Currently, the full suite of autmated tests take 15 to 25 mn to run, depending on the hardware. We propose to reduce it to a more reasonable set of tests that would take much less time, while still providing a meaningful coverage.
We also propose to run this lightweight suite as part of the Travis build setup in github, in addition to making it available locally on the command line. The currentm CI build with Jenkins would still run the full suite.
The Jenkins build is showing intermittent failures for some of our automated tests on multiserver topologies. I'm not sure yet what the problem is exactly, but it seems to occur frequently enough (1 time out of 5) to warrant a dedicated bug report.
The failures I've been seeing recently, and for which a set of logs is attached to this defect:
We propose to add Docker images for JPPF, for drivers, nodes and web admin console. The configuration of a JPPF grid with docker should allow any kind of JPPF topology, including multi-server topologies.
Antoher objectivve is to provide ready-to-use tools and configurations to run JPPF grids within a Docker swarm/Kubernetes infrastructure.
The class [https://www.jppf.org/javadoc/6.2/index.html?org/jppf/scheduling/JPPFSchedule.html JPPFSchedule] is used to specify the start or expiration schedule of a job, as well as the expiration schedule of a task. It currently has 2 basic constructors, one that takes an epoch time in millis, the other that takes a string which represents a date, along with a SimpleDateFormat-compliant format to parse it.
We propose to extends this class to enable building JPPFSchedule objects based on the classes in java.time.*, such as ZOnedDateTime, Duration, etc.
When a job is dispatched to multiple nodes in parallel, this can result in the same class loading request being issued to the same client in parallel or in sequence. This would happen when identical requests are forwarded to the same client, before the first response is received by the server, and therefore before in can be added to the server-side cache. It could be worthwhile, from a perfromance perspective, to use a cache of class definitions, such that identical requests (same client-side class loader and same resource path) only result in a single lookup in the classpath.
To this effect, we propose to implement a cache in the client, as follows:
* an identity hash map whose keys are class loaders
* the values are hash maps where the key is a path in the classpath and the value is the byte for the resource located at that path. These could be implemented as [https://www.jppf.org/javadoc/6.2/index.html?org/jppf/utils/collections/SoftReferenceValuesMap.html SoftReferenceValueMap]s to avoid out of memory conditions due to the cache
The online and offline dosc still show, in the tutorial, code snippe that use the "blocking" job attribute, which is now deprecated, as well as job submission with JPPFClient.submitJob(), now deprecated and replace with submit() and submitAsync
This is about refactoring the distributed class loader communication model into the more efficient and scalable model introduced in JPPF 6.1 (see feature request JPPF-549 and feature request JPPF-564). This includes both driver/node and driver/client communication channels.
These are the last components to switch to the new model. Once it is done, we expect a number of benefits:
* we will be able to get rid of the old nio code, which should reduce the maintenance burden
* this should also increase the performance, simply because we will remove parts of the code inherited from the old model, which are still present but not used in the new model
* increased performance and scalability, because the new nio model is more efficient
We propose to enable dependencies between tasks in the same job. For instance, we envision the ability to express that task A depends on the completion of tasks B and C, who each depend on the completion of task D (diamond dependency graph):
This implies a number of challenges, including but not limited to:
* decide how to schedule and parallelize the execution of the tasks: a task with dependencies cannot be scheduled before its dependenices have completed
* handle failures / cancellation of tasks on which other tasks depend
* provide an expressive and intuitive API to specify the dependencies
We have a [https://www.jppf.org/samples-pack/JobDependencies/ job dependencies sample] which illustrates an ad-hoc way of executing acyclic graphs of dependent jobs.
We propose to make this an actual feature instead of a sample, and to explore other possibilities such as:
* a broader set of relationshups between jobs than just "depends on", e.g. split/join (or map/reduce). This implies to be able to gather global results for an entire job and apply transformations to these results.
* an expressive and intuitive way to build the job graph
Currently, [https://www.jppf.org/doc/6.1/index.php?title=The_Location_API#MavenCentralLocation MavenCentralLocation] only allows to download artifacts from Maven Central. We propose to add the ability to specify a different repository, as well as the ability to download SNAPSHOT artifacts, For instnce in a class nmamed MavenLocation, of which MavenCentralLocation could be a specialized subclass.
[https://www.jppf.org/doc/6.1/index.php?title=Monitoring_data_providers '''Monitoring data providers'''] allow to define properties of various types that are monitored over time. However, there is currently no way to specify how these values should be displayed in the JVM health view of the desktop and web administration consoles. JPPF currently uses default conversions based on the type of each property, but this may not be always convenient.
For instance, let's say we want to monitor the JVM uptime. This value is expressed in millisecons as a long integer value. However, in the GUI we'd rather have it displayed as days:hours:minutes:seconds.millis.
We propose to implement the ability to configure a value converter for each defined property to this effect.
For instance (just for example purpose, this is not what the actual design will be):
I would like to embed jppf-admin-web into my own embedded webserver as an executable jar. I need jppf-admin-web as a jar dependency instead of war to make this work. I would define my own web.xml for this and ignore the one inside
See description here: https://pragmaticintegrator.wordpress.com/2010/10/22/using-a-war-module-as-dependency-in-maven/
You would need to add:
So that I could use:
Also it would be nice if you could define the jppf.css and and images/ as maven resources behind a package name and add those resources into the classes folder. You could then mount those resources in your wicket application under your current paths using PackageResourceReferences to serve them from the classpath. This makes embedding easier and I don't have to copy these resources myself then.