PAW-ATM Panel

Julian Samaroo (MIT JuliaLab)

Who am I?

I am a Research Software Engineer at MIT's JuliaLab. I work 95% on the Julia language, with a focus on developing libraries for distributed, multi-threaded, and GPU paradigms. I built and maintain Julia's AMDGPU support, and currently focus on developing the Dagger.jl HPC library.

And yes, that is my Golden Retriever in the picture :)

My MPI+X Alternative:

• Pure-Julia task parallelism library
• Unified task API supports multiple high-level APIs (arrays, tables, graphs, streaming, data dependencies)
• Supports multi-threading, multi-node, multi-GPU out of the box
• Machine learning-based scheduler optimizes computational DAGs based on learned task behavior
• Advanced data dependency system automatically parallelizes operations on overlapping data regions (like OpenMP's "task depend" on steroids)

Applications of Dagger.jl

• Quantum Tensor Network Contraction (BSC)
• HEP Streaming Framework (CERN)
• Reinsurance Pricing Engine (Ark)
• ...and more

Advantages

• Removes cognitive burden of manual task scheduling and data transfers
• No/minimal code changes to support different accelerators
• Scales automatically to available hardware
• Variety of built-in parallelized operations
• Multiple APIs for different problems

Known Issues

• Relatively high per-task overhead
  
• Missing optimizations for NUMA, optimal scheduling, many small tasks, etc.
  
• High memory allocations (count and size)
  
• Missing extensive (necessary) docs for new users
  
• Bus factor == 1 (approximately)