne          100  avgt   20  1433.371  ±  3.253  ns/op
> > SelectOne.testSelectOne         1000  avgt   20  1435.007  ± 12.420  ns/op
> > SelectOne.testSelectOne        10000  avgt   20  1445.201  ±  7.902  ns/op
> > 
> > 
> > Replacing the Selector implementation after 19 years of service is a big step. \
> > The proposal is to leave the existing implementation until there is confidence \
> > with the new implementation. It also serves as a backup in case Windows were to \
> > significantly change the AFD interface. The old Selector can be used by running \
> > with -Djava.nio.channels.spi.SelectorProvider=sun.nio.ch.WindowsSelectorProvider. \
> > A selection of existing tests have been updated to run a second time with the old \
> > Selector so that it is continues to be tested. 
> > CSR and Release Note planned.
> 
> Alan Bateman has updated the pull request with a new target base due to a merge or \
> a rebase. The incremental webrev excludes the unrelated changes brought in by the \
> merge/rebase. The pull request contains 14 additional commits since the last \
> revision: 
> - Merge
> - Windows 32-bit support
> - Update micro to bind to loopback address
> - Merge
> - Bring back getDescriptor, bzero poll array when initially created
> - epoll_create1 returns NULL on error
> - Remove unused WEPoll.getDescriptor. Update micro to close listener
> - Fix copyright header on SelectOne benchmark
> - Merge
> - Propagate exception from discardOOB, cleanup
> - ... and 4 more: https://git.openjdk.java.net/jdk/compare/29d66ac3...d2be1c3a

A good improvement. Approved.

-------------

Marked as reviewed by bpb (Reviewer).

PR: https://git.openjdk.java.net/jdk/pull/3816


[prev in list] [next in list] [prev in thread] [next in thread]