Projects

Hardware Stacks

Hardware Data Structures

Implementation and protocol for efficient host/accelerator-to-accelerator intercommunication.

Sub-projects

FSHMEM: GasNet-FPGAApache Arrow*
System Software

Hybrid Programming Model

Performance prediction model and software abstractions for accelerator-HPC tools integration.

Sub-projects

PGAS/OpenSHMEMHybrid SPMD-MPSD
Data Engineering

Edge Neural Network

Graph partitioning, optimization, and dataflow schemes for large DNNs.

Sub-projects

CNN AND RNNRecommender

Principles

Workflow-oriented

Easily define your application as a workflow then integrate your data all in a user-friendly programming environment.

Automated

Workflows are composed of either statically or dynamically generated tasks. PCS maps these tasks for efficient execution.

Heterogeneous

Program execution occurs as sets of hardware microservices, leveraging high-performance compute backends like FPGA, CPU and ASIC.

Features

Domain-specific language (DSL)

Source-to-source translation and remote memory access (RMA) abstractions incorporated with familiar frameworks like TensorFlow and Airflow.

Scalable accelerator-aware orchestration

Data/task parallel processing under a single, unified view of compute, communication, and memory resources.

Throughput-aware distributed runtime system

Load-balancing of intra-node against inter-node communication while accommodating for compute-pool heterogeneity and bulk synchronizations.

Latency-aware accelerator-to-accelerator intercommunication

Exploitation of one-sided communication and reduced intercommunication overhead through GasNet and PGAS hardware-based libraries.

Feature Overview

Organizations using and contributing to PCS

Flapmax

University of Florida

Oak Ridge National Lab

Brookhaven National Lab

Lawrence Livermore National Lab