Effective Use of Lossy Compression for Numerical Linear Algebra Resilience and Performance

Research topic and goals

Prior work has shown the effectiveness of using a multigrid hierarchy to compress the iterative solution to reduce checkpoint size. Recently specially designed floating-point lossy compression algorithms have gained increased popularity due to their ability to significantly reduce floating-point data size and bound the error at each data point. The focus of this collaboration is to explore how to use compression techniques to improve checkpoint-based recovery routines for multigrid methods used either as a stand alone linear system solver or as a preconditioner for other methods such as GMRES or CG. In particular, we explore hybrid compression techniques that switch between the lossy compressor SZ developed at ANL and multi-level multigrid compression in order to minimize the resilience overhead. When using SZ we explore various methods of setting the compressor’s error bounds such as a static tolerance and one that varies the tolerance based on the current accuracy of the solver. To improve the quality of the compressed checkpoint we create and solve a local problem based on the patch of the solution that we are recovering. Finally, we plan to create performance models to explore theoretical possibilities for new compressor designs and system architectures.

The project involves JLESC Fellow Jon Calhoun, JSC affiliated graduate student Mirco Altenbernd, Robert Speck from JSC, and Franck Cappello from Argonne. This project was formed though collaborative discussions at the 2018 JLESC workshop in Barcelona.

Results for 2019/2020

In-progress results from comparing SZ lossy compression vs multigrid compression for checkpointing iterative linear solvers was presented at the SPPEXA workshop in Dresden (October 2019). Currently, a journal article based on the results in being prepared for submission to the IEEE Transactions on Parallel and Distributed Systems.

Preliminary results on evaluating current state-of-the-art compressors for use inside the pySDC application from Robert Speck were accepted and presented as an ACM student research competition (SRC) poster at SC’19. Results showed that specially designed lossy compressors are effective at reducing the runtime memory size, but long compression/decompression times results in slowdown.

Results for 2020/2021

Integration of in-line lossy compression was explored using ZFP compressed arrays. ZFP compressed arrays use a block-based compression scheme to enable random access into the compressed data. Moreover, ZFP compressed arrays use a software cache to mitigate the impact of decompression time on the critical path. During this year, we ran several HPC computation kernels (e.g. matrix-matrix multiplication, sparse matrix-vector multiplication) and two Department of Energy mini-apps Branson and Pennant. Results show that use of in-line compression reduces runtime memory requirements, but does degrade runtime performance by an order-of-magnitude. Error in the mini-app solutions compared to a version without lossy compression grows logarithmically with compressions in each iteration. From this exploration, we plan to build a generic compression runtime system to support any compressor as an in-line compressor. This generic framework will be integrated into the pySDC application of Robert Speck (JSC) for further testing.

To protect compressed data from catastrophic corruption due to silent data corruption caused by a soft error, we develop the automated resiliency for compression (ARC) tool. ARC applies error-correcting codes to lossy compressed data to detect/correct silent data corruption. Results show that ARC is able to protect against potentially100-1000s of bit-flips in the compressed arrays with encoding and decoding bandwidths up to 3.6 GB/s on a modern multi-core CPU.

Results for 2021/2022

The ARC library developed in the previous year was published at HPDC 2021. The source code is publicly available on github. This work is being extended to work for in-compression and data transfer scenarios.

Development of an in-line compression library for large data arrays is in development. This year’s progress explores the performance impact of operating on in-line compressed arrays in HPC kernels (e.g., matrix-matrix multiplication, FFT, SpMV, Jacobi). We explore several strategies of using in-line compressed arrays and determine how applicable different types of parallelization are. This work is being compiled into a publication for 2022. To improve the performance of algorithms that operation on in-line compressed arrays, we can cache in a software managed cache decompressed regions that are part of the current working set. We are developing a cache simulator for in-line compressed arrays to determine the best configuration to maximize performance for certain kernels and access patterns. This work is being compiled into a publication for 2022. These works will be used in the current integration of lossy compression into pySDC.

