/
2024-07-25 Linaro Data Center Group steering committee Meeting Agenda/Minutes

2024-07-25 Linaro Data Center Group steering committee Meeting Agenda/Minutes

Owned by Leonardo Garcia
Last updated: Aug 01, 2024

Meeting Date: 2024-07-25

Attendance
Steering Committee Members 


Name

Alternate
Leonardo Garcia, Linaro(tick)
Tom Gall, Linaro(error)
Andy Wafaa, Arm(tick)
Ashok Bhat, Arm(error)
Shun Kamatsuka, Fujitsu(tick)


Jonathan Cameron, Huawei(tick)
Kangkang Shen, FutureWei(error)
, Red Hat(error)


Jon Masters, Google(error)



Guests

Name

Kevin Zhao, Linaro

(tick)
Gen Shimada, Linaro(error)
Mikael Rolfhamre, Linaro(tick)
Marc Meunier, Arm(tick)

Guodong Xu, Linaro

(tick)
Randy Linnell, Linaro(tick)
Joyce Qi, Linaro(tick)
Kaly Xin, Linaro(tick)

Meeting Logistics

When? 4th Thursday of each month.

Zoom connection details

 Agenda

Documents

Action Items

  •  

Minutes

  • Project updates
    • Deploy CCA on Arm Platforms (DCAP)
      • Upgraded build environment for our build recipes.
      • Key broker API design and formalisation.
      • Configured and set up CI machines for Kata Containers on Arm64 in the community.
    • System Ready Compliant Platform Collaboration (SRCPC)
      • Added CPU topology into Trusted Firmware -A and QEMU.
      • Handled EFI_RNG_PROTOCOL in EDK2. Neoverse-N2 and max (in QEMU) processors will have it enabled.

    • Big Data and Data Science (BDDS)

      • Ongoing hiring process.

    • Enhanced Software-Defined-Storage on Arm Servers (STOR)

      • Provided Lustre rpms on openEuler 22.03 LTS SP4.
      • Reviewed patches from Lustre community.
  • Host Security ID (HSI)
    • This topic was initially discussed in the Data Center Group steering committee meeting that happened in Madrid, during Linaro Connect MAD24, on 2024-05-16.
    • Recap:
      • HSI is being used by customers to list requirements for x86 servers. This is not well supported on Arm platforms.
      • Arm platforms would probably need to work with the HSI standard to include their features in the appropriate HSI category.
    • This discussion has moved to SystemArchAC Security working group. They are discussing directly with Richard Hughes, from Red Hat, who maintains HSI.
  • UXL
    • Members to bring specific requirements around UXL and whether they can provide engineers to help on this work. The discussion was mainly around library optimization.
      • Andy:
        • oneDNN is working, mainly enabled and optimized by Fujitsu.
        • oneDAL patches from Fujitsu being discussed upstream.
        • oneMKL: there is some work being done.
        • Arm just pushed out an open source project called openRNG that we should be able to use with oneMKL as a reference back-end.
        • Most things work, but there is still tunning to be done.
        • There are also gaps in CI. CI was all held and run behind Intel's firewall. Arm is working to add Arm instances to the CI.
        • There are discussions around having a C library to get around all the SYCL requirements. There is also all the SYCL upstream effort going on, to get its support into LLVM.
        • No concrete requirements for the Linaro Data Center Group yet. Trying to figure the gaps that need to be filled.
        • UXL will be hosting oneAPI dev summits in Asia and Europe/USA.
  • Optimization potential future work
    • We had discussed many times which frameworks or projects we should be optimizing. Even though everyone agrees this is an important area to explore, there hasn't been a clear consensus on which direction we should take.
    • Some direct projects, such as STOR and BDDS in the past, had worked on optimizing specific areas that were not well performant.
    • Arm provides very little information about the work they are doing in project Interstellar.
    • In the last steering committee meeting, Fujitsu talked about the work they are doing to optimize AI frameworks for Arm (oneDNN and oneDAL).
    • Everybody agreed that Linaro should be coordinating any collaboration in this area. So I decided to build a list of potential projects that might be good targets for optimization on Arm64 platforms and we can go from there. I got valuable input from Linaro employes Carlos Seo and Everton Constantino (Deactivated), who have expertise on performance optimization, to build this list.
    • Important reminder: any optimization work starts with an investigation phase. It is not possible to jump directly in the optimization work without making a good performance analysis of the target project.
    • Projects with potential performance optimization work on Arm64 platforms (I didn't check current status of these projects):
      • UXL projects
        • oneDNN: Fujitsu done optimization work. Anything else left to explore?
        • oneDAL: Fujitsu done optimization work. Anything else left to explore?

