look into ray distributed computing framework

[vvksh]

came out of RISELab at Berkeley, supposed to be a general purpose distributed computing framework.

arxiv paper here

notes:

• special focus on RL systems since RL workloads require training, simulation and serving

"""elative inflexibility of the BSP model, the high per-task overhead, and the lack of an actor abstraction led us to develop a new system"""

fault tolerance helps save money since it allows us to run on cheap resources like spot instances on AWS

extra systems discussed:

• Clipper (for model serving),
• Orleans+Akka (actor systems)
• Sparrow, Mesos (schedulers)
• Cilk (parallel programming language)

Qs:

• what is BSP execution model? spark and mapreduce implement it

tasks:

• [ ] #21 sample cluster run
• [ ] #22 autoscaling

[vvksh]

also this medium blog gives a nice intro starting from hadoop -> spark -> ray; mainly focuses on why spark is not suited for async tasks

there;s also this report by oreily

[vvksh]

this blog shows benchmark comparing Ray vs Spark vs Dask for a given task; might be interesting to replicate it on rpi cluster or even azure.

[vvksh]

look into ray cluster and play with autoscaling: https://docs.ray.io/en/master/cluster/index.html