[ 
https://issues.apache.org/jira/browse/AVRO-746?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
 ]

Douglas Creager updated AVRO-746:
---------------------------------

    Attachment: 0001-Performance-test-program.patch

Here's a patch that does a performance test of the reference counting 
mechanism.  It allocates a single int value, and then references and 
dereferences that value 100,000,000 times.  It runs the test 3 times and prints 
out the average time across the 3 runs.

It turns out that the atomic counts are a bigger performance hit than I 
thought.  With the original implementation, which just uses C's ++ and -- 
operators, the test case ran in 0.460 seconds.  Using the new atomic operation, 
it took 1.820 seconds.

That said, I don't know if reference counting is the bottleneck in a "typical" 
use of the library.  I'm going to add some more performance tests to the test 
program to see what we get.

> Atomic reference counts
> -----------------------
>
>                 Key: AVRO-746
>                 URL: https://issues.apache.org/jira/browse/AVRO-746
>             Project: Avro
>          Issue Type: Bug
>          Components: c
>            Reporter: Douglas Creager
>            Assignee: Douglas Creager
>         Attachments: 0001-Atomic-reference-counts.patch, 
> 0001-Performance-test-program.patch, fixed-make-dist.patch
>
>
> In one of my projects that uses Avro, I pass avro_datum_t instances between 
> threads, using the reference count mechanism to make sure that they're not 
> freed while any thread still has a reference to them.  I was getting some 
> spurious segfaults, which were caused by the fact that the reference counts 
> aren't updated atomically.  I've created a patch that implements atomic 
> reference counts, using the [OpenPA 
> library|http://trac.mcs.anl.gov/projects/openpa/] to provide the atomic 
> operations themselves.  (That library is MIT licensed, so it can be included 
> in the source tree.)
> Note that only avro_XXX_incref and avro_XXX_decref are thread-safe as a 
> result of this patch.  For all of the other library functions, the caller is 
> still responsible for ensuring thread safety.
> The patch makes sure that the OpenPA code works in both the old autotools 
> build and the newer CMake build.

-- 
This message is automatically generated by JIRA.
-
For more information on JIRA, see: http://www.atlassian.com/software/jira

        

Reply via email to