      • Computer Vision

        • OpenCV: SIMD enablement via OpenCV's HAL (Hardware Abstraction Layer)?

        • FFMPEG2: SIMD enablement?

      • BLAS

        • OpenBLAS: SIMD and SME enablement?

        • Eigen: SIMD and SME enablement?

        • BLIS: SIMD and SME enablement?

      • Boost

        • XGBoost

        • LightGBM: CUDA only today. SYCL version? UXL support?

      • Other maths

        • NumPy: is there anything left to do?

        • SciPy: is there anything left to do?

      • DNN

        • MKL-DNN

      • Algorithms that might be important (need to find relevant libraries):
        • Classification Analysis
          • Decision Tree Analysis
          • K Nearest Neighbour
          • Linear Discriminant Analysis (LDA)
          • Confusion Matrix
        • Cluster Analysis
          • K-Means
          • Agglomerate Hierarchical
          • Affinity Propagation
          • Latent Dirichlet Allocation (LDA)
          • Gaussian Mixture Model (GMM)
        • Cryptographic algorithms
          • OpenSSL
          • MBedTLS
          • Language runtimes (Golang)?

        • Data preparation

          • Sampling

          • Principal Component Analysis

        • GPU offload
          • OpenMP target
          • OpenACC?
          • SYCL?
        • Time Series Analysis
          • Fast Fourier Transform (FFT)
          • Correlation Function
        • Regression
          • Generalized Linear Model (GLM)
        • Compression / decompression
          • Zlib?
        • Association Analysis
        • Probability Distribution
        • Outlier Detection
        • Recommender Systems
        • Link Prediction
        • Statistical Functions
      • Kleidi
        • Arm recently introduced Kleidi, a low-level CPU instruction library targeting Artificial Intelligence and Computer Vision workloads for the moment. This library is specifically targeting CPUs, by providing a common layer with algorithms in these areas optimized for specific CPU implementations (Armv8 NEON and Armv9 SVE2 and SME2).
        • Right now, KleidiAI has just a small number of NEON based kernels (matrix-multiply) upstreamed. Is this an area the Data Center Group should be looking at? My understanding is that current algorithms are optimized only for Arm microarchitectures.
        • Andy: initial focus is client, not server.
    • Andy: a lot of the work being done at Arm w.r.t. performance optimizaiton is related to making sure PyTorch runs and performs well on Arm platforms.
    • Jonathan: Huawei has been working on https://www.hikunpeng.com/en/developer/boostkit/library that overlaps some of the projects and algorithms listed above.
    • Kamatsuka-san:
      • Apart from oneDNN and oneDAL, Fujitsu is also working to optimize OpenBLAS.
      • Fujitsu has also interest in libXSMM, which is a library for small matrix multiplication, but is not working on it yet.
      • Also interested in OpenSSL, but also not working on it yet.
  • Events
  • AOB
    • Marc Meunier
      • Paul Howard (Arm) and Thomas Fossati (Linaro) are working on the creation of an endorsement framework for Confidential Computing. They want to pave the way for a standard framework for endorsement feeds in the Confidential Computing verifier and want feedback from the community. Leonardo will invite them to present their work in the next steering committee meeting.

Related content