Releases: FederatedAI/eggroll
Releases · FederatedAI/eggroll
v2.3.1
v2.2.1
v2.3.0
v2.2.0
- RollSite: Redesigned RollSite to support polling mode. One federated learning party between the two can have no public listening port.
- RollSite: Reimplemented provide better performance, better robustness and better memory usage.
- RollPair: Removed RollPairMaster to boost job schedule performance.
v2.0.2
v2.1.0
v2.0.1
Optimizations and robustness improvements.
Main changes:
- Supports different
total_partitionbefore / aftermapoperation. Shuffles correctly whentotal_partitionof binary op are different. - Reimplements GC in rollpair.
- Optimizes creations of gRPC servers / clients.
- Optimizes retry mechanism in rollsite.
- Optimizes deployment and operation scripts.
- Fixes deadlock issues in database connections.
- Removes dependencies on MyBatis and many python libraries.
- Lots of other minor enhancements e.g. logs, more parameter configurations, memory optimizations, LMDB usage optimizations.
v2.0.0
Next generation of Eggroll - THE HPC infrastructure.
- New module: Cluster Manager. Cluster manager manages all physical resources of the cluster.
- New module: Node Manager. Node manager manages life cycles of Eggroll computing session.
- Roll Pair: Brand new design of kv based computing. Hugely optimises performances of shuffle.
- Roll Site: Enhancements and integration of 1.x's proxy + clustercomm (i.e. federation).