[
https://issues.apache.org/jira/browse/CASSANDRA-15396?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17098200#comment-17098200
]
David Capwell commented on CASSANDRA-15396:
-------------------------------------------
Ran a read only test on a 2 dc (6 nodes per dc) cluster with a k/v schema where
the values ranged from 100-102,400 bytes and saw an improvement at p99 and
p99.9 (4 9s has too much jitter in all runs, so ignoring the results, p99 and
p99.9 are mostly consistent across multiple runs); my next steps are to run the
same test with a smaller skew in data size, but still with larger values.
The test was the following
{code}
max_blob_length=102400
num_partitions="500k"
args=(
--compaction "{'class': 'LeveledCompactionStrategy'}"
--compression "{'class': 'LZ4Compressor'}"
--replication "{'class': 'NetworkTopologyStrategy', 'DC1': 3, 'DC2': 3}"
--cl LOCAL_QUORUM
--hdr-dir $PWD
--concurrency 100
--threads 1
--dc DC1
--readrate 1
--deleterate 0
--workload.maxBlobLength=$max_blob_length
--field.keyvaluelargeblob.value="random(100,$max_blob_length)"
--partitions "$num_partitions"
--populate "$num_partitions"
)
tlp-stress run KeyValueLargeBlob "${args[@]}" --duration 30m
{code}
> RandomAccessReader does not override skip or skipBytes so will do a lot of
> disk io when n is large
> --------------------------------------------------------------------------------------------------
>
> Key: CASSANDRA-15396
> URL: https://issues.apache.org/jira/browse/CASSANDRA-15396
> Project: Cassandra
> Issue Type: Improvement
> Components: Local/SSTable
> Reporter: David Capwell
> Assignee: David Capwell
> Priority: Normal
> Attachments: Histogram-3.png
>
>
> RandomAccessReader does not override skip or skipBytes which becomes a
> problem when the size of n (the bytes to skip) is larger than a single
> buffer; in these cases we can rely on seek to avoid the extra disk io.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
