[jira] [Resolved] (IGNITE-1615) .Net: Perform AtomicLong.get() without JNI if possible.

["Vladimir Ozerov (JIRA)" <ji...@apache.org>]

     [ https://issues.apache.org/jira/browse/IGNITE-1615?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]

Vladimir Ozerov resolved IGNITE-1615.
-------------------------------------
    Resolution: Won't Fix

This cannot be done because another node might update the value. So any local optimizations do not work by default.

> .Net: Perform AtomicLong.get() without JNI if possible.
> -------------------------------------------------------
>
>                 Key: IGNITE-1615
>                 URL: https://issues.apache.org/jira/browse/IGNITE-1615
>             Project: Ignite
>          Issue Type: Task
>          Components: interop
>    Affects Versions: ignite-1.4
>            Reporter: Vladimir Ozerov
>             Fix For: ignite-1.5
>
>
> Variables with atomic/interlocked semantics are frequently used in mostly-read scenarios. E.g. in spin-loops, non-blocking alogrithms, as regular volatiles, etc..
> With current implementation we perform JNI call on every read which is too expensive, especially with poor Java performance when performing (native -> JVM) transition.
> We can optimize it with the following non-blocking algorithm:
> 1) Add atomic "stamp" field.
> 2) Add atomic "cached" field.
> 3) On any update:
> - Do the update;
> - Atomically increment the stamp;
> 4) On any read:
> - Read stamp (oldStamp);
> - Read cached value;
> - Read stamp again (newStamp);
> - if (oldStamp == newStamp == cache.stamp()), return cached value.
> - Otherwise perform real read through JNI and update cached value with a pair of (readValue + oldStamp);



--
This message was sent by Atlassian JIRA
(v6.3.4#6332)