[ 
https://issues.apache.org/jira/browse/HDFS-6287?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13989791#comment-13989791
 ] 

Colin Patrick McCabe commented on HDFS-6287:
--------------------------------------------

bq. Hi, Colin. Thanks for posting this. Did you find that you needed to use SSE 
to get the addition fast enough so that the benchmark highlights read 
throughput instead of sum computation? IOW, could we potentially simplify this 
patch to not use SSE at all and still have a valid benchmark?

Without that optimization, the benchmark quickly becomes CPU-bound and you 
don't get true numbers for ZCR and other fast read methods.  I just benchmarked 
1.5 GB/s for the un-optimized version versus 5.7 GB/s for the optimized.

bq. I think it would be helpful to add a comment with a high-level summary of 
what vecsum does, maybe right before the main.

Added

bq. I have one minor comment on the code itself so far. I think you can remove 
the hdfsFreeBuilder call. hdfsBuilderConnect always frees the builder, whether 
it succeeds or fails. The only time you would need to call hdfsFreeBuilder 
directly is if you allocated a builder but then never attempted to connect with 
it. I don't see any way for that to happen in the libhdfs_data_create code.

Yeah, that is deadcode.  Let me remove that

> Add vecsum test of libhdfs read access times
> --------------------------------------------
>
>                 Key: HDFS-6287
>                 URL: https://issues.apache.org/jira/browse/HDFS-6287
>             Project: Hadoop HDFS
>          Issue Type: Test
>          Components: libhdfs, test
>    Affects Versions: 2.5.0
>            Reporter: Colin Patrick McCabe
>            Assignee: Colin Patrick McCabe
>            Priority: Minor
>         Attachments: HDFS-6282.001.patch, HDFS-6287.002.patch, 
> HDFS-6287.003.patch, HDFS-6287.004.patch, HDFS-6287.005.patch
>
>
> Add vecsum, a benchmark that tests libhdfs access times.  This includes 
> short-circuit, zero-copy, and standard libhdfs access modes.  It also has a 
> local filesystem mode for comparison.



--
This message was sent by Atlassian JIRA
(v6.2#6252)

Reply via email to