[
https://issues.apache.org/jira/browse/CASSANDRA-8464?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14260733#comment-14260733
]
Benjamin Janssen commented on CASSANDRA-8464:
---------------------------------------------
Out of curiosity, how does this ticket play with CASSANDRA-6890 which proposes
removing mmapped I/O from the read path? Is removing mmapped still the plan?
We're using compression for our SSTables and we've seen some issues with
compaction being unable to keep up, not because we're disk I/O bound but
because we are spending so much time in system calls since memory mapped I/O
doesn't seem to be used when compression is enabled. We'd love to see memory
mapped support for compressed SSTables (which our understanding is this ticket
would provide) but would hate to see memory mapped I/O removed, which the other
ticket seems to suggest will happen.
> Support direct buffer decompression for reads
> ---------------------------------------------
>
> Key: CASSANDRA-8464
> URL: https://issues.apache.org/jira/browse/CASSANDRA-8464
> Project: Cassandra
> Issue Type: Improvement
> Reporter: T Jake Luciani
> Assignee: T Jake Luciani
> Labels: performance
> Fix For: 3.0
>
> Attachments: compression_direct.png
>
>
> Currently when we read a compressed sstable we copy the data on heap then
> send it to be de-compressed to another on heap buffer (albeit pooled).
> But now both snappy and lz4 (with CASSANDRA-7039) allow decompression of
> direct byte buffers. This lets us mmap the data and decompress completely
> off heap (and avoids moving bytes over JNI).
> One issue is performing the checksum offheap but the Adler32 does support in
> java 8 (it's also in java 7 but marked private?!)
> This change yields a > 10% boost in read performance on cstar. Locally I see
> upto 30% improvement.
> http://cstar.datastax.com/graph?stats=5ebcdd70-816b-11e4-aed6-42010af0688f&metric=op_rate&operation=2_read&smoothing=1&show_aggregates=true&xmin=0&xmax=200.09&ymin=0&ymax=135908.3
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
