Github user nicolas-kourtellis commented on the issue:
https://github.com/apache/incubator-samoa/pull/54
Hi @manuzhang,
I managed to get the adapter working. Here are some notes that I would ask
you take into consideration:
- There are some inherent difficulties compiling gearpump from source. It
would be good to have a compiled version to use directly.
- Assuming this is given (which was my case because @manuzhang provided a
compiled version), I managed to get samoa to compile/package with gearpump and
run the package.
- However, it would be good for the adapter to be upgraded to the new
version of samoa in incubation, which is 0.5.0. But it should be fairly
straightforward. This will allow us to test it with some more generators and ML
methods added in the recent past.
- Feedback when executing VHT:
=> The engine seems to continue executing the topology long after it has
been created, used for the task and finished. Is there any way to pass a signal
at the end of the execution to shut it down? (note: not the engine itself, but
the topology). It was occupying resources on my computer for no reason at full
CPU consumption. I found a manual way to kill it using the command "gear kill
-appid X" with X being the id of the task, but I wonder if there is a more
automatic way.
=> After I killed the jobs manually, the java processes that were created
for the execution (I will assume they are the containers of the topologies)
were still alive, just not consuming much resources. Shouldn't they have been
terminated and removed? Is there a way to do that?
=> When I run new tasks, they just keep getting added on the engine (which
is logical), even though I had killed the other ones earlier.
=> Multiple executions of the same experiment with the same seed for the
random generator using the parameter -r which should yield the same random
tree, perform differently with respect to accuracy. Is that expected?
=> Using a different seed for the random tree generator (r=1,...,5), the
performance of the execution of VHT on local GearPump is fairly low (average
over 5 different seeds: 65.39% accuracy) in comparison to running the topology
on local Storm (84.046% accuracy). Any explanation why so much reduction in
performance?
---
If your project is set up for it, you can reply to this email and have your
reply appear on GitHub as well. If your project does not have this feature
enabled and wishes so, or if the feature is enabled but not working, please
contact infrastructure at [email protected] or file a JIRA ticket
with INFRA.
---
