Logo image
Back
Designing Domain-Specific Compilers for Lossy Compression: A Case Study on Wafer-Scale Engine
Conference proceeding

Designing Domain-Specific Compilers for Lossy Compression: A Case Study on Wafer-Scale Engine

Shihui Song, Robert Underwood, Sheng Di, Peng Jiang and Franck Cappello
Proceedings - IEEE International Parallel and Distributed Processing Symposium, pp.1091-1104
05/25/2026
DOI: 10.1109/IPDPS65963.2026.00092

View Online

Abstract

Modern HPC applications generate data at extreme rates, making error-bounded lossy compression essential to reduce the data scale. Emerging accelerators such as the Cerebras Wafer-Scale Engine (WSE) offer tremendous computational potential for these workloads, yet developing high-performance compressors on the WSE remains challenging due to its distributed dataflow architecture, limited on-chip memory, and complex inter-PE communication. In this paper, we propose P3Z, a domain-specific compiler that lets users define compressors in a set of Python definitions and automatically generates optimized code for both the Cerebras WSE and CPU backends. P3Z represents programs as an Abstract Syntax Graph (ASG), lowers to an intermediate representation (IR), applies a series of IR-level and hardware-aware optimizations, and then generates the platform code. P3Z also supports semantic-aware automatic decompression, which derives decompression kernels directly from the compression pipeline. Across seven real-world datasets, P3Z achieves 1.24×-1.5× higher throughput than hand-written CSL while reducing development effort by over 14×.
 Data Compression High Performance Computing Optimization Codes Compiler Compressors Conferences Dataflow Architecture Domain Specific Language Information rates Memory Printing Program processors Throughput

Details

Metrics

1 Record Views
Logo image