OK, thanks. Will investigate the possibility.
If this is caused by a network timeout or long GC pause as you
suggested, shouldn't the node be expected to rejoin the cluster once the
problem has gone away?
On 01/08/18 15:46, Denis Mekhanikov wrote:
Tim,
This is strange, since /failureDetectionTimeout/ and
/clientFailureDetectionTimeout/ are pretty important properties
and Ignite maintainers encourage users to tune timeouts using them.
I think, you should send a feature request to Nextflow and ask them to
add this property to configuration.
But once again, this functionality is currently broken in Ignite, and
it's going to be fixed in 2.7
Denis
ср, 1 авг. 2018 г. в 17:25, Tim Dudgeon <tdudgeon...@gmail.com
<mailto:tdudgeon...@gmail.com>>:
Hi
Yes, I saw that property, but:
1. I wasn't sure what the default was and what to use for a more
tolerant value
2. The Nextflow framework does not seem to allow to set this. Only
these Ignite params seem to be controllable:
https://www.nextflow.io/docs/latest/ignite.html#advanced-options
TIm
On 01/08/18 15:17, Denis Mekhanikov wrote:
As per failing node: looks like it failed due to a network issue
or a long GC pause.
Try increasing value of
IgniteConfiguration.html#clientFailureDetectionTimeout
<https://ignite.apache.org/releases/latest/javadoc/org/apache/ignite/configuration/IgniteConfiguration.html#setClientFailureDetectionTimeout-long->
property.
Denis
ср, 1 авг. 2018 г. в 17:00, Denis Mekhanikov
<dmekhani...@gmail.com <mailto:dmekhani...@gmail.com>>:
Tim,
By default IP finder cleans unreachable addresses from the
registry once per minute.
You can change this frequency by setting a different value to
TcpDiscoverySpi.html#setIpFinderCleanFrequency
<https://ignite.apache.org/releases/latest/javadoc/org/apache/ignite/spi/discovery/tcp/TcpDiscoverySpi.html#setIpFinderCleanFrequency-long->
configuration
property.
You shouldn't be concerned about these files too much.
Unreachable addresses in IP finder usually don't have any
negative impact.
IP finder addresses are used only during initial node lookup.
After that nodes exchange their addresses, and use only
those, which are actually connected to the cluster.
Denis
ср, 1 авг. 2018 г. в 11:51, Tim Dudgeon
<tdudgeon...@gmail.com <mailto:tdudgeon...@gmail.com>>:
I'm hitting a strange problem when using an ignite
cluster for
performing compute jobs.
Ignite is being managed by the underlying Nextflow tool
(https://www.nextflow.io/docs/latest/ignite.html) and I
don't understand
the precise details of how this is set up, but I believe
there's nothing
unusual going on.
Cluster discovery is being done using a shared directory.
The cluster is set up fine, and jobs are executed on all
the nodes as
expected. Then some random event happens on a node which
results in it
leaving the cluster. The nextflow/ignite process is still
running on the
node, and the tasks that are currently executing continue
to execute to
completion, but no new tasks are started. And the node is
still
registered in the shared cluster directory. But on the
master the node
is seen to have left the cluster and no longer consumes
jobs, and never
rejoins.
We are seeing this on an OpenStack environment. When we
do the same on
AWS the problem is not encountered. So presumably there
is something
strange going on at the network level to cause this.
Possibly changing
some of the timeouts might help. The ones that Nextflow
allows to change
are listed here:
https://www.nextflow.io/docs/latest/ignite.html#advanced-options
But its not clear to me which timeouts should be changed,
and what new
values to try. Any advice here would be most welcome.
For an example of this in action look here for the
Nextflow log file on
the worker node, which includes various log output from
Ignite:
https://gist.github.com/tdudgeon/2940b8b1d1df03aecb7d13395cfb16a8#file-node-nextflow-log-L1109-L1265