Hello, Fault tolerance patch should be in this week.
The other features I want to throw for consideration are written below. Please feel free to prioritize for 0.6: BSPPeer API Improvements - Broadcast API We have to experiment a broadcast API that does not use the Messaging queue. Few strategies for the purpose could be brainstormed. - HAMA-546 Task Scheduler Improvements - Task Allocation as feature of BSP job instead of system (Coming with the Fault tolerance patch) Today the task allocation responsibility is wrongly implemented in TaskInProgress class. With this feature we can have pluggable task allocation strategy with data-local allocation as default strategy. - To support real time jobs, we would need to implement resource local scheduler, where resource here is more than the file input split. Based on the nature of job, it could be nodes with bigger pipe for communication, nature of the machine(GPU), Nodes with SSDs, etc. Job Submission Improvements - A Job Submission API, that gives users the flexibility to define the nature of messaging queue and other attributes per superstep Improve Performance And Scalability - Benchmark on more nodes - Try Asynchronous communication - HAMA-593 - Spilling message queue and External Sorting Queue. Apart from this, there could be more graph algorithms to be implemented, ml algorithms to be implemented and Mikalai's work integrated in the release. When are we planning to have next release? I can take a shot at release management, but I would definitely need help for the first time. Thanks, Suraj On Mon, Jul 9, 2012 at 3:22 AM, Thomas Jungblut <[email protected]>wrote: > So now we released 0.5.0 what are our plans for 0.6.0? > And who is going to do the release managing? >
