A Performance and Scalability Analysis of Actor Message Passing and Migration in SALSA Lite
This paper presents a newly developed implementation of remote message passing, remote actor creation and actor migration in SALSA Lite. The new runtime and protocols are implemented using SALSA Lite’s lightweight actors and asynchronous message passing, and provide significant performance improvements over SALSA version 1.1.5. Actors in SALSA Lite can now be local, the default lightweight actor implementation; remote, actors which can be referenced remotely and send remote messages, but cannot migrate; or mobile, actors that can be remotely referenced, send remote messages and migrate to different locations. Remote message passing in SALSA Lite is twice as fast, actor migration is over 17 times as fast, and remote actor creation is two orders of magnitude faster. Two new benchmarks for remote message passing and migration show this implementation has strong scalability in terms of concurrent actor message passing and migration. The costs of using remote and mobile actors are also investigated. For local message passing, remote actors resulted in no overhead, and mobile actors resulted in 30% overhead. Local creation of remote and mobile actors was more expensive with 54% overhead for remote actors and 438% for mobile actors. In distributed scenarios, creating mobile actors remotely was only 6% slower than creating remote actors remotely, and passing messages between mobile actors on different theaters was only 5.55% slower than passing messages between remote actors. These results highlight the benefits of our approach in implementing the distributed runtime over a core set of efficient lightweight actors, as well as provide insights into the costs of implementing remote message passing and actor mobility.
Conference DayMon 26 OctDisplayed time zone: Eastern Time (US & Canada) change
10:30 - 12:00
|Bulk-Synchronous Communication Mechanisms in Diderot|
|Programming Abstractions for Augmented Worlds|
|Manyfold Actors: Extending the C++ Actor Framework to Heterogeneous Many-Core Machines using OpenCL|
|A Performance and Scalability Analysis of Actor Message Passing and Migration in SALSA Lite|