[
https://issues.apache.org/jira/browse/CASSANDRA-14765?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Dinesh Joshi updated CASSANDRA-14765:
-------------------------------------
Description:
*Setup:*
* Cassandra: 6 (2*3 rack) node i3.8xlarge AWS instance (32 cpu cores, 240GB
ram) running cassandra trunk with Jason's 14503 changes vs the same footprint
running 3.0.17
* One datacenter, single tokens
* No compression, encryption, or coalescing turned on
*Test #1:*
ndbench loaded ~150GB of data per node into a LCS table. Then we killed a node
and let a new node stream. With a single token this should be a worst case
recovery scenario (only a few peers to stream from).
*Result:*
As the table used LCS and we didn't not have encryption on, the zero copy
transfer was used via CASSANDRA-14556. We recovered *150GB in 5 minutes,* going
at a consistent rate of about 3 gigabit per second. Theoretically we should be
able to get 10 gigabit, but this is still something like an estimated 16x
improvement over 3.0.x. We're still running the 3.0.x test for a hard
comparison.
*Follow Ups:*
We need to get more rigorous measurements (over more terminations), as well as
finishing the 3.0.x test. [~sumanth.pasupuleti] and [~djoshi3] are driving this.
was:
*Setup:*
* Cassandra: 6 (2*3 rack) node i3.8xlarge AWS instance (32 cpu cores, 240GB
ram) running cassandra trunk with Jason's 14503 changes vs the same footprint
running 3.0.17
* One datacenter, single tokens
* No compression, encryption, or coalescing turned on
*Test #1:*
ndbench loaded ~150GB of data per node into a LCS table. Then we killed a node
and let a new node stream. With a single token this should be a worst case
recovery scenario (only a few peers to stream from).
*Result:*
As the table used LCS and we didn't not have encryption on, the zero copy
transfer was used via CASSANDRA-14556. We recovered *150GB in 5 minutes,* going
at a consistent rate of about 3 gigabit per second. Theoretically we should be
able to get 10 gigabit, but this is still something like an estimated 16x
improvement over 3.0.x. We're still running the 3.0.x test for a hard
comparison.
*Follow Ups:*
We need to get more rigorous measurements (over more terminations), as well as
finishing the 3.0.x test. [~sumanth.pasupuleti] and [~djoshi] are driving this.
> Evaluate Recovery Time on Single Token Cluster Test
> ---------------------------------------------------
>
> Key: CASSANDRA-14765
> URL: https://issues.apache.org/jira/browse/CASSANDRA-14765
> Project: Cassandra
> Issue Type: Sub-task
> Reporter: Joseph Lynch
> Priority: Major
>
> *Setup:*
> * Cassandra: 6 (2*3 rack) node i3.8xlarge AWS instance (32 cpu cores, 240GB
> ram) running cassandra trunk with Jason's 14503 changes vs the same footprint
> running 3.0.17
> * One datacenter, single tokens
> * No compression, encryption, or coalescing turned on
> *Test #1:*
> ndbench loaded ~150GB of data per node into a LCS table. Then we killed a
> node and let a new node stream. With a single token this should be a worst
> case recovery scenario (only a few peers to stream from).
> *Result:*
> As the table used LCS and we didn't not have encryption on, the zero copy
> transfer was used via CASSANDRA-14556. We recovered *150GB in 5 minutes,*
> going at a consistent rate of about 3 gigabit per second. Theoretically we
> should be able to get 10 gigabit, but this is still something like an
> estimated 16x improvement over 3.0.x. We're still running the 3.0.x test for
> a hard comparison.
> *Follow Ups:*
> We need to get more rigorous measurements (over more terminations), as well
> as finishing the 3.0.x test. [~sumanth.pasupuleti] and [~djoshi3] are driving
> this.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