Results for 2023/2024

We experimented with compressing parts of the collocation problem, which requires keeping multiple solution-size objects in memory. Sansriti Ranjan at Clemson University implemented a version with a cache in pySDC. Once the cache fills up, older objects are compressed and decompressed only when needed. She analysed this with lossless compression using zstd for various types of problems and wrote it up in her masters thesis. Unfortunately, she was unable to significantly reduce the memory footprint and the computational overhead due to compression was very large. We showed in some simple experiments that SDC can still converge to limited accuracy when using lossy compression. Continuing Sansriti’s work with lossy compression is planned by Jon as this is more promising for reducing the memory requirements.

Visits and meetings

  • Completed: Visit by Mirco Altenbernd to Clemson, Spring 2019 for 2 months.
  • Completed: Present talk at JLESC Meeting in Knoxville, April 2019.
  • Completed: Brief meetup and presented in progress work at SPPEXA workshop in Dresden, October 2019.
  • Completed: Visit by Robert and Thomas to Clemson, Spring 2023 for 1 week.

Impact and publications

  • One ACM student research competition poster at Supercomputing 2019.
  • Publication at HPDC 2021 on Resilience of lossy compressed data (Fulp et al. 2021).
  • One ACM student research competition poster at Supercomputing 2021.

    Future plans

    Create performance models for using lossy compression inside running HPC applications. Integrate lossy compression algorithm into applications and measure the impact on performance / accuracy.

    References

    1. Fulp, Dakota, Alexandra Poulos, Robert Underwood, and Jon C. Calhoun. 2021. “ARC: An Automated Approach to Resiliency for Lossy Compressed Data via Error Correcting Codes.” In Proceedings of the 30th International Symposium on High-Performance Parallel and Distributed Computing, 57–68. HPDC ’21. New York, NY, USA: Association for Computing Machinery. doi:10.1145/3431379.3460638.
      @inproceedings{Fulp:HPDC2021:ARC,
        author = {Fulp, Dakota and Poulos, Alexandra and Underwood, Robert and Calhoun, Jon C.},
        title = {ARC: An Automated Approach to Resiliency for Lossy Compressed Data via Error Correcting Codes},
        year = {2021},
        isbn = {9781450382175},
        publisher = {Association for Computing Machinery},
        address = {New York, NY, USA},
        url = {https://doi.org/10.1145/3431379.3460638},
        doi = {10.1145/3431379.3460638},
        booktitle = {Proceedings of the 30th International Symposium on High-Performance Parallel and Distributed Computing},
        pages = {57–68},
        numpages = {12},
        keywords = {sz, error correcting codes, zfp, error propagation, error-bounded lossy compression, silent data corruption, soft error},
        location = {Virtual Event, Sweden},
        series = {HPDC '21}
      }
      
      Progress in high-performance computing (HPC) systems has led to complex applications that stress the I/O subsystem by creating vast amounts of data. Lossy compression reduces data size considerably, but a single error renders lossy compressed data unusable. This sensitivity stems from the high information content per bit in compressed data and is a critical issue as soft errors that cause bit-flips have become increasingly commonplace in HPC systems. While many works have improved lossy compressor performance, few have sought to address this critical weakness.This paper presents ARC: Automated Resiliency for Compression. Given user-defined constraints on storage, throughput, and resiliency, ARC automatically determines the optimal error-correcting code (ECC) configuration before encoding data. We conduct an extensive fault injection study to fully understand the effects of soft errors on lossy compressed data and how to best protect it. We evaluate ARC’s scalability, performance, resiliency, and ease of use. We find on a 40 core node that encoding and decoding demonstrate throughput up to 3730 MB/s and 3602 MB/s. ARC also detects and corrects multi-bit errors with a tunable overhead in terms of storage and throughput. Finally, we display the ease of using ARC and how to consider a systems failure rate when determining the constraints.
    2. Elmore, Donald, and Jon Calhoun. 2019. “Evaluating Lossy Compressors for Inline Compression.” In Poster Session of the 2019 ACM/IEEE International Conference for High Performance Computing, Networking, Storage and Analysis. SC ’19. Washington, DC, USA: IEEE Computer Society.
      @inproceedings{ElmoreCalhoun2019,
        title = {Evaluating Lossy Compressors for Inline Compression},
        author = {Elmore, Donald and Calhoun, Jon},
        booktitle = {Poster Session of the 2019 ACM/IEEE International Conference for High Performance
              Computing, Networking, Storage and Analysis},
        series = {SC '19},
        year = {2019},
        publisher = {IEEE Computer Society},
        address = {Washington, DC, USA}
      